The Vesper Bell backlog

Why

During parallel agent work on 2026-10-03, pnpm test failed 1–2 tests intermittently at load average 11–33 (seen on the mw-e11.7 rebase and the mw-e11.9 branch), then passed on rerun. CI has stayed green, but a load-sensitive test will eventually flake in CI too.

What

Reproduce with CPU stress (e.g. run vitest with high --pool concurrency alongside a busy loop, or --repeat), identify the tests (likely wall-clock or timing-budget assertions outside tests/bench), and make them deterministic. No retries.

AC-1 unit The full unit suite passes 5 consecutive runs under a documented CPU-stress reproduction.

AC-2 unit The tests found are fixed to not depend on wall-clock timing (or moved into tests/bench).

Why

Spike prototypes sit outside every quality gate on purpose (spikes/README.md). Left on main they rot, confuse search results and invite imports into game code. Depth before breadth applies to the repo too: one maintained implementation beats two.

What

As a developer, I want finished spikes moved off main so that the only AI and engine code I find is the maintained code in src/. Tag the current spikes on an archive ref (archive/spikes-2026), delete spikes/engine-*, spikes/determinism, spikes/bench, spikes/shared (ADR-0001, mw-e00.13) and spikes/ai-architecture (ADR-0005, mw-e11.1) from main, and update the ADRs' evidence links to point at the archive ref so the evidence stays reproducible.

Scope

• In: Archive tag/branch, deletion, ADR link updates, the spikes ignore entries in eslint.config.js and .prettierignore (keep them if spikes/ stays as the home for future spikes)
• Out: Future spikes (they follow the same pattern and get their own removal)

Notes

Do it after mw-e11.2 has its own tests for the patrol → investigate → return scenario, so the spike is no longer the only executable evidence for ADR-0005.

• AC-1 manual Given the archive ref, when checked out, then the ADR-0001 and ADR-0005 spike commands in each spike README still run.
• AC-2 unit Given main after the change, when git ls-files spikes/ runs, then it lists only spikes/README.md (or nothing).
• AC-3 manual Given ADR-0001 and ADR-0005, when their evidence links are followed, then every link resolves on the archive ref.
• AC-4 unit Given main after the change, when pnpm lint, typecheck, test:coverage and build run, then they pass with no coverage change.

Why

Automated tests prove the loop works; only people prove it is understandable.

What

Five internal testers play the slice cold with no instructions beyond controls. Observe: time to first attack, whether they find the key, whether death/reload is understood, whether they save voluntarily. Produce docs/playtests/m1-slice.md and beads for every blocker.

Scope

• In: Session plan, observation notes, findings doc, follow-up beads
• Out: Combat fun validation (e35-greybox-combat-test, e04-skeleton-fun-validation)

Notes

Run in the same sessions as e35-greybox-combat-test to save tester time.

• AC-1 manual Given 5 testers, when they play, then >= 4 reach slice complete within 8 minutes without help.
• AC-2 manual Given testers who die, when the death screen appears, then >= 4 of 5 return to play without asking what to do.
• AC-3 content Given the findings doc, when complete, then every issue observed by >= 2 testers has a bead key listed next to it.

Why

Manual AC-4 of mw-e01.5 cannot be automated: readable telegraphs are a playfeel claim.

What

Run in the same sessions as mw-e01.10. Have 3 internal testers play the slice; record whether each can name at least one skeleton attack by its telegraph. Findings go in docs/playtests/m1-slice.md.

Scope

• In: session, per-tester result, follow-up beads for unreadable telegraphs
• Out: automated telegraph tests (mw-e04.20, closed)

• AC-1 manual Given 3 internal testers, when they play the slice, then each can name at least one skeleton attack by its telegraph.
• AC-2 content Given the findings doc, when complete, then each tester's answer is recorded and any failure has a bead key next to it.

Why

M2 must prove each class fantasy in grey-box: "a locked tower isn't simply requires level 8".

What

As a player of any class, I want a grey-box course whose every gate has a route for my class so that my class's core verbs are proven.

Five chambers, each with one gate and four routes: (1) Parry gate: skeleton sentry — knight parry/riposte or smash a weak wall; archer headshot from a gallery; sorcerer Gust it off a ledge; thief sneak/backstab. (2) Rope bridge chasm: rope arrow anchor; Levitate/Blink; climb route; knight kicks a beam across. (3) Webbed door: Ember/fire arrow burns web; knight cuts; thief finds vent. (4) Water channel: Frostglass freeze; float crates; climb beams; blunt-arrow the sluice. (5) Guard room with locked chest and lever: lockpick; shoot the lever; Mage Hand it; shield-bash the lock.

Scope

• In: Scene data for 5 chambers, gate/route wiring, route declarations per class
• Out: The verbs (E04/E05/E08/E10)
• Out: Route audit for the real MVP (E16)

Notes

Each chamber is also a reusable e16 obstacle template. Every route must work via properties (flammable, climbable, shootable), never pairings.

• AC-1 content Given the gauntlet scene, when validated, then every gate declares >= 3 class routes referencing existing capability ids.
• AC-2 integration Given chamber 3's web, when any heat source above its ignition threshold touches it, then it burns away regardless of source (spell, arrow, torch).
• AC-3 integration Given a player breaks the chamber 2 beam before crossing, when the chasm is re-evaluated, then at least one remaining route still exists for every class (no softlock, edge).
• AC-4 perf Given the densest gauntlet chamber with 4 creatures and fire, when measured, then frame time p95 <= 16.7 ms (steady 60 fps) on the reference machine (MacBook Pro M1 Pro, Chrome, High preset, 1440p-equivalent).

Why

Class fantasies must stay proven as systems change; a regression in any class verb should fail CI.

What

Four recorded replays (knight, archer, sorcerer, thief) that complete the gauntlet using only class-granted capabilities, asserted headlessly in CI and via Playwright on Chromium nightly.

Scope

• In: 4 replays, headless CI job, nightly Playwright job, capability-use assertion
• Out: Playtests of class feel (e35-class-fantasy-tests)

Notes

The capability-use assertion reads capability.used events and fails if a replay used a capability outside its class kit.

• AC-1 integration Given each class replay, when run headless, then the gauntlet.complete fact is set and the final hash equals the golden hash.
• AC-2 integration Given the thief replay, when capability.used events are inspected, then no event references a capability not granted by the thief class (edge).
• AC-3 e2e Given the nightly Chromium run, when all four replays execute, then each finishes with zero console errors.

Why

Bible §4.5: dying in the Knot wakes you in the Fold, dying in the Briarwood wakes you at the Sleeping Ox. Death must move the story sideways.

What

As a player, I want to wake somewhere meaningful after I die so that failure becomes part of the story.

Extends the respawn resolver with mode wake-in-place: no reload; the world state is kept; the player is moved to the destination spawn with restored health, set facts (e.g. rescues.byHorn += 1), an optional cost (coin/trinket fee), and a wake text id. Death screen still offers "Load last save" as a secondary option. Rules for Knot → Fold, Knot (Horn dead) → goblin camp or Forgotten laying-out, Briarwood → Sleeping Ox are content in the respective area items.

Scope

• In: wake-in-place mode, fee/cost effects, counters, wake text hook, death-screen option ordering
• Out: Rule content per area (e26-horns-hall-fold, e25-briarwood-greybox)
• Out: Wake texts (e39-failure-narratives)

Notes

Creatures that killed the player return to routine; hostile creatures near the wake point are pushed to Searching elsewhere so the player never wakes into a lethal loop.

• AC-1 unit Given a wake-in-place rule for region knot with fact horn.alive, when the player dies, then the player is placed at the Fold spawn at full health and rescues.byHorn increments by exactly 1.
• AC-2 unit Given a rule with a fee of 20 coins and a player holding 5, when applied, then the fee takes 5 coins and sets fact debt.kettleback = 15 (edge: no negative gold).
• AC-3 integration Given the player wakes in place, when inspected, then no world facts or entity deltas are rolled back other than those in factsToSet.
• AC-4 integration Given a hostile creature within 15 m of the wake spawn, when the wake completes, then that creature is relocated or set to a non-combat state before input resumes (edge).
• AC-5 e2e Given the death screen after a wake-eligible death, when shown, then "Wake" is the default focused option and "Load last save" is available.

Why

When collision pulls the orbit camera (mw-e02.4) in tight, for example behind a pillar in the testbed corridor or with the player backed against a wall, the placeholder capsule can fill half the screen and hide the path ahead. mw-e02.4 scoped a transparency fade of occluders near the camera, but it shipped without one.

What

As a player, I want the character and anything right in front of the camera to fade out when the camera is squeezed in close, so that I can still see where I am going.

Scope

• In: Fade the player's body by camera distance, and fade scene pieces that the camera's near sphere or the camera-to-player line passes through, with a smooth fade in and out.
• Out: Camera placement and collision (done in mw-e02.4).

Notes

src/game/camera/orbit-camera.ts already publishes boom length and ideal distance each frame. The fade is presentation only and never touches the sim.

• AC-1 unit Given the boom pulled in below the fade start distance, when a frame is drawn, then the player's opacity falls monotonically with distance and is fully opaque at or beyond it.
• AC-2 unit Given a scene piece between the camera and the player, when a frame is drawn, then that piece fades out within 0.2 s and fades back in once it is clear.
• AC-3 e2e Given the testbed corridor, when the bot walks past the pillars, then there are no console errors and the frame time stays within the perf budget.

Why

mw-e02.16 shipped lock-on with documented default numbers the bead did not specify (distance weighting, priority weight, turn rate, flick thresholds, camera framing). Feel is the owner's call.

What

As a designer, I want the lock-on feel reviewed in the testbed arena so that the defaults are confirmed or retuned.

Scope

• In: src/content/data/lock-on/player.json values (distanceWeight 0.5, priorityWeight 0.25, turnRate 720°/s, stick flick 0.7/0.3, mouse flick 40/5 counts, framing 0.2 s / 0.5 / −10°).
• Out: rule changes (new beads).

Notes

Open the testbed, walk into the arena, lock on (Q / RS click), cycle (Tab, mouse swipe, RS flick), strafe with A/D.

• AC-1 manual Given the testbed arena with three dummies, when the owner locks, cycles and strafes with keyboard+mouse and with a pad, then each tuning value is confirmed or a new value is recorded in the content file.

Follow-up from mw-e02.18 (PR #313). Push sim (src/sim/push) ships but: (1) friction comes from the object or an optional surface hook, not the surface under the object; (2) strength defaults to 100 for all classes (knight should be stronger, via class data); (3) push.config (axis, grid, rotateStep) cannot be authored in scene data; (4) prompt still reads Push while grabbing; (5) rotation is not swept and can clip walls; (6) PUSH_LOAD_FACTOR=4 was added to the bead formula so stone vs ice differ — confirm or retune. Place at least one pushable crate in the testbed.

AC-1 unit A crate on an ice surface pushes faster than on stone using the surface under it.

AC-2 unit Knight strength from class data exceeds the default and pushes faster.

AC-3 unit Scene data can author axis, grid and rotateStep, validated by schema.

AC-4 unit Rotation that would penetrate a wall is refused.

AC-5 e2e The testbed has a pushable crate the player can grab, push and release.

Why

The lock-on marker and camera framing (mw-e02.16) read the locked target's lock point from the sim's ViewAnchor as of the last tick. Training dummies stand still, but a moving creature would make the marker step at 60 Hz instead of gliding with the interpolated render.

What

As a player, I want the lock ring to stay glued to a moving target so that locked combat reads cleanly at any refresh rate.

Scope

• In: interpolate the target's lock point with RenderSync alpha (or read the drawn transform) for the HUD marker and LockFraming.
• Out: lock-on rules (src/sim/targeting).

Notes

src/game/player/testbed-player.ts lockTarget(); src/game/player/lock-marker.ts.

• AC-1 unit Given a target moving at 5 m/s and a 144 Hz display, when frames are drawn between ticks, then the marker position follows the interpolated lock point (error ≤ 1 px at 1080p).
• AC-2 unit Given a stationary target, when frames are drawn, then the marker writes no DOM changes (idle HUD stays free).

From PR #100: InteractPrompt is not in src/ui/gallery.ts, so gallery axe and text-scale checks don't cover it. Add it.

From PR #100: interaction system applies the Interact button to every interactor entity by default. Fine while only the player interacts; future NPC interactors must supply input via the system's input option.

Why

mw-e02.12 notes that heavy encumbrance may block pull-up (data). Mantling and hangs shipped without it because the sim has no carried weight yet.

What

As a player carrying too much, I want pulling up from a hang to fail (with a cue), so that weight matters on climbs.

Scope

• In: an encumbrance threshold in the controller's ledge tuning (content); the ledge hook refuses the pull-up above it
• Out: the inventory and weight model itself

• AC-1 unit Given a hanging player above the pull-up weight threshold, when jump is pressed, then the player keeps hanging; below it, the pull-up runs.

Why

Found while testing mw-e02.12: in the in-memory collision world, a falling capsule held against a wall whose top edge is within the capsule's height hangs in mid-air (velocity 0, not grounded) forever. The fake's square-edge Minkowski approximation reports the wall's top face (normal +y) for a diagonal down-and-in sweep beside the edge, and the controller's slide then cancels the fall. Rapier rounds edges, so the game is unaffected, but unit tests and goldens on the fake can hit it.

What

As a developer, I want the fake to behave like Rapier at wall top edges, so that controller tests never pin a test-double artifact.

Scope

• In: fix the fake's contact choice at edges (prefer the face the core is beside), plus a contract test run against both worlds
• Out: changes to the Rapier port

• AC-1 unit Given a capsule falling while pushing into a 2 m wall, beside its top edge, when stepped, then it slides down and lands, in both the fake and the Rapier collision world (contract suite).

Why

mw-e02.17 (PR #301) marks shoulder-carried things with a carry.bulky marker component to avoid a property-schema version bump; ADR-0003 names bulky as a world property.

What

Add bulky to the property vocabulary (schema version bump + migration), switch src/sim/carry to read it, drop the marker component.

AC-1 unit Given an entity with the bulky property, when carried, then the carry is a shoulder carry (x0.5 speed, no jump).

Why

Automated tests keep the controller correct; only humans can tell if it feels excellent. Numeric targets need validating against the tone references (Souls weight, Zelda responsiveness).

What

Time-boxed (1 day) playtest of the greybox traversal course with 3+ testers, comparing jump arc, acceleration, camera lag and mantle timing against reference feel notes. Output: tuning changes committed to controller data and a short findings doc.

Scope

• In: Playtest script, questionnaire, tuning deltas
• In: docs/design/controller-feel.md with before/after values
• Out: Combat feel — see mw-e04

Notes

Time-box: 1 day. Record sessions (e34 telemetry optional).

• AC-1 manual Given 3 testers on the traversal course, when surveyed, then median "controls feel responsive" score is ≥ 4/5.
• AC-2 content Given the review, when finished, then all tuning changes are committed to controller data and golden replays are re-blessed with diffs attached.
• AC-3 manual Given any tester reports a specific frustration (e.g. missed mantle), when reviewed, then it is either fixed or filed as a bead with a failing test.

Why

mw-e03.21 signal graphs bind nodes to level entities by name ("entity": "lower-lock-gate"). Nothing checks those names at content load yet because levels are not content; a typo only fails when the graph is placed at runtime.

What

As a designer, I want the content loader to reject a graph whose bindings name entities missing from the level that places it, so that typos fail in CI rather than in play.

Scope

• In: cross-reference check from the level/placement data to signal-graph bindings, with file path and pointer
• Out: the level format — see mw-e01.4 greybox level

Notes

The sim already throws on a missing binding when the graph is placed (addSignalGraph).

• AC-1 content Given a level placing a graph whose node binds an entity id absent from the level, when validated, then validation fails naming the id and the level file.

Why

Since mw-e03.35 a World with physics carries PhysicsState (engine id plus Rapier's snapshot as base64 and a handle table) in its WorldSnapshot, and restore() requires it. The save world section's schema (src/game/save/format/registry.ts) is a strictObject without a physics field, so saving a world that has physics would be rejected, and physics would be lost across save/load.

What

Add the optional physics field to the world section (a new section version with an identity migration, like facts in v2), with round-trip and fixture tests. Consider compressing the Rapier bytes (ADR-0001: about 350 KB per 200 bodies) and whether static level geometry should be rebuilt from the level rather than saved (mw-e27.3 baselines).

Scope

• In: world section schema/version, save fixtures, round-trip test with a physics world
• Out: dynamic bodies (mw-e03.10)

• AC-1 unit Given a World with Rapier physics and a loaded scene, when saved and loaded into a new World, then the state hash is identical and stepping on matches the uninterrupted run.
• AC-2 unit Given a save from before this change (no physics), when loaded into a World without physics, then it loads unchanged.

Why

mw-e03.21 filters volumes by element presence on entities (fire = burning, water = wetness > 0, ice = frozen, charge = charge > 0). A room flooding with gas or a patch of burning oil on the floor lives in the element grid (mw-e03.4), not on entities, so a volume cannot see it yet.

What

As a designer, I want a trigger volume filter that tests element-field cells inside the volume (fire, gas, water, charge above a threshold) so that burning oil or seeping gas can trip a mechanism.

Scope

• In: a field-presence predicate for volume nodes (element + min concentration), content schema + sim evaluation
• Out: the element grid itself — see mw-e03.4

Notes

Reuse shapeBounds to find the cells; keep evaluation order canonical.

• AC-1 unit Given a volume filtered to gas ≥ 0.3, when gas cells above that concentration overlap it, then active is true; below it, false.
• AC-2 content Given a field predicate naming an unknown element, when validated, then validation fails naming the element.

Follow-up from mw-e03.5 (PR #68). Fire rules currently write property changes with source null: a fire lit by the player's torch and spread through three crates carries no attribution, so quests, crime and AI blame (mw-e04.19, mw-e11) cannot tell who burned the barn. Record the heat source that ignited an entity (e.g. the attributed stimulus that raised its temperature, or the burning entity whose flames reached it) in a snapshotted component, and pass it as the source of ignition/burn/burnout property writes and fire events.

AC-1 unit Given a player heat stimulus that ignites crate A whose flames ignite crate B, when both burn, then fireIgnited for B and B's burning propertyChanged carry the player as source.

AC-2 unit Given the attribution state, when the world is snapshotted and restored mid-fire, then attribution survives and replays hash identically.

Why

Shadows under a hedge or behind a curtain should be dimmer, not pitch black or fully lit. Line of sight (mw-e09.1) already treats partial cover as transmittance; the light field (mw-e03.15) only knows opaque blockers.

What

As a thief, I want foliage and cloth to cast soft shadows so that partial cover hides me partly.

Light paths through transparent (glass, water, ice) and partial-cover entities are attenuated by a transmittance factor, using the same occlusion vocabulary as line of sight (opaque | transparent | partial:<factor>), instead of only opaque entities blocking fully.

Scope

• In: transmittance for dynamic light occluders; sharing the occlusion factor source with mw-e09.1
• Out: coloured light and caustics (render only)

Notes

Follow-up from mw-e03.15. Keep the sampling budget (AC-6 of mw-e03.15).

• AC-1 unit Given a partial:0.5 bush between an emitter and a sample, when sampled, then the emitter's contribution is halved.
• AC-2 unit Given a transparent pane between an emitter and a sample, when sampled, then the contribution is unchanged.
• AC-3 perf Given the mw-e03.15 light benchmark, when partial occluders are added, then it stays within budget.

Why

mw-e03.22 extracts ledges once at scene load from static greybox parts. Movable props (physics-object crates from testprops with bodies) and runtime-spawned geometry (propped ladders, debug-console spawns) have no ledges, so a player cannot mantle a pushed crate.

What

Ledges for physics-object boxes derived from their current pose (top edges of upright boxes), refreshed as they move/sleep, exposed through the same LedgeIndex query (LedgeHit) as scene ledges.

Notes

Build on src/sim/climb/ledges.ts (extractLedges, LedgeIndex). Only upright boxes need ledges; tilted ones can be skipped.

• AC-1 unit Given a 1.4 m physics crate at rest, when ledges are queried near its top, then its four top edges are returned with its entity.
• AC-2 unit Given the crate is pushed 2 m, when queried on the next tick, then the ledges follow its new pose.
• AC-3 unit Given the crate is tipped over past 20°, when queried, then it has no ledges (edge).

Why

PR #138 (mw-e03.11) added a breakable profile type with its own fragile threshold and resistances, but the existing world properties toughness (single-hit threshold per kind) and bashable remain unused. Two ways to express the same thing breaks the single property-based model (contract §2).

What

Either derive breakable thresholds from toughness/bashable or remove those properties; also decide whether physics impacts should wear down structural hp (today they only test fragility) and whether break loudness belongs on the material rather than the profile.

• AC-1 unit Given a content entity, when it declares breakability, then exactly one property path determines its break thresholds and the validator rejects the other.

Why

mw-e17.6 shipped thrown oil flasks that land whole and burn where they lie. The bead's intent (a thrown flask shatters and leaves a flammable puddle) needs breakables that spill their liquid contents.

What

Fragile containers break on impact above a threshold and spawn a puddle entity inheriting the container's liquid worldProperties (flammable, wet, conductive...), with a data-driven radius. Purely property-based — no flask-specific code.

Notes

Follow-up from PR #154 (mw-e17.6).

AC-1 unit A fragile container with liquid properties breaking emits a puddle spawn carrying exactly those properties.

AC-2 integration In the greybox testbed a thrown oil flask shatters, leaves a puddle, and igniting it burns the area via e03 propagation.

PR #279 (mw-akb) gives every flammable thing, including wooden floors and walls, an ambient temperature, so Ember can now light level geometry. An idle probe showed no spread, but sustained fire beside wooden architecture is untested: decide whether structural pieces burn, char or are fireproof, and make sure burning one can't remove a required floor/door/path.

• AC-1 integration: burn the wooden pieces in the Sleeping Ox and the slice; required paths stay traversable; result documented as a design rule

Follow-up from mw-e03.20 (PR #322). physicsImpact from thrown/dropped bodies is always blunt, so a thrown dagger cannot trip a pierce-only bullseye. Read a damage type (and optional element) from the thrown item's content data and carry it on the impact.

AC-1 unit A thrown dagger (pierce) at >= 20 J activates a pierce-filtered bullseye.

AC-2 unit A thrown crate (no damage type) stays blunt.

AC-3 unit Hit-zone check optionally respects the switch's facing plane (front hits only) when configured.

mw-e03.28 maps burning/wet/frozen/charged to VFX and mesh looks, and the element-feedback table has an optional audioCue per state, but the sound manifest has no crackle loop, sizzle, freeze crack or zap cue. Add the cues (mw-e38 asset beads) and set audioCue in src/content/data/element-feedback/element-feedback.json. A looping cue (crackle while burning) needs a loop hook in ElementFeedbackBridge; today it plays one-shots on entering a state.

mw-e03.28's bridge reacts to propertyChanged, so an entity that loads burning, wet or frozen from scene data (addProperties fires no event) gets no VFX or look, e.g. lanterns and braziers in briar-glen-lane (their flame is the light rig's). Seed the bridge from a world scan at scene load, skipping entities whose flame the light rig already draws.

mw-e03.28 screenshot review: wet (darker bluish albedo + blue drips) is the weakest of the four states at game camera distance and is easily mistaken for a dark crate. Strengthen it (wet sheen, puddle decal vfx-decal-wet-01, larger drips) and give every state a non-colour cue (shape or motion), coordinated with mw-e31.8.

mw-e03.28 AC-3 and AC-5 refer to a chaos room that does not exist. The e2e uses crates spawned into the testbed. Add a purpose-built scene with spaced props for fire, freeze, wet and charge scenarios and the 6-object identification playtest.

Found in mw-e15.9: sunbeam-hall expresses 'light the sensor by igniting a brazier near it' in the puzzle goal (any of sensor lit, brazier lit) because lightReceiver reads beams only (lightActivated property is declared but unread), and the bead's 'light spell' has no spell content. Let receivers (or a sensor node) read light-field level, and add a light spell.

Found in mw-e15.9: the frozen channel cannot declare 'climb the ceiling beams' because the climb system has ledge hang/shimmy and wall/rope climbing but no overhead traversal. Related open: mw-e10.9 (climb tagged surfaces), mw-e03.14 (buoyancy: crates sink, so 'float crates' is not declarable), mw-e08.5 (no cold source until Frostglass), mw-e08.13 (no self-levitation).

Found in mw-e15.10: the dark crossing was meant to be solvable by following drafts in smoke to the safe path. There is no smoke, draft or airflow field to read or render (wind only blows flames out as a force stimulus). Add an airflow source (vents, gaps) the stimulus API and renderer can read, then declare the route in src/content/data/puzzle/dark-crossing.json.

Found in mw-e15.9: four trapdoors (state open) in a flooded trench, wired to close from a lever, stay 'blocked' with door.blockedBy set to the liquid layer entity (placement radius 2 at the volume centre). A drawbridge over water would be a data-only second route for the frozen channel.

Why

We need living proof that "Wait… I can do that?" moments emerge from properties, and a regression net so they keep emerging as rules change.

What

As a designer, I want a catalogue of emergent combos, each a replayable greybox scenario, so that I can see and protect the game's surprising interactions.

A chaos-room greybox level plus ≥ 12 scenario files, each with setup, input log and asserted outcome expressed via properties/facts. Initial catalogue: oil puddle ignition chain; burn rope → dropped chandelier; freeze river → walk across; soak enemies + charge in puddle; frozen actor shattered by blunt hit; levitated crate weighs 0 on plate; wedge portcullis with crate; gust extinguishes candles but not brazier; smoke blocks sight line; burning haystack ejects hider; focused beam ignites web barrier; ice block on pressure plate melts → plate releases (timer puzzle for free).

Scope

• In: Chaos room level, scenario schema, runner integrated with e00-replay-harness
• In: 12+ scenarios with property-based outcome assertions
• In: Catalogue doc docs/design/emergence-catalogue.md generated from scenario metadata
• Out: Randomised coverage — see e03-interaction-matrix

Notes

Scenario outcomes are phrased as properties ("some entity with flammable ends burnt"), never "the rope entity". New combos discovered in playtests get added here.

• AC-1 integration Given each scenario, when replayed in CI, then the asserted property outcome holds and the final state hash matches its golden.
• AC-2 content Given the scenario directory, when validated, then there are ≥ 12 scenarios and each has setup, inputs, outcome and a player-facing description.
• AC-3 integration Given a rule change breaks a scenario, when CI runs, then the failure names the scenario and its outcome description.
• AC-4 integration Given the ice-on-plate scenario, when simulated, then the plate deactivates only after the ice melts below the weight threshold (edge: time-based emergent state).
• AC-5 e2e Given the chaos room in the browser, when the Playwright bot runs the rope/chandelier scenario, then it completes with no console errors.

Why

Gust clearing poisonous gas, igniting swamp gas, smoke blocking sight lines for a thief escape — gas turns rooms into systemic spaces.

What

As a thief, I want smoke to block vision and flammable gas to explode when lit so that I can escape or set traps.

Gas types in the element field (smoke: blocks vision; steam: obscures, scalds lightly; poison: damages actors; flammable/marsh gas: ignites with flash + force stimulus when concentration ≥ threshold). Gas diffuses, rises/sinks by type, is blocked by walls/doors, vents through openings, and is pushed by force stimuli (Gust).

Scope

• In: Gas species data (buoyancy, diffusion, decay, effects)
• In: Ignition of flammable gas → explosion stimulus
• In: Visibility occlusion query for stealth/AI (density along a ray)
• Out: AI reaction to smoke — see mw-e11
• Out: Gas VFX — see e29-vfx-framework

Notes

Closed doors block gas (door entities are field blockers) — opening a door lets smoke pour out.

• AC-1 unit Given flammable gas at concentration ≥ threshold in a cell and a heat source ≥ ignition, when resolved, then the gas is consumed and a force + heat stimulus is applied with radius scaled by the consumed volume.
• AC-2 unit Given smoke between an observer and a target, when densityAlongRay is queried, then it returns the integrated density and exceeds the vision-block threshold for a 2 m thick cloud.
• AC-3 unit Given a closed door separating two rooms, when gas fills room A, then room B stays at 0 until the door opens, after which concentration equalises over time.
• AC-4 unit Given a force stimulus through a gas cloud, when applied, then gas is displaced along the force direction by the documented amount.
• AC-5 unit Given gas in a fully sealed room with no decay source, when simulated for 30 minutes, then total gas mass is conserved and no cell exceeds max concentration (edge).

Why

Floating crates as stepping stones, sinking a heavy chest to trip a submerged plate, or ice slabs drifting are cheap emergent traversal once density is a property.

What

As a player, I want wooden objects to float and iron to sink so that I can improvise rafts and stepping stones.

Buoyancy force from submerged volume × liquid density vs. object density, with drag; objects adjust to the liquid level; wet/soaked materials may change density (data).

Scope

• In: Buoyancy + drag forces for bodies in liquid volumes
• In: Standing on floating bodies (moving platform)
• In: Density changes from waterlogging (data)
• Out: Currents / flowing rivers — out of scope for MVP

Notes

Ice (density 0.92) floats automatically once frozen slabs break free.

• AC-1 unit Given an object with density 0.6 in water (1.0), when settled, then 60% ± 5% of its volume is submerged.
• AC-2 unit Given an iron object (density 7.8), when placed in water, then it sinks to the bottom and can rest on a submerged pressure plate.
• AC-3 unit Given a floating crate with a 20 kg load exceeding its buoyancy reserve, when loaded, then it sinks (edge).
• AC-4 unit Given the water level drops (drain), when resolved, then floating objects descend with the surface and settle on the floor.

Why

mw-e04.5 AC-6: running out of stamina must be readable without the HUD (souls-like fairness). The sim now emits StaminaExhausted / StaminaRecovered (src/sim/combat/stamina.ts) but nothing presents them yet.

What

As a knight, I want to hear and see that I am out of breath so that I stop swinging before the HUD tells me.

On StaminaExhausted: heavy-breathing loop on the player bus and a hunched/winded idle and slower recovery pose; on StaminaRecovered: breath settles and the loop stops. ActionRejected{reason:"stamina"} plays a short strained grunt (once per press). Placeholder assets via e28 cue sheets until final audio lands.

Scope

• In: cue-sheet entries and game-layer bindings for the three events, placeholder animation state, manual playtest
• Out: HUD stamina bar (mw-e04.10), final SFX/animation assets (audio/art beads)

• AC-1 unit Given a StaminaExhausted event, when the game layer processes it, then the breathing cue starts once and stops on StaminaRecovered.
• AC-2 unit Given ActionRejected{reason:"stamina"}, when processed, then the strain cue plays once.
• AC-3 manual (carried from mw-e04.5 AC-6) Given a 5-minute sandbox session, when a playtester fights the attacker dummy, then they can describe when they are "out of breath" without looking at the HUD.

Why

mw-e04.8 AC-6 is a manual feel check that needs the combat sandbox's attacker dummy (mw-e04.9), which did not exist when the dodge shipped.

What

As a player, I want dodges that feel fair so that timing, not luck, keeps me alive. Once the sandbox exists, three playtesters practise rolling through the attacker dummy's metronome swing for 3 minutes each.

Scope

• In: the playtest, logging counterexamples with replay files
• Out: tuning changes (file separately with evidence)

Notes

Dodge rules: src/sim/combat/dodge; numbers in src/content/data/move/dodge-roll.json and backstep.json. Golden replays of an on-time vs one-tick-early roll: tests/replays/dodge-*.json.

• AC-1 manual Given the sandbox attacker dummy, when three playtesters practise for 3 minutes, then each reports dodges "felt fair" (no hits they believed they had dodged); any counterexample is logged with a replay file.

Why

Found while wiring creature spawning (mw-e12.4): the knight's light attack, swung at a fixture-hound (0.9 m tall, hurtbox capsule 0–0.9 m) standing 1.2 m in front of him, never connects, while the same swing at the 1.8 m fixture-guard at the same distance does. Hounds, rats, goblins and other low creatures must be hittable with the knight's basic attack, or the player has no answer to them.

What

Decide (owner/design) how low creatures are hit: angle the light chain's socket track lower, add a downward/low attack, or auto-aim the swing's height at the locked target. Then make it so.

Scope

• In: knight melee reaching low targets; a test with a short creature.
• Out: creature hurtbox regions per rig (asset beads).

• AC-1 integration Given fixture-hound 1.2 m in front of the knight, when he performs the light attack (locked on or not), then it deals damage to the hound.
• AC-2 integration Given fixture-guard at the same distance, the light attack still hits as today.

Why

mw-e04.8 shipped numbers and controls the bead did not fix; they are documented defaults awaiting the owner's call.

What

Review and confirm or change: backstep travel 1.2 m (backstep.json); roll attack 10/4/20 ticks, 14 stamina, 16 slash, 12 poise (roll-attack.json); keyboard dodge on R (F stays the remap example; Ctrl/Alt trip browser shortcuts); pad dodge on B with crouch moved to D-pad Down and ability 3 on LB only; a swing dodged once cannot hit the dodger later in the same active window; a dodge and an attack pressed on the same tick start the attack; the roll moves only on its 13 active ticks and stands still in recovery.

Scope

• In: decisions, data edits, docs/design/controls.md
• Out: new mechanics

• AC-1 manual Given the list in the description, when the owner has reviewed it, then each item is confirmed or changed in content data/bindings with the decision noted in the move notes or controls doc.

Why

mw-e04.6 AC-8 is a manual playtest (three testers, median ≥ 4/5 for 'swings feel heavy and committed', none say floaty) that needs the combat sandbox (mw-e04.9).

What

Run the 5-minute sandbox session and record ratings.

Scope

• In: the playtest and its notes

Notes

Also review the defaults chosen in mw-e04.6: planted during attacks, facing locked through recovery, chain memory forgotten on interrupt, training dummy 200 hp.

• AC-1 manual Given a 5-minute sandbox session, when three playtesters use the chain, then the median rating for 'swings feel heavy and committed' is ≥ 4/5 and none describe attacks as floaty.

mw-e04.11 AC-5 manual: given A/B builds with hit-stop on vs off, three playtesters try both blind; at least two of three prefer hit-stop on for impact. Hit-stop off = set every tier to 0 in src/content/data/hit-stop/hit-stop.json. Also a chance to tune the PLACEHOLDER tick values.

Why

"The knight might discover that an old wall can simply be smashed apart." Force-based breaking turns the knight's strength into a world verb (Classes change possibilities).

What

As a knight, I want to smash cracked walls, barricades and crates with heavy blows so that brute force opens routes other classes find differently.

Any object with world property breakable and a toughness breaks when a single hit's impactForce (J) reaches its toughness for that kind of hit (query breaksUnder(type, energy), e03 world properties). impactForce = move base × charge multiplier × weapon class (blunt ×1.5). Hits below toughness advance a visible crack stage (max 2) but never break it — a readable hint that more force is needed. Breaking spawns debris physics objects, emits a loud noise event for stealth, can reveal passages, and persists via e27-world-state-store. The rule is property-based: any source dealing impactForce (bash, blunt arrow, gust spell, rolling boulder) works.

Scope

• In: Toughness rule, crack stages, debris spawn, noise event, persistence
• Out: Destruction visuals (e29-destruction-fx)
• Out: Placing weak walls in levels (e26 dungeon)

Notes

Suggested toughness: crate 20, barricade 60, weak masonry wall 110 (charged heavy blunt or equivalent only).

• AC-1 unit Given any breakable object with toughness 110, when a hit with impactForce 120 lands, then it breaks and ObjectBroken is emitted with the instigator.
• AC-2 unit Given the same object, when a hit with impactForce 60 lands, then its crack stage increments (capped at 2) and it does not break.
• AC-3 integration Given a weak wall hiding a passage, when broken, then debris physics objects spawn, the passage collider is traversable and a noise event of loudness ≥ 80 is emitted.
• AC-4 integration Given a broken wall, when the game is saved and reloaded, then the wall remains broken.
• AC-5 unit Given a non-knight source (e.g. a blunt projectile) producing impactForce ≥ toughness, when it hits, then the object breaks identically (no class check).
• AC-6 manual Given a greybox room with one cracked wall, when three fresh playtesters play as knight, then at least two try hitting it within 2 minutes without being told.

Why

Timing windows are the game; golden replays guarantee a refactor never silently shifts a parry by one tick.

What

A set of recorded input logs + expected state hashes in tests/replays/combat/: 3-hit chain; parry first/last/one-late tick; guard break; i-frame boundary (tick 14 vs 15); charged release at 11/12/60/90 ticks; hit-stop stacking; wall recoil. CI runs them on every PR; a failure prints the first divergent tick and the differing component.

Scope

• In: Fixtures, CI wiring, divergence reporter, fixture re-record script with review checklist
• Out: Replay harness itself (e00-replay-harness)

Notes

Re-recording a fixture requires a PR note explaining the intended timing change.

• AC-1 integration Given the combat fixtures, when run in CI, then all replays match their stored hashes.
• AC-2 integration Given a deliberate 1-tick change to the parry window, when the suite runs, then the parry-last-tick fixture fails and reports the first divergent tick and component ParryState.
• AC-3 unit Given src/sim/combat, when coverage runs, then it reports 100% on all metrics.
• AC-4 integration Given a fixture recorded with an older move-data version, when run, then the harness fails with a clear "fixture data version mismatch" message rather than a hash diff.

Why

Constitution: "Fighting a skeleton should be enjoyable even when the player doesn't need XP or loot." This spike proves or disproves that before content scales.

What

Question: with placeholder art/audio, is repeatedly fighting a greybox skeleton (3-move set from e04-creature-attack-hooks) intrinsically enjoyable for the knight with zero XP/loot? Time-box: 2 days. Protocol: ≥ 5 playtesters (internal acceptable), short tutorial, then an arena that spawns a new skeleton on demand; observe, record replays and telemetry (parries, dodges, voluntary re-fights, session length), short survey.

Scope

• In: Test protocol, arena variant of the sandbox, telemetry events, write-up with tuning recommendations
• Out: Implementing the recommendations (follow-up beads)
• Out: External m4 playtests (e35)

Notes

Pass bar: ≥ 4 of 5 testers voluntarily fight ≥ 3 skeletons beyond the required one and median "fun" ≥ 4/5. Failing is a valid outcome and must produce concrete tuning beads. Related: e13 bestiary skeleton, e35 fun validation.

• AC-1 manual Given ≥ 5 playtest sessions, when completed, then docs/playtests/skeleton-fun-01.md records per-tester voluntary re-fight counts, fun rating (1–5) and quotes.
• AC-2 integration Given the arena build, when a session runs, then telemetry events for parry attempts/successes, dodges, deaths and re-fights are exported via e34-telemetry-events.
• AC-3 manual Given the results, when compared to the pass bar (≥ 4/5 testers fight ≥ 3 extra skeletons; median fun ≥ 4), then the write-up states PASS or FAIL explicitly.
• AC-4 manual Given a FAIL or any rating ≤ 2, when the write-up is finished, then at least three tuning beads are proposed with the replay files that motivated them.

Why

mw-e04.19 says fire volumes deal 8 fire/s *and apply burning per creature rules*. The hazard damage shipped (src/sim/combat/environment, 8 fire/s in 250 ms pulses through the damage model), but creatures have no burning status yet: fire rules (mw-e03.5) only ignite world objects with flammable/temperature properties.

What

Give creatures a burning state (duration, damage over time, extinguished by water/rolling) applied when a creature is hit by a hazard whose property is burning, respecting fire immunity (resistance 0 never ignites).

Discovered from mw-e04.19.

• AC-1 unit Given a creature with fire resistance 1 in a fire volume, when a pulse lands, then it gains the burning state per its creature data.
• AC-2 unit Given a creature with fire resistance 0, when it stands in fire, then it never gains burning.
• AC-3 unit Given a burning creature entering water, then burning ends.

Why

mw-e04.12 AC-8 is a manual playtest the agent cannot run: parrying must feel learnable (Knight fantasy: mastery).

What

As the owner, I want three playtesters to practise parrying the combat sandbox attacker dummy (?scene=combat-sandbox, 2 s metronome, parry on 3 / LB, F3 frame data shows the window, the Parried stun and 'riposte ready') so that the 10-tick window and the placeholder riposte tuning are validated.

Also judge these placeholder decisions from mw-e04.12: (1) parry bound to ability 3 (LB / 3) while block holds LT / right click; (2) a successful parry still plays out its 30 ticks before the knight can riposte (no early exit); (3) the riposte's 8/4/24 frames, 2.0 m thrust and critical-tier hit-stop (src/content/data/move/sword-riposte.json).

Scope

• In: playtest, tuning notes back into move data
• Out: creature attacks through the executor are not yet tagged parryable (they route through hit volumes in mw-e04.20)

• AC-1 manual Given the sandbox attacker dummy with a 2 s metronome, when three playtesters practise for 5 minutes, then each lands >= 5 parries and rates the window 'learnable' >= 4/5 (carried over from mw-e04.12 AC-8).

mw-e04.11 freezes the action timeline, its hitbox sweeps and hit-reaction clocks. Not yet frozen: a knocked-back/launched victim keeps flying via the character controller during the freeze, and attacks run by the creature attack executor (src/sim/combat/attacks) have their own clock and ignore hit-stop. Decide whether they should follow local time (isHitStopped) and implement.

Why

Accessibility: more players should reach the joy of a parry without removing the skill ceiling for others.

What

As a player, I want optional combat assists so that timing-heavy combat is accessible to me.

Options via e31-settings-framework: parry window ×1.0/×1.5/×2.0 (extends the end tick), i-frames +0/+3/+6 ticks, stamina costs ×1.0/×0.75, extended creature telegraph cue (+10 ticks warning). Values are injected into the sim at session start and recorded in the replay header for determinism.

Scope

• In: Assist options, sim injection, replay header recording, settings UI entries
• Out: Difficulty levels / enemy damage scaling (later if needed)

Notes

Timing convention: sim runs at a fixed 60 Hz (1 tick = 16.67 ms, per e00-sim-core). All frame numbers are sim ticks; ms values are for readability only.

• AC-1 unit Given parry assist ×2.0, when parry is pressed on tick 0, then hits on ticks 4–23 are parried (window 20 ticks).
• AC-2 unit Given i-frame assist +3, when a medium-load roll executes, then i-frames cover ticks 2–17.
• AC-3 integration Given a replay recorded with assists on, when replayed on a machine with assists off, then the header values are used and hashes match.
• AC-4 unit Given assists changed mid-session, when applied, then they take effect at the next safe point (no action in progress) and an AssistsChanged event is logged.

Why

Archery must feel excellent from the first shot: a satisfying draw, precise aim and immediate, physical release (Archer fantasy: precision).

What

As an archer, I want to draw, aim and release arrows so that I can hit precise targets at range.

Hold fire to draw: full draw at 48 ticks (800 ms, shortbow). Launch speed lerps 30%→100% of 60 m/s by draw fraction. Releasing before 12 ticks cancels the shot and returns the arrow. Drawing costs 6 stamina; holding full draw beyond 120 ticks drains 8 stamina/s and at 0 the draw collapses (arrow returned). Aim mode uses the over-shoulder camera from e02-camera (FOV 70→55°), walk speed 50% while drawn. Consumes the selected arrow type from the quiver (e17-inventory-core); a cycle input switches type. Dodge or block cancels the draw and returns the arrow.

Scope

• In: Draw/release/cancel, launch speed curve, stamina, aim mode hookup, quiver selection, bow data (shortbow)
• Out: Sweet-spot release (e05-draw-timing)
• Out: Sway and reticle (e05-aim-sway)
• Out: Ballistics (e05-arrow-projectile)

Notes

Anchor e05-bow-core: with e05-arrow-projectile (a dependency) this completes draw, aim, release and arrow physics.

Timing convention: sim runs at a fixed 60 Hz (1 tick = 16.67 ms, per e00-sim-core). All frame numbers are sim ticks; ms values are for readability only.

• AC-1 unit Given draw held for 24 ticks, when released, then an arrow spawns at 65% of full launch speed (30% + 0.5 × 70%).
• AC-2 unit Given draw held for 8 ticks, when released, then no arrow spawns and the quiver count is unchanged.
• AC-3 unit Given full draw held for 120 ticks, when held further, then stamina drains 8/s and at 0 the draw collapses with the arrow returned.
• AC-4 unit Given an empty quiver for the selected type, when fire is pressed, then the draw does not start and ActionRejected{reason:"ammo"} is emitted.
• AC-5 unit Given a drawn bow, when dodge is pressed, then the draw cancels, the arrow returns to the quiver and the dodge executes.
• AC-6 manual Given a 5-minute archery range session, when three playtesters shoot, then median rating for "the bow feels satisfying" is ≥ 4/5.

Why

Draw timing gives archery a mastery curve comparable to the knight's parry (Archer fantasy: draw timing).

What

As an archer, I want a brief sweet spot at full draw so that well-timed releases are rewarded.

Focused release: releasing within 6 ticks (100 ms) after reaching full draw → +25% damage, +15% launch speed, zero sway, FocusedShot event (reticle pulse + chime cue). Rhythm: consecutive focused releases reduce nock time from 20 to 14 ticks, stacking up to 3 times; any unfocused shot or 3 s idle resets stacks.

Scope

• In: Focused window, bonuses, rhythm stacks, events for presentation
• Out: Reticle visuals (e05-aim-sway)
• Out: Audio chime (e28-bow-audio)

Notes

Timing convention: sim runs at a fixed 60 Hz (1 tick = 16.67 ms, per e00-sim-core). All frame numbers are sim ticks; ms values are for readability only.

• AC-1 unit Given full draw reached on tick 48, when released on tick 48 or tick 53, then the shot is focused; on tick 47 or tick 54, then it is not.
• AC-2 unit Given a focused release, when the arrow spawns, then its damage template is ×1.25 and launch speed ×1.15.
• AC-3 unit Given three consecutive focused shots, when the fourth arrow is nocked, then nock time is 14 ticks; after one unfocused shot, then it returns to 20.
• AC-4 unit Given 2 rhythm stacks and 3 s without firing, when the next draw starts, then stacks are 0.
• AC-5 manual Given the archery range, when three playtesters shoot for 5 minutes, then each can hit the focused window on ≥ 50% of shots by the end and describes it as "satisfying".

Why

Sway and breath control make long shots a skill and keep the arrow honest to what the reticle shows.

What

As an archer, I want my aim to steady or waver based on what I am doing so that long shots demand composure.

Sway is simulated (seeded from sim RNG, so replays match): 0° for the first 60 ticks at full draw, then grows 0.1°/s up to 1.5°; walking adds 0.6°; being hit adds a 2° kick decaying over 300 ms; crouched & stationary ("braced") −40%. Hold-breath input at full draw zeroes sway for up to 3 s at 10 stamina/s. Launch direction = reticle direction + current sway offset. Reticle HUD shows draw progress and sway; optional trajectory arc preview (assist, off by default) via settings.

Scope

• In: Sway model, breath hold, braced modifier, reticle HUD, arc preview assist
• Out: Camera aim mode (e02-camera)
• Out: Lead indicator (e05-moving-targets)

Notes

Timing convention: sim runs at a fixed 60 Hz (1 tick = 16.67 ms, per e00-sim-core). All frame numbers are sim ticks; ms values are for readability only.

• AC-1 unit Given full draw held for 60 ticks while stationary, when released, then sway offset is exactly 0°.
• AC-2 unit Given full draw held 20 s, when sampled, then sway amplitude is capped at 1.5° (0.9° while braced).
• AC-3 unit Given breath hold active with 25 stamina, when 3 s pass, then sway stays 0 for 2.5 s and resumes when stamina hits 0.
• AC-4 integration Given the same seed and inputs, when two runs fire 10 swaying shots, then launch directions match bit-for-bit.
• AC-5 unit Given the arc preview assist disabled (default), when drawing, then no arc renders; enabled, then the arc matches e05-arrow-projectile's prediction within 0.2 m at 30 m.

Why

Setting a distant oil slick or rope alight is a classic archer "wait, I can do that?" — and must emerge from properties, not scripts.

What

As an archer, I want fire arrows and the ability to light any arrow in a flame so that I can ignite things at range.

Fire arrows carry the burning property in flight; on impact they apply burning in a 0.3 m radius through e03-element-propagation — any flammable object or creature there reacts per its own rules. Dip to light: while drawing, if the arrow tip is within 0.8 m of any burning entity for 30 ticks, the nocked arrow becomes burning. Burning arrows passing through a water volume or heavy rain are extinguished.

Scope

• In: Fire payload behaviour, dip-to-light rule, extinguish in water
• Out: Fire spread rules (e03-element-propagation)
• Out: Fire visuals (e29-arrow-fx)

Notes

Timing convention: sim runs at a fixed 60 Hz (1 tick = 16.67 ms, per e00-sim-core). All frame numbers are sim ticks; ms values are for readability only.

• AC-1 integration Given any flammable object, when a burning arrow impacts within 0.3 m, then the object gains burning via element propagation on that tick.
• AC-2 integration Given a non-flammable stone wall, when a burning arrow impacts, then nothing ignites and the arrow burns out after 3 s.
• AC-3 unit Given a drawn standard arrow held 0.6 m from a burning brazier for 30 ticks, when released, then the arrow carries burning; at 29 ticks, then it does not.
• AC-4 integration Given a burning arrow that passes through a water volume, when it exits, then it no longer carries burning and ignites nothing on impact.
• AC-5 manual Given a greybox room with a brazier and an oil-soaked rope bridge, when three fresh archer playtesters play, then at least one discovers dip-to-light unprompted within 5 minutes.

Why

Dousing a torch from the shadows serves stealth, puzzles and fire-fighting through one property (multiple solutions).

What

As an archer, I want water arrows that shatter into a splash so that I can put out fires and lights and make surfaces wet.

On impact the arrow shatters and applies wet in a 1.5 m radius and removes burning from anything inside it (torches, braziers, burning creatures). Extinguished light sources lower the local light level via e03-light-field. Wet surfaces then behave per element rules (e.g. conduct shock). Hitting extreme heat produces whatever e03 rules dictate (e.g. steam). Water arrows never stick.

Scope

• In: Water payload, extinguish, wet application
• Out: Light-level recomputation (e03-light-field)
• Out: Steam/conduction rules (e03-element-propagation)

Notes

Related: stealth group — dousing lights increases thief/archer concealment.

• AC-1 integration Given any burning entity within 1.5 m of impact, when the water arrow shatters, then burning is removed from it that tick.
• AC-2 integration Given a lit torch providing light, when extinguished by a water arrow, then the light level queried at the torch's position decreases.
• AC-3 integration Given a wet floor and a shock source per e03 rules, when shock is applied, then conduction occurs across the wetted area (property-based, not arrow-specific).
• AC-4 unit Given a water arrow, when it hits anything, then it never sticks and is not retrievable.

Why

Opening a vertical route with a rope arrow is the archer's answer to the locked tower (Classes change possibilities).

What

As an archer, I want to shoot a rope arrow into a wooden beam and climb it so that I can reach places others cannot.

Rope arrows stick only into surfaces tagged softAnchor (the wood and earth material presets) whose normal is within 60° of horizontal or facing downward (beams, ceilings). On stone/metal they bounce off with sparks (readable hint). A hit spawns an 8 m rope entity hanging from the anchor, climbable via e02-player-controller climb hooks. Ropes are flammable (burning rope falls and drops climbers) and cuttable. Max 3 active ropes per area (oldest retracts).

Scope

• In: Anchor rule, rope spawn, climb hookup, rope properties, active cap, retrieval at anchor
• Out: Climbing mechanics (e02-player-controller)
• Out: Rope visuals (e29-arrow-fx)

Notes

Ropes persist while the area is loaded; persistence across saves is not required for MVP.

• AC-1 integration Given a wooden beam overhead, when a rope arrow hits it, then an 8 m rope spawns and the player can climb to the beam.
• AC-2 integration Given a stone wall, when a rope arrow hits it, then it bounces, no rope spawns and the arrow remains retrievable.
• AC-3 integration Given a player climbing a rope, when the rope gains burning, then within 3 s it detaches and the player falls (environment fall rules apply).
• AC-4 unit Given 3 active ropes, when a 4th is created, then the oldest retracts and returns a rope arrow at its anchor position.
• AC-5 unit Given a wooden surface whose normal points 80° upward (a floor), when a rope arrow hits it, then no rope spawns.

Why

Luring a guard away with a rattling arrow is a clean, systemic distraction shared with the stealth model.

What

As an archer, I want a noise arrow that rattles where it lands so that I can pull creatures toward a spot.

On impact it emits a noise event (loudness 70, duration 2 s) into e09-sound-propagation at the impact point; how creatures react is entirely their perception/AI. In water the loudness is ×0.3. Deals no damage (blunt head, 1 blunt).

Scope

• In: Noise payload emission, water muffling, positional cue event
• Out: Noise propagation and creature reaction (stealth group)

Notes

Timing convention: sim runs at a fixed 60 Hz (1 tick = 16.67 ms, per e00-sim-core). All frame numbers are sim ticks; ms values are for readability only.

• AC-1 integration Given a noise arrow impact, when resolved, then exactly one noise event of loudness 70 lasting 120 ticks is emitted at the impact position.
• AC-2 integration Given impact inside a water volume, when resolved, then the noise event loudness is 21.
• AC-3 integration Given a creature within hearing range per e09 rules, when the noise event propagates, then the creature receives it (reaction owned by AI; test asserts delivery only).
• AC-4 unit Given a noise arrow hitting a creature, when resolved, then it deals ≤ 1 damage and causes no reaction beyond flinch rules.

Why

Non-lethal, physics-first arrows let the archer solve problems by shoving the world rather than killing it.

What

As an archer, I want blunt arrows that hit hard without sticking so that I can knock things over, press switches at range and stun unaware foes.

Blunt damage, penetration 0, poise damage ×2.5, impactForceMul 2.5, never sticks (drops at impact; always retrievable unless it lands in a hazard). Applies impulse to physics objects (knock a bucket off a shelf). A head hit on an unaware creature whose schema allows knockout renders it unconscious for 60 s (non-lethal).

Scope

• In: Blunt behaviour, impulse, knockout rule
• Out: Switch/mechanism triggering (e05-shoot-mechanisms)
• Out: Unconscious state handling in AI (e11)

Notes

Timing convention: sim runs at a fixed 60 Hz (1 tick = 16.67 ms, per e00-sim-core). All frame numbers are sim ticks; ms values are for readability only.

• AC-1 integration Given a 3 kg object on a shelf, when hit by a blunt arrow, then it receives an impulse and falls.
• AC-2 unit Given an unaware creature with knockout: true, when hit in the head by a blunt arrow, then it becomes unconscious for 3600 ticks.
• AC-3 unit Given an aware creature, when hit in the head by a blunt arrow, then it is not knocked out and takes blunt damage with ×2.5 poise damage.
• AC-4 unit Given a blunt arrow impact on any surface, when resolved, then it does not stick and is retrievable at the impact point.

Why

Draw timing, drop and trick arrows need a fast, repeatable test space just as melee has the combat sandbox.

What

As a developer, I want an archery range scene so that I can tune ballistics and verify trick arrows quickly.

Static targets at 10/20/40/60 m, targets on rails at 2/4/6 m/s, a 6 m elevated platform, an overhead wooden beam (rope anchor), a brazier, a shootable bell wired to a door, a cuttable rope holding a crate, a pool of water. Stats overlay: hit %, flight time, impact speed, focused-shot rate.

Scope

• In: Scene, fixtures, stats overlay, deterministic seed
• Out: Final art for the range (none planned)

Notes

Lives alongside the combat sandbox under /dev/archery.

• AC-1 e2e Given the range loaded in Chromium, when it boots, then all fixtures exist, no console errors occur and the overlay reads 0 shots.
• AC-2 integration Given the rail targets, when 10 s elapse, then their positions match the deterministic schedule within 0.01 m.
• AC-3 integration Given 10 scripted shots at the 20 m target, when complete, then the overlay hit % equals hits/shots from sim events.
• AC-4 e2e Given the bell/door fixture, when an arrow hits the bell, then the door opens within 1 s.
• AC-5 e2e Given the archery range in Chromium, when Playwright draws for 48 ticks and releases at the 20 m target, then the target registers a hit and no console errors occur. (Moved from mw-e05.3 AC-6 by owner decision 2026-10-02.)

Why

High ground should matter to an archer: climbing a tower should change the fight (Archer fantasy: elevation).

What

As an archer, I want shooting from above to hit harder and reach farther so that seeking high ground is a real tactic.

Range gain is natural from ballistics. Additionally: shooter at ≥ 3 m above target → +15% damage; ≥ 6 m → +25% (data-tunable). Implemented as an attacker-stage modifier in the damage pipeline tagged elevation.

Scope

• In: Elevation modifier, thresholds in data
• Out: AI reaction to unreachable attackers (e11)

Notes

Related: AI group should give melee creatures a behaviour for unreachable archers (seek path, throw rocks, take cover).

• AC-1 unit Given a shooter 4 m above the target, when a 20-pierce arrow hits, then 23 damage applies and the hit is tagged elevation.
• AC-2 unit Given a shooter 2.9 m above, when an arrow hits, then no elevation bonus applies.
• AC-3 unit Given a 30%-draw shot from 10 m height versus ground level, when simulated, then horizontal range increases by ≥ 20%.
• AC-4 unit Given a shooter below the target, when an arrow hits, then no bonus or penalty applies.

Why

Hitting a running target should be hard and rewarding (Archer fantasy: moving targets).

What

As an archer, I want well-led shots on moving creatures to trip or ground them so that tracking a target is worth mastering.

A limb (leg) hit on a creature moving > 3 m/s causes a stumble (60-tick knockdown for bipeds; quadrupeds slowed 50% for 3 s). A wing region hit on a flying creature grounds it for 4 s. Optional lead-indicator assist (off by default): shows the intercept point for the lock-on target from its velocity and arrow flight time (3-iteration solver).

Scope

• In: Moving-limb stumble rule, wing grounding, lead solver + indicator assist
• Out: Flying locomotion (e12)

Notes

Timing convention: sim runs at a fixed 60 Hz (1 tick = 16.67 ms, per e00-sim-core). All frame numbers are sim ticks; ms values are for readability only.

• AC-1 unit Given a biped moving at 4 m/s, when a leg region is hit, then it is knocked down for 60 ticks; moving at 2 m/s, then normal hit reactions apply.
• AC-2 unit Given a flying creature, when its wing region is hit, then it is grounded for 240 ticks.
• AC-3 unit Given a target moving at a constant 5 m/s at 30 m, when the lead solver runs, then its predicted intercept is within 0.3 m of the simulated arrow impact.
• AC-4 unit Given the lead assist disabled (default), when locked on, then no indicator renders.

Why

The archer's fantasy is precision and movement; they should reposition without losing a draw.

What

As an archer, I want to sidestep while holding a draw and shoot mid-jump so that I can stay mobile under pressure.

Quickstep: 12-tick sidestep with i-frames ticks 2–7, 15 stamina, retains up to 60% of current draw. Aerial shot: firing while airborne is allowed with +1.0° additional spread; draw progress continues in the air.

Scope

• In: Quickstep move, draw retention, aerial shot spread
• Out: Knight roll (e04-dodge-iframes)

Notes

Timing convention: sim runs at a fixed 60 Hz (1 tick = 16.67 ms, per e00-sim-core). All frame numbers are sim ticks; ms values are for readability only.

• AC-1 unit Given full draw, when quickstep executes, then draw fraction after it is 0.6 and i-frames cover ticks 2–7.
• AC-2 unit Given 40% draw, when quickstep executes, then draw fraction remains 0.4 (retention caps at 60%, never increases).
• AC-3 unit Given the player airborne, when an arrow is released, then its spread includes +1.0°.
• AC-4 manual Given the range with rail targets, when three playtesters use quickstep, then median rating for "I feel mobile as an archer" is ≥ 4/5.

Why

Creature projectiles (mw-e12.5) fly in straight lines and pass through walls (the executor says level geometry does not stop them yet), while arrows (mw-e05.2) arc, collide with the level and use region-tagged hurtboxes. Two projectile models drift apart; a skeleton archer's arrow should behave like the player's.

What

Route projectile attacks through the arrow system: the attack's projectile data names an arrow content id (or a drag-free projectile body), runActive looses it with fireArrow (shooter = attacker), and hits go through the hurtbox query and CollisionWorld like any arrow. Keep AttackProjectileLaunched/AttackHit for AI and cues. Retire the straight-line projectileSystem once no content uses it.

Scope

• In: executor projectile launch via fireArrow, attack schema field, migration of shipped projectile attacks, tests.
• Out: creature bow animation, AI aiming with lead.

Discovered from mw-e05.2.

• AC-1 unit Given a creature projectile attack at a target behind a wall, when fired, then the projectile stops at the wall and deals no damage.
• AC-2 unit Given a projectile attack at a target in range, when it hits, then AttackHit is emitted with the damage results of the struck hurtbox's region.
• AC-3 integration Given the creature attack replays, when re-recorded, then only content hashes change or the change is explained.

Why

Proves the systemic promise: new arrows need data, not code, and immediately interact with existing world properties.

What

As a designer, I want to add frost, shock and smoke arrows purely in data so that the arrow roster grows without engineering.

Frost applies frozen (e.g. water surfaces become walkable ice per e03 rules); shock applies electrified (conducts through wet/conductive); smoke spawns a vision-blocking volume. No new sim code beyond what payload types already support.

Scope

• In: Three arrow data entries, tests exercising each
• Out: New payload primitives (would need a separate engineering bead)

Notes

If any arrow needs code, split that capability into its own bead first.

• AC-1 content Given the three new arrow entries, when validated, then they pass with no schema changes.
• AC-2 integration Given a water surface, when a frost arrow impacts it, then it gains frozen per e03 rules.
• AC-3 integration Given a smoke arrow impact, when resolved, then a vision-blocking volume exists for its declared duration and blocks sight queries through it.
• AC-4 unit Given the change set for this bead, when reviewed, then no files under src/sim are modified (asserted by a CI path check on the PR).

Why

Battlefield manipulation needs areas, cones and beams, and beams that reflect enable light puzzles (puzzle constitution).

What

As a sorcerer, I want sphere, cone, line/wall and ground-decal volumes plus continuous beams so that shockwaves, fire walls and light rays are all data.

Volumes: instant or persistent with tick rate, falloff curve, line-of-sight check option. Beams: raycast each tick, length, width, reflect off objects tagged reflective up to N bounces, blocked by opaque solids.

Scope

• In: volume shapes, persistence, tick, falloff, LOS option; beam raycast, reflection, max bounces
• Out: stimulus semantics (e06-world-stimulus-emitters)
• Out: specific spells (mw-e08)

Notes

Beam reflection uses the reflective property only — any mirror, polished shield or ice surface tagged reflective works.

• AC-1 unit Given a 4 m sphere with linear falloff, when targets stand at 0, 2 and 4.1 m, then they receive 100%, 50% and 0% magnitude.
• AC-2 unit Given a 60° cone of range 6 m, when a target is at 35° off-axis, then it is not affected; at 25°, it is.
• AC-3 unit Given a persistent volume with tick 500 ms and duration 3 s, when it expires, then exactly 6 ticks were applied.
• AC-4 unit Given a beam hitting an object tagged reflective, when maxBounces=2, then the beam reflects by the surface normal and stops after the second reflection.
• AC-5 unit Given an AoE with requiresLOS, when a target is fully behind an opaque wall, then it receives no effect.
• AC-6 perf Given 8 persistent volumes and 4 beams active in the stress scene, when simulated, then magic sim cost p95 ≤ 1.0 ms per tick on the reference machine.

Why

Summoning (Animate Bones, Mirror Self decoys, conjured objects) is a qualitatively different verb from damage.

What

As a sorcerer, I want summon ops that spawn a creature or object archetype for a duration, optionally consuming nearby world material, so that summons feel grounded in the world.

Parameters: archetype id (from creature schema), duration, allegiance (caster/neutral), maxActive per caster, requiresMaterial {property tag, radius, count} (e.g. remains), on-expire ops, command hooks (follow/stay/attack).

Scope

• In: summon op, material consumption by property tag, lifetime, cap enforcement (oldest dismissed), allegiance faction assignment
• Out: summon AI behaviours (E11/E12 — Related: follow/guard behaviours)
• Out: specific summons (mw-e08)

Notes

Material requirement uses property tags such as remains, not specific corpse ids.

• AC-1 unit Given maxActive=2 and two active summons, when a third is summoned, then the oldest is dismissed with its on-expire ops and the new one spawns.
• AC-2 unit Given requiresMaterial {tag: remains, radius: 6, count: 1} and no remains within 6 m, when cast, then the cast fails with reason "no-material", emits a SpellFizzled event and refunds 100% mana.
• AC-3 unit Given remains within range, when the summon succeeds, then the consumed remains entity is removed and the summon spawns at its position.
• AC-4 unit Given duration 30 s, when 30 s pass, then the summon is removed and on-expire ops apply at its location.
• AC-5 integration Given a summon with allegiance caster, when hostile creatures perceive it, then they treat it as hostile per faction rules.

Why

Polymorph-style effects are iconic "wait, I can do that?" magic and must restore state safely.

What

As a sorcerer, I want a transform op that swaps a creature's archetype (body, senses, attacks, properties) for a duration and restores it exactly so that turning a guard into a chicken is safe and systemic.

Preserves identity, health ratio, inventory, quest/dialogue state, alert memory; the new form uses its own senses/locomotion/properties (a transformed creature that is now flammable burns).

Scope

• In: transform op, snapshot/restore, health ratio mapping, size-safe restore (find free space), resist rules via creature tags
• Out: player self-transformation (post-MVP, Related: future spell)
• Out: specific forms (mw-e08 Polymorph)

Notes

Restore must handle the form being killed (original dies? or reverts with 1 HP — configurable per spell) and not enough space to restore (push to nearest free space).

• AC-1 unit Given a creature at 60% health transformed into a form with different max HP, when transformed and restored, then health ratio is 60% in both forms (minus damage taken).
• AC-2 unit Given a transformed creature, when the duration ends, then inventory, faction, dialogue flags and identity id are byte-identical to the pre-transform snapshot.
• AC-3 unit Given a creature tagged immune-transform, when targeted, then the op fails with reason "resisted" and emits SpellResisted.
• AC-4 unit Given the transformed form is killed and the spell sets onFormDeath=revert, when killed, then the original reverts with 1 HP and a stagger status.
• AC-5 unit Given insufficient space to restore a large creature, when reverting, then it is placed at the nearest free position within 3 m, or the revert is deferred until space exists.
• AC-6 integration Given a save made while a creature is transformed, when loaded, then the transform remaining duration and snapshot restore correctly.

Why

Teleportation opens traversal routes (classes change possibilities) and must never put the player inside walls.

What

As a sorcerer, I want displace ops (teleport-to-point, swap-with-target, short dash) with safe placement so that Blink-like spells are reliable and exploitable.

Validates destination: capsule fits, not inside solid geometry, optional line-of-sight requirement, ledge/floor snapping, falls through if no floor (story!).

Scope

• In: displace op handler, capsule fit test, LOS option, nearest-valid fallback search, swap semantics
• Out: specific spells (mw-e08 Blink)

Notes

Teleporting over a chasm with no floor intentionally drops the caster — failure creates stories.

• AC-1 unit Given a teleport target inside a wall, when resolved, then the caster lands at the nearest valid point along the aim ray before the wall, never inside geometry (1000 seeded fuzz trials).
• AC-2 unit Given requiresLOS and a target point behind a closed door, when cast, then it fails with reason "no-line-of-sight".
• AC-3 unit Given a swap-with-target op on a creature too large to fit at the caster's position, when cast, then the swap is rejected with reason "no-space" and mana is refunded.
• AC-4 unit Given a teleport destination with no floor within 1 m, when cast, then the caster arrives at the destination and begins falling.

Why

Failure creates stories: a botched spell should produce a new situation (smoke, a wild blink, a sudden frost) instead of a dull "fizzle".

What

As a sorcerer, I want interrupted, over-cast or poorly-understood spells to sometimes misfire in school-flavoured ways so that mistakes are memorable.

Miscast triggers: over-cast (mana deficit), late interruption, low comprehension (from E07), casting while a status like drenched/charged conflicts. A weighted miscast table per school (and optional per spell) picks outcomes composed from existing ops: backfire on self at reduced magnitude, random-direction delivery, harmless cosmetic, loud noise event, stimulus burst at caster (e.g. frost puddle, smoke).

Scope

• In: miscast request handling, severity calculation, weighted tables in content, seeded selection, MiscastOccurred event
• Out: comprehension model (e07-comprehension-study)
• Out: VFX for miscasts (placeholder cue per school)

Notes

Miscast outcomes must never softlock or instakill: cap self-damage at 25% max HP. Every outcome reuses existing ops so miscasts interact with the world too (a fire miscast can ignite the bookshelf you are standing next to).

• AC-1 unit Given a miscast request with severity 0.1, when resolved with a fixed seed, then the outcome is drawn from the table's low-severity band and is deterministic across runs.
• AC-2 unit Given any miscast outcome, when applied to a caster at full HP, then self-damage never exceeds 25% max HP (property test over all tables).
• AC-3 integration Given a Fire-school miscast producing a heat burst at the caster, when a flammable object is within the burst radius, then it ignites via E03 rules.
• AC-4 content Given a miscast table whose weights sum to 0 or referencing an unknown op, when validated, then validation fails.
• AC-5 unit Given a miscast outcome "loud", when applied, then a noise event of the configured loudness is emitted for stealth/AI.
• AC-6 manual Given a playtest of 20 deliberate over-casts, when observed, then at least 3 distinct outcome types occur and testers describe them as funny or interesting rather than punishing.

Why

Magic must look and sound magical even in grey-box; placeholders unblock gameplay until the art pass (gameplay is never blocked by final art).

What

As a developer, I want spell events (wind-up, release, travel, impact, persistent, miscast) to trigger named VFX and SFX cues so that swapping placeholders for final assets is a data change.

Cue registry maps cue id → placeholder effect (coloured particles per school, primitive meshes) and placeholder sound; missing cue logs a warning and plays a school default.

Scope

• In: cue registry, event→cue binding in src/game, school-coloured placeholder particle presets, placeholder SFX hooks
• Out: final VFX (E29 asset beads)
• Out: final SFX (E28/E38 asset beads)

Notes

Cue naming follows asset ID convention: vfx-spell-<spell>-<phase>, sfx-spell-<spell>-<phase>.

• AC-1 integration Given a spell impact event with vfxCue "vfx-spell-ember-impact", when emitted, then the VFX system receives a spawn request for that cue at the impact position.
• AC-2 integration Given a cue id with no registered asset, when triggered, then the school default placeholder plays and one warning is logged per cue id (not per trigger).
• AC-3 content Given every spell in the catalogue, when validated, then each referenced cue id matches the naming convention.
• AC-4 e2e Given the greybox testbed, when every catalogue spell is cast in sequence, then no console errors occur and a screenshot sanity check detects non-background pixels at each impact.

Why

Save-game stability is a technical pillar; a fire wall or summon must survive save/load exactly.

What

Serialize mana, cooldowns, in-progress channel state (cancelled on load), persistent volumes, summons and statuses into the versioned save format with round-trip and migration tests.

Scope

• In: magic save section, round-trip tests, migration test scaffold
• Out: known spells and study progress (e07-learning-save-state)
• Out: transform snapshots (e06-transformation-framework)

Notes

In-flight projectiles are dropped on save (documented) unless trivial to keep.

• AC-1 unit Given a sim with 2 persistent volumes, 1 summon and 3 statuses, when saved and loaded, then state hashes of the magic subsystem match.
• AC-2 unit Given a save while channeling, when loaded, then the channel is cancelled and no mana is lost beyond ticks already applied.
• AC-3 unit Given a save from magic schema v1, when loaded by v2 code, then the migration test passes.

Split from mw-e06.5 (its AC-2) because the freeze rule is mw-e03.7, not yet built. Once e03.7 lands, add the integration test through the spell pipeline.

AC-1 integration Given any wet or waterSurface tile inside a spell cold stimulus >= its freeze threshold, when applied, then it gains the frozen property for the stimulus duration plus melt time. Generic property-tagged fixtures only (no named props).

Split from mw-e06.5 (its AC-3) because conduction is mw-e03.8, not yet built. Once e03.8 lands, add the integration test through the spell pipeline.

AC-1 integration Given a spell charge stimulus hitting a conductive object that touches other conductive or wet objects, when applied, then charge propagates to the connected objects per E03 rules. Generic property-tagged fixtures only.

Why

Enemy mages should obey the same rules as the player so that interrupts, miscasts and world interactions apply to them too.

What

As a designer, I want creature definitions to list spells with usage hints so that goblin shamans and liches cast via the standard pipeline.

Exposes a cast-request API for AI; creature mana optional; AI can be interrupted and miscast.

Scope

• In: creature schema extension for spells (via E12 hooks), AI cast API, tests
• Out: AI decision-making about when to cast (E11 — Related: combat behaviours)

Notes

Player-learnable spells dropped as books by enemy casters is handled in e07-book-acquisition-hooks.

• AC-1 unit Given a creature with a fixture projectile spell, when AI requests a cast at a target, then the pipeline processes it identically to a player cast (same events).
• AC-2 unit Given a creature caster hit hard during wind-up, when poise breaks, then its cast is interrupted.
• AC-3 content Given a creature referencing an unknown spell id, when validated, then validation fails.

Why

Fast content creation needs designers to grant spells, ignore mana and see stimuli while testing combinations.

What

Debug console commands: spell.grant <id|all>, spell.infiniteMana, spell.noCooldown, spell.forceMiscast <table>, spell.reload (hot-reload spell JSON); overlay drawing active stimulus volumes, beams and summon lifetimes.

Scope

• In: console commands, hot reload of spell content in dev builds, debug overlay
• Out: general console (e33-debug-console)

Notes

Hot reload re-validates; invalid file keeps the previous version and prints errors.

• AC-1 integration Given the debug console, when "spell.grant all" runs, then every catalogue spell is known and assignable.
• AC-2 integration Given a spell file edited with an invalid value, when hot-reloaded, then the previous definition remains active and the validation error is printed.
• AC-3 integration Given spell.infiniteMana on, when casting 100 spells, then mana never decreases.

Why

Difficult books should be a challenge, not a lock; casting a half-understood spell is a botched-cast story generator.

What

As a sorcerer, I want difficult books to be learnable early at low comprehension so that I can cast advanced spells unreliably until I understand them.

Comprehension = f(book difficulty, related known spells in same school, insights, reference books owned). Below threshold 1.0 the spell is "unstable": miscast chance scales with (1 − comprehension). Re-study or related learning raises comprehension.

Scope

• In: comprehension formula in data, unstable spell state, miscast chance hook into e06-botched-casts, UI-facing values
• Out: miscast outcomes (e06-botched-casts)

Notes

Anti-grind: comprehension rises from knowledge (books, insights, related spells), not from casting count alone.

• AC-1 unit Given difficulty 0.8 and no related knowledge, when learned, then comprehension is 0.2 and the spell is flagged unstable.
• AC-2 unit Given comprehension 0.2, when cast 1000 times with a seeded RNG, then miscast rate is 80% ± 3%.
• AC-3 unit Given the player learns two more spells of the same school, when comprehension is recomputed, then it increases by the configured per-related-spell bonus and caps at 1.0.
• AC-4 unit Given comprehension reaches 1.0, when computed, then the unstable flag clears and no comprehension miscasts occur.
• AC-5 content Given a comprehension config with negative bonuses, when validated, then validation fails.

Why

Forbidden knowledge is tempting; carrying or reading it should create risk and stories rather than a simple lock.

What

As a sorcerer, I want forbidden books to teach powerful spells at a social cost so that I decide whether to be seen with them.

Forbidden books are contraband: guards/NPCs who perceive the player reading or openly casting forbidden-school spells react (disposition drop, confiscation attempt, alert). Some carry "whispers" side effects (noise events at night, attracting certain creatures).

Scope

• In: legality flag handling, witness detection via perception hooks, disposition/world-state consequences, whisper effect
• Out: guard dialogue content (E22/E24)
• Out: buying/selling forbidden stock (e21-stolen-forbidden-stock)

Notes

Necromancy school spells (e.g. Animate Bones) come from forbidden books in the MVP.

Integration note: Promoted — Animate Bones (P1/m2) is taught by a forbidden book.

• AC-1 integration Given an NPC with a lawful disposition that perceives the player reading a forbidden book, when perception succeeds, then a ContrabandWitnessed event fires and the NPC's disposition drops by the configured amount.
• AC-2 integration Given no witness has line of sight, when the player reads a forbidden book, then no consequence fires.
• AC-3 unit Given a book with whispers, when carried in inventory between 22:00 and 04:00 in-game, then a quiet noise event is emitted every 60 s at the player position.
• AC-4 unit Given a book dropped or stashed, when whispers would emit, then they emit at the book's location instead.
• AC-5 unit Given a forbidden spell learned, when cast unseen, then no ContrabandWitnessed event fires.

Why

"Wait… I can do that?" — some magic should be discovered by experimenting, not bought.

What

As a sorcerer, I want casting my spells in unusual world situations to occasionally reveal an insight so that experimentation teaches new verbs.

Insight rules are data: trigger = (spell-sourced stimulus type, target property/state, optional second stimulus) observed by the sim; reward = new spell, spell variant, or comprehension boost. Example rules (property-based): cold stimulus on a water surface → insight "ice bridge" (unlocks Frostglass surface variant); force stimulus on a burning object → insight "flung flame".

Scope

• In: insight rule schema, observer on stimulus-resolution events, one-time unlocks, InsightDiscovered event + journal entry
• Out: the unlocked spells themselves (mw-e08)

Notes

Rules may only reference stimulus types and property tags, never object or spell ids for triggers (validator enforces).

• AC-1 unit Given an insight rule {stimulus: cold, targetProperty: waterSurface}, when a player-sourced cold stimulus freezes any water surface, then InsightDiscovered fires once and the reward is granted.
• AC-2 unit Given the same condition caused by a non-player source (a frost trap), when resolved, then no insight fires.
• AC-3 unit Given an insight already discovered, when its trigger recurs, then nothing is granted again.
• AC-4 content Given an insight rule whose trigger references an object id, when validated, then validation fails ("triggers must use properties").
• AC-5 integration Given a save after discovering an insight, when loaded, then the insight remains discovered.

Why

The anti-grind rule: progression unlocks new verbs (Firebolt → Flame Jet → Fire Wall), not +12% tiers.

What

As a sorcerer, I want mastering a spell to open its next verb, learnable by study, insight or advanced book, so that growth changes what I can do.

Verb chain data: spell.next[] with unlock requirements (comprehension 1.0, N meaningful uses — e.g. hits on distinct target types, an insight, or owning an advanced book). A content validator rejects chain steps that differ only numerically from their predecessor (same delivery + same op types).

Scope

• In: verb chain data, unlock evaluation, integration with e19 progression model, anti-grind validator
• Out: specific chains (mw-e08: Ember/Firebolt/Flame Jet/Fire Wall, Mage Hand → Telekinetic Throw & Catch)

Notes

"Meaningful uses" counts distinct contexts (distinct property tags affected), not raw cast count.

• AC-1 unit Given Firebolt with next Flame Jet requiring comprehension 1.0 and 5 distinct affected property tags, when both are met, then Flame Jet becomes studyable and a VerbUnlocked event fires.
• AC-2 unit Given 50 casts affecting only 1 property tag, when evaluated, then Flame Jet remains locked.
• AC-3 content Given a chain step with the same delivery and op types as its predecessor differing only in magnitudes, when validated, then validation fails with "numeric-only progression".
• AC-4 unit Given an advanced book that teaches a chain step directly, when studied before requirements are met, then the spell is learned but unstable (comprehension ≤ 0.5).
• AC-5 integration Given the e19 progression model, when a verb unlocks, then it is recorded as a capability unlock.

Why

Books are the sorcerer's core reward; reading one should feel like opening a treasure, not a tooltip.

What

As a sorcerer, I want to open a book into a two-page reader showing lore, diagrams and a Study button with time estimate so that reading is a moment.

Shows garbled text for unknown language, torn-page placeholders for missing pages, forbidden warning, comprehension estimate, "already known" state.

Scope

• In: reader UI, study action with time estimate, placeholder parchment/page art, keyboard + gamepad navigation, text scaling via settings
• Out: final book/parchment art (asset prompt beads)

Notes

Honour E31 text sizing and dyslexia-friendly font option if available.

• AC-1 integration Given a book with 6 pages, when opened, then pages 1–2 show and next/prev keys navigate in pairs without going past the last page.
• AC-2 integration Given a damaged book, when a missing page is shown, then a torn-page placeholder renders with the text "Pages missing".
• AC-3 integration Given the Study button, when pressed, then the displayed estimate matches the sim's computed remaining study minutes.
• AC-4 integration Given the player is in combat (Alerted/Combat nearby), when Study is pressed, then it is disabled with a tooltip reason.
• AC-5 e2e Given the greybox testbed, when a fixture book is opened and closed with the keyboard, then no console errors occur.

Why

The grimoire is the sorcerer's trophy case and planning tool — seeing locked next verbs motivates exploration.

What

As a sorcerer, I want a grimoire listing known spells by school with description, cost, comprehension, world interactions discovered and next verbs (silhouetted if locked) so that I plan experiments and hotbar loadouts.

Drag spell to hotbar; filter by school; insights journal page; "rumoured spells" section fed by NPC rumour hooks.

Scope

• In: grimoire UI, school tabs, drag-to-hotbar, insight journal page, locked-verb silhouettes, rumoured spells list
• Out: hotbar (e06-spell-hotbar-ui)
• Out: final grimoire art (asset prompt beads)

Notes

"World interactions discovered" lists property tags the player has personally affected with that spell — reinforcing the systemic model.

• AC-1 integration Given 5 known spells across 3 schools, when the grimoire opens on the Fire tab, then only Fire spells are listed.
• AC-2 integration Given a spell with a locked next verb, when viewed, then the next verb shows as a silhouette with its unlock requirement text.
• AC-3 integration Given a spell dragged onto hotbar slot 4, when dropped, then slot 4 holds that spell.
• AC-4 integration Given no known spells (non-sorcerer start), when opened, then an empty state is shown without errors.
• AC-5 integration Given the player ignited 3 distinct property-tagged objects with Ember, when viewing Ember, then those 3 property tags are listed as discovered interactions.

Why

The world rewards curiosity: bookshelves can contain books, and casters carry the books they learned from.

What

As a sorcerer, I want to search bookshelves, loot books from enemy casters and receive books from quests or characters so that books come from everywhere, not just shops.

Bookshelf-tagged objects have searchable loot tables (mostly mundane lore, occasional spell book); creature casters optionally drop a book for one of their spells; quest/dialogue reward actions can grant books.

Scope

• In: bookshelf search interaction, book entries in loot tables, caster book drop rule, grant-book action for quest/dialogue
• Out: specific placements (E24–E26 world content)
• Out: shops (mw-e21)

Notes

Searching a bookshelf takes 2 s and makes a quiet noise (thief interplay).

• AC-1 integration Given a bookshelf-tagged object with a loot table containing a spell book, when searched with a fixed seed, then the book is added to inventory and the shelf becomes "searched" in world state.
• AC-2 integration Given a searched bookshelf, when searched again, then nothing new is granted.
• AC-3 unit Given a creature caster with dropsSpellBook=true, when killed, then its loot contains a book teaching one of its spells.
• AC-4 integration Given a dialogue/quest action grantBook, when executed, then the book is added to inventory.
• AC-5 unit Given a full inventory (if capacity limits exist), when a book is found, then it is placed on the ground at the player position rather than lost.

Why

Losing a learned spell to a save bug would betray the core sorcerer reward loop.

What

Serialize known spells + comprehension, per-book study progress, known languages, discovered insights, verb unlock counters into the versioned save with round-trip and migration tests.

Scope

• In: learning save section, round-trip tests, migration scaffold
• Out: inventory contents (E17/E30)

Notes

Unknown spell ids in a save (content removed) are dropped with a warning, not a crash.

• AC-1 unit Given 4 known spells, 2 in-progress studies, 1 language and 2 insights, when saved and loaded, then all values are identical.
• AC-2 unit Given a save referencing a spell id no longer in content, when loaded, then the spell is dropped, a warning is logged and load succeeds.
• AC-3 unit Given a v1 learning save, when loaded by v2, then the migration test passes.

Why

Damaged books turn a spell into an exploration hook: the rest of the pages are somewhere in the monastery.

What

As a sorcerer, I want a damaged book to teach a partial version of its spell and point me towards missing pages so that finding pages completes it.

Missing pages are separate page items (loot) linked by bookId; a damaged book teaches nothing, lore-only, or an unstable variant as configured; combining pages restores the book.

Scope

• In: page items, combine action, partial-teach behaviour, restoration event
• Out: placement of pages in the world (E24–E26 content)

Notes

Missing page text may hint at location (Related: world-content group authors hints).

• AC-1 unit Given a damaged book missing pages 3 and 7 configured partialTeach=unstable, when studied, then the spell is learned with comprehension capped at 0.5.
• AC-2 unit Given the player holds both missing page items, when combining, then the book becomes intact, the page items are consumed and the comprehension cap lifts.
• AC-3 unit Given the player holds only page 3, when combining, then the book remains damaged with only page 7 missing.
• AC-4 content Given a page item referencing a book that is not damaged, when validated, then validation fails.

Why

Untranslated books create quests and reasons to talk to scholars and find lexicons — curiosity rewarded.

What

As a sorcerer, I want books in unknown languages to be unreadable until I obtain a lexicon or a translation so that decoding knowledge feels earned.

Languages are data; the player knows languages via lexicon items or NPC translation service (dialogue action); partially known languages show garbled text proportional to knowledge.

Scope

• In: language registry, known-language state, lexicon items, translate action hook for dialogue, garbled text renderer data
• Out: NPC translator dialogue content (E22/E24)

Notes

Garbling is deterministic per book + knowledge level so the same text appears on re-read.

• AC-1 unit Given a book in "Old Tessic" and no knowledge, when read, then study is disabled with reason "language" and pages render garbled.
• AC-2 unit Given the player acquires the Old Tessic lexicon item, when reading, then the book is studyable.
• AC-3 unit Given 50% language knowledge, when rendering page text, then the same deterministic ~50% of words are garbled on every read.
• AC-4 integration Given a dialogue action "translate book X", when executed, then the book gains a translated flag and is studyable regardless of language knowledge.

Why

Twenty spells authored by different people must feel like one coherent magic system.

What

docs/design/magic-schools.md: each school's fantasy, primary stimuli, colour tag for placeholder VFX, default loudness, miscast flavour; spell file template; naming for cue/icon/cover ids; testing template (effects + ≥2 property interactions); explicit "too useful is allowed" policy and the too-useful label.

Scope

• In: design doc
• In: spell file template in src/content/spells/_template
• In: test file template
• Out: individual spells

Notes

Schools in MVP: Fire, Frost, Storm, Arcane, Illusion, Alteration, Conjuration, Necromancy, Nature, Light, Shadow, Gravity, Time.

• AC-1 content Given the spell template file, when validated by the spell schema, then it passes.
• AC-2 content Given the magic-schools doc, when checked by a docs test, then every school enum value in the schema has a section.
• AC-3 manual Given a designer new to the project, when authoring a fixture spell from the template, then they finish in ≤ 30 minutes without code changes.

Why

The canonical anti-grind chain starts here: Firebolt is a verb, Flame Jet is a new verb, not +12%.

What

As a sorcerer, I want Firebolt (fast damaging bolt that ignites on impact) and, once unlocked, Flame Jet (channeled short cone that sets everything in front ablaze and pushes gas) so that fire progression changes how I fight.

Flame Jet is a chain step of Firebolt with a different delivery (cone channel).

Scope

• In: Firebolt spell
• In: Flame Jet spell
• In: verb chain link Firebolt → Flame Jet
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Fire. Placeholder cues: vfx-spell-firebolt-{cast,travel,impact}, vfx-spell-flame-jet-channel, sfx-spell-firebolt-{cast,impact}, sfx-spell-flame-jet-loop. Icon: icon-spell-firebolt. Book cover: ui-book-cover-firebolt.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given any flammable object hit by Firebolt, when resolved, then it ignites and a 1 m heat splash applies around the impact.
• AC-3 integration Given Flame Jet channeled for 2 s at a group of objects, when resolved, then every flammable object in the cone ignites and non-flammable ones take no fire state.
• AC-4 integration Given a gas volume tagged flammableGas in the Flame Jet cone, when channeled, then the gas ignites per E03 propagation (explosion allowed).
• AC-5 content Given the Firebolt → Flame Jet chain, when checked by the anti-grind validator, then it passes (different delivery).
• AC-6 unit Given Flame Jet channeled with mana running out mid-channel, when mana reaches 0, then channeling stops cleanly (over-cast disallowed).

Why

Constitution example; freezing water is a key puzzle verb and ice surfaces are traversal and combat tools.

What

As a sorcerer, I want a frost bolt that freezes what it hits and, aimed at water or ground, creates a temporary slick ice sheet so that I can freeze enemies, cross water and make foes slip.

Cold stimulus at impact; creatures accumulate chill → slowed → frozen-solid (shatterable by force); water surfaces become walkable ice for 20 s; ground becomes slippery (low friction).

Scope

• In: Frostglass spell
• In: ice sheet surface volume
• In: frozen-solid shatter rule via force
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Frost. Placeholder cues: vfx-spell-frostglass-{cast,travel,impact}, vfx-spell-frostglass-ice-sheet, sfx-spell-frostglass-{cast,impact,shatter}. Icon: icon-spell-frostglass. Book cover: ui-book-cover-frostglass.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given any waterSurface area, when hit by Frostglass, then a walkable frozen patch of 3 m radius exists for 20 s then melts back to water.
• AC-3 integration Given a creature hit 3 times within 4 s, when resolved, then it becomes frozen-solid for 3 s; a force ≥ threshold during that time shatters the ice and deals bonus damage.
• AC-4 integration Given a creature walking onto an ice sheet, when moving, then its friction is reduced per E03 and it can slip (knockdown chance per creature balance data).
• AC-5 integration Given a burning object hit by Frostglass, when resolved, then it is extinguished.
• AC-6 unit Given a creature tagged cold-immune, when hit, then no chill accumulates.

Why

Conjured water is the connective tissue for combinations: douse fires, fill basins, soak foes for frost and lightning.

What

As a sorcerer, I want to conjure a splash of water at a point so that I can extinguish fires, make things wet and fill containers.

AoE moisture stimulus 3 m; creates a temporary puddle (waterSurface) for 15 s; fills container-tagged objects (basins, scales) with water weight.

Scope

• In: Wellspring spell
• In: temporary puddle volume
• In: container fill via moisture
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Conjuration. Placeholder cues: vfx-spell-wellspring-{cast,burst}, vfx-spell-wellspring-puddle, sfx-spell-wellspring-{cast,splash}. Icon: icon-spell-wellspring. Book cover: ui-book-cover-wellspring.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given burning objects inside the 3 m area, when Wellspring bursts, then they are extinguished and gain wet.
• AC-3 integration Given a creature in the area, when hit, then it gains wet for 20 s, which E03 rules make conductive and freezable.
• AC-4 integration Given an object tagged container under the burst, when filled, then its weight increases by the configured water mass (can trigger pressure plates).
• AC-5 unit Given Wellspring cast in a space with no floor within 10 m, when resolved, then water falls and the puddle forms where it lands or nowhere if it falls out of world.

Why

Constitution example; force + gas-clearing gives a cheap, flexible physics verb.

What

As a sorcerer, I want a cone of wind that pushes creatures and loose objects and disperses gas and smoke so that I can shove foes off ledges or clear poison.

Cone force impulse scaled by weight; displaces gas volumes; fans existing fires (spreads heat downwind) and blows out small flames.

Scope

• In: Gust spell
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Storm. Placeholder cues: vfx-spell-gust-cone, sfx-spell-gust-whoosh. Icon: icon-spell-gust. Book cover: ui-book-cover-gust.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given loose objects of 5 kg and 80 kg in the cone, when Gust fires, then the 5 kg object is thrown ≥ 4 m and the 80 kg object moves ≤ 0.5 m.
• AC-3 integration Given a poisonous gas volume in the cone, when Gust fires, then the gas concentration in the cone drops by ≥ 80% and relocates downwind.
• AC-4 integration Given a small flame (candle-scale light source) in the cone, when Gust fires, then it is extinguished; a large fire instead spreads 1 m downwind.
• AC-5 integration Given a creature near a ledge, when pushed past the edge, then it falls and takes fall damage per creature rules.
• AC-6 unit Given a creature tagged anchored/heavy, when hit, then it staggers but is not displaced.

Why

Constitution example: short-range shockwave with substantial environmental effects — and it is loud.

What

As a sorcerer, I want a point-blank radial shockwave that knocks creatures down, breaks fragile objects and shakes loose unstable things so that I can reshape a room in one cast.

Radial force + fragile-break + "unstable" property trigger (loose rubble, chandeliers) + very loud noise event + short deafened status on creatures (reduced hearing).

Scope

• In: Thunderclap spell
• In: deafened status
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Storm. Placeholder cues: vfx-spell-thunderclap-wave, sfx-spell-thunderclap-boom. Icon: icon-spell-thunderclap. Book cover: ui-book-cover-thunderclap.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given any object tagged fragile within 4 m, when Thunderclap fires, then it breaks.
• AC-3 integration Given any object tagged unstable (hanging, precarious) within 6 m, when Thunderclap fires, then it detaches and falls under physics.
• AC-4 integration Given creatures within 4 m, when hit, then they are knocked down if poise is exceeded and gain deafened for 5 s.
• AC-5 integration Given sleeping or Unaware creatures within audible range of loudness 100, when Thunderclap fires, then they transition to at least Suspicious via the alert state machine.
• AC-6 unit Given the caster standing on an ice sheet, when cast, then the caster is not displaced by their own shockwave (self-exempt) but the ice cracks if tagged fragile.

Why

Electricity rewards reading the environment: wet enemies, chains, puddles and metal armour become conductors.

What

As a sorcerer, I want a short beam of lightning that jumps between conductive or wet targets so that soaking enemies or luring them onto metal pays off.

Beam with charge stimulus; E03 conduction propagates through connected conductive/wet objects and water surfaces; powers chargeActivated mechanisms.

Scope

• In: Arc Lash spell
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Storm. Placeholder cues: vfx-spell-arc-lash-beam, vfx-spell-arc-lash-chain, sfx-spell-arc-lash-crackle. Icon: icon-spell-arc-lash. Book cover: ui-book-cover-arc-lash.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given three wet creatures standing in a connected puddle, when one is hit, then all three receive charge damage via propagation.
• AC-3 integration Given a dry creature not touching conductive objects, when hit, then only that creature is affected.
• AC-4 integration Given any mechanism tagged chargeActivated, when hit by Arc Lash, then it activates.
• AC-5 integration Given the caster standing in the same puddle as the target, when cast, then the caster also receives charge damage (failure creates stories).
• AC-6 unit Given a creature tagged insulated, when in a conductive chain, then propagation stops at it.

Why

Anti-grind example from the constitution: telekinesis → throwing creatures → catching projectiles.

What

As a sorcerer, I want Mage Hand's follow-on verb to fling held objects (or small creatures) and snatch incoming projectiles from the air so that telekinesis becomes a combat tool.

Throw applies a large impulse to the held entity; creatures under a size tag can be grabbed; catch window 0.4 s freezes an incoming projectile, which can then be thrown back.

Scope

• In: Telekinetic Throw spell (chain step of Mage Hand)
• In: projectile catch
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Arcane. Placeholder cues: vfx-spell-telekinetic-throw-{launch,catch}, sfx-spell-telekinetic-throw-{launch,catch}. Icon: icon-spell-telekinetic-throw. Book cover: ui-book-cover-telekinetic-throw.

TOO USEFUL — INTENTIONALLY ALLOWED (constitution §4). Do not nerf this spell to prevent sequence breaks or cheese; encounters and puzzles must tolerate it. Balance only via mana/cast time/cooldown, never by adding pairwise exceptions.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given a held 10 kg object thrown at a creature, when it hits at ≥ 8 m/s, then collision damage and stagger apply per force rules.
• AC-3 integration Given a small-tagged creature within range, when grabbed and thrown into any fragile object, then the object breaks and the creature takes collision damage.
• AC-4 integration Given an incoming arrow or spell projectile, when catch is timed within the 0.4 s window, then the projectile stops and can be re-thrown with its on-hit ops intact.
• AC-5 unit Given catch pressed outside the window, when the projectile arrives, then the catch fails and the projectile hits normally.
• AC-6 content Given the Mage Hand → Telekinetic Throw chain, when checked by the anti-grind validator, then it passes.

Why

Constitution example; opens balconies, gaps and escapes — a classic sequence-breaker we deliberately allow.

What

As a sorcerer, I want to teleport up to 8 m to an aimed point with line of sight so that I can cross gaps, escape grapples and reach ledges.

Uses displacement op with LOS; breaks rooted/grappled statuses; leaves a brief light flash and a quiet noise at origin and destination.

Scope

• In: Blink spell
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Arcane. Placeholder cues: vfx-spell-blink-{depart,arrive}, sfx-spell-blink-{depart,arrive}. Icon: icon-spell-blink. Book cover: ui-book-cover-blink.

TOO USEFUL — INTENTIONALLY ALLOWED (constitution §4). Do not nerf this spell to prevent sequence breaks or cheese; encounters and puzzles must tolerate it. Balance only via mana/cast time/cooldown, never by adding pairwise exceptions.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given a ledge 7 m away and 3 m up with line of sight, when Blink is cast at it, then the caster stands on the ledge.
• AC-3 integration Given a closed window or bars between caster and target, when cast, then Blink fails with reason "no-line-of-sight" and mana is refunded.
• AC-4 integration Given the caster is rooted by any restraint status, when Blink succeeds, then the rooted status is removed.
• AC-5 integration Given a guard watching the origin point, when the caster blinks out of sight, then the guard's last-known position is the origin (it did not see the destination).
• AC-6 unit Given aim into a wall, when cast, then the caster lands at the nearest safe point before the wall.

Why

Constitution: "the sorcerer might levitate onto a balcony" — vertical traversal as a class-specific solution.

What

As a sorcerer, I want to levitate myself (or a targeted object) upward at 1.5 m/s for up to 6 s so that I can reach balconies, cross pits slowly or lift a platform.

Channeled lift status; drains mana per tick; ends gently (slow-fall for 2 s) when mana runs out or cancelled.

Scope

• In: Levitate spell
• In: slow-fall ending
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Gravity. Placeholder cues: vfx-spell-levitate-loop, sfx-spell-levitate-loop. Icon: icon-spell-levitate. Book cover: ui-book-cover-levitate.

TOO USEFUL — INTENTIONALLY ALLOWED (constitution §4). Do not nerf this spell to prevent sequence breaks or cheese; encounters and puzzles must tolerate it. Balance only via mana/cast time/cooldown, never by adding pairwise exceptions.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given a balcony 6 m above, when Levitate is channeled for 4 s with lateral steering, then the caster reaches the balcony.
• AC-3 integration Given mana runs out mid-levitation, when it ends, then the caster slow-falls for 2 s and takes no fall damage from ≤ 8 m.
• AC-4 integration Given Levitate targeted at any object below the lift weight limit, when channeled, then it rises (e.g. lifting a platform or pressure-plate weight).
• AC-5 unit Given the caster is hit and poise breaks while levitating, when interrupted, then levitation ends and slow-fall applies.

Why

Constitution example; clumping enemies and debris sets up spectacular combinations.

What

As a sorcerer, I want to place a 5 s gravity point that drags creatures and loose objects inward so that I can group enemies, pull them off ledges or yank items from shelves.

Persistent volume with damped point pull scaled by weight; objects pulled into each other collide (fragile breaks); combos with Thunderclap/Flame Jet.

Scope

• In: Gravity Well spell
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Gravity. Placeholder cues: vfx-spell-gravity-well-loop, vfx-spell-gravity-well-collapse, sfx-spell-gravity-well-{loop,collapse}. Icon: icon-spell-gravity-well. Book cover: ui-book-cover-gravity-well.

TOO USEFUL — INTENTIONALLY ALLOWED (constitution §4). Do not nerf this spell to prevent sequence breaks or cheese; encounters and puzzles must tolerate it. Balance only via mana/cast time/cooldown, never by adding pairwise exceptions.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given 4 creatures within 6 m, when the well persists 5 s, then all creatures under the heavy tag end within 1.5 m of the centre.
• AC-3 integration Given loose objects on a shelf within range, when the well activates, then they are pulled off and fly toward the centre, breaking if fragile on collision.
• AC-4 integration Given the well placed beyond a ledge, when creatures are pulled, then they fall.
• AC-5 unit Given the caster within the radius, when active, then the caster is also pulled (not exempt) — intentionally allowed hazard.
• AC-6 perf Given 30 loose objects and 8 creatures in one well, when simulated, then magic + physics cost p95 ≤ 3 ms per tick on the reference machine.

Why

Constitution example; deception that plugs into the stealth/AI simulation instead of damage.

What

As a sorcerer, I want to project an illusory copy of myself at a point so that enemies investigate or attack it while I slip away.

Summons a decoy entity perceivable by sight and hearing (emits footsteps/noise), attracting perception priority; dispels when struck or on duration end; creatures tagged true-sight ignore it.

Scope

• In: Mirror Self spell
• In: decoy entity archetype
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Illusion. Placeholder cues: vfx-spell-mirror-self-{spawn,shimmer,dispel}, sfx-spell-mirror-self-{spawn,dispel}. Icon: icon-spell-mirror-self. Book cover: ui-book-cover-mirror-self.

TOO USEFUL — INTENTIONALLY ALLOWED (constitution §4). Do not nerf this spell to prevent sequence breaks or cheese; encounters and puzzles must tolerate it. Balance only via mana/cast time/cooldown, never by adding pairwise exceptions.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given a guard in Unaware state with the decoy in view, when perceived, then the guard targets the decoy as the player-equivalent stimulus (Suspicious → Investigating toward decoy).
• AC-3 integration Given the decoy is struck by any attack, when hit, then it dispels and the attacker enters Searching with last-known position at the decoy.
• AC-4 integration Given a creature tagged true-sight, when the decoy is in view, then it is ignored.
• AC-5 integration Given the decoy placed in darkness, when guards look, then its visibility uses the same light-field-based visibility model as the player.
• AC-6 unit Given maxActive 1, when cast again, then the previous decoy dispels.

Why

Navigating darkness is a puzzle pillar; a light spell trades stealth for vision and scares light-averse creatures.

What

As a sorcerer, I want to summon a floating light orb that follows me or can be sent to a point so that I can explore dark places, light up puzzles and drive back light-averse creatures.

Light op attached to a small summoned entity (follow or fixed); 90 s duration; sends creatures tagged light-averse into fear via their own properties.

Scope

• In: Wisp Light spell
• In: wisp entity
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Light. Placeholder cues: vfx-spell-wisp-light-orb, sfx-spell-wisp-light-{spawn,hum-loop}. Icon: icon-spell-wisp-light. Book cover: ui-book-cover-wisp-light.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given a dark room with light level 0.05, when the wisp is present, then the light field within 5 m of it is ≥ 0.5.
• AC-3 integration Given any light-sensitive mechanism (tagged lightActivated) within the wisp's radius above its threshold, when the wisp arrives, then it activates.
• AC-4 integration Given a creature tagged light-averse within 5 m, when the wisp approaches, then it retreats per its disposition rules.
• AC-5 integration Given the player carries a follow-wisp while sneaking, when a guard evaluates visibility, then the player's visibility score increases per the visibility model (stealth trade-off).
• AC-6 unit Given a Snuff/darkness effect overlapping the wisp, when resolved, then the wisp is extinguished early.

Why

Puzzle constitution: redirecting light. A reflecting beam turns mirrors and polished shields into puzzle pieces.

What

As a sorcerer, I want a channeled beam of focused light that bounces off reflective surfaces so that I can power light mechanisms around corners and dazzle creatures.

Beam with light op and small heat; reflects up to 3 times off reflective-tagged objects; dazzles (blinded 2 s) creatures that look into it.

Scope

• In: Prism Ray spell
• In: blinded status
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Light. Placeholder cues: vfx-spell-prism-ray-beam, vfx-spell-prism-ray-reflect, sfx-spell-prism-ray-loop. Icon: icon-spell-prism-ray. Book cover: ui-book-cover-prism-ray.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given a reflective object angled at 45°, when the beam hits it, then the reflected beam continues at 90° and can activate a lightActivated mechanism on the reflected path.
• AC-3 integration Given any object tagged reflective including a Frostglass ice sheet, when hit, then the beam reflects (property-based, not object-based).
• AC-4 integration Given a creature facing the caster within the beam, when hit, then it gains blinded for 2 s and its sight perception is disabled for that time.
• AC-5 unit Given more than 3 reflective surfaces in a row, when channeled, then the beam stops after the third reflection.

Why

Constitution example; restraint as a verb — and vines are organic, so fire and frost interact naturally.

What

As a sorcerer, I want roots to erupt in a 3 m area and bind creatures for up to 6 s so that I can pin enemies, hold doors or stop a fleeing thief.

Spawns vine entities (tagged organic, flammable) that apply rooted to creatures inside; rooted breaks when its vine entity is destroyed or burns; only works on soil/stone ground, not water or wood floors above empty space.

Scope

• In: Grasping Vines spell
• In: vine entity archetype
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Nature. Placeholder cues: vfx-spell-grasping-vines-{erupt,hold,wither}, sfx-spell-grasping-vines-{erupt,creak}. Icon: icon-spell-grasping-vines. Book cover: ui-book-cover-grasping-vines.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given 3 creatures inside the area, when cast, then each is rooted for 6 s unless tagged too-large-to-root.
• AC-3 integration Given the vines are hit by any heat stimulus above ignition threshold, when they burn, then they are destroyed within 2 s and rooted ends early.
• AC-4 integration Given a strong creature with the break-free trait, when rooted, then it breaks free after 2 s.
• AC-5 unit Given cast onto a water surface, when resolved, then the cast fails with reason "invalid-surface" and mana is refunded.
• AC-6 integration Given a life stimulus source nearby (e.g. a second Vines cast), when overlapping, then vine duration refreshes (organic growth rule).

Why

Constitution example; summoning grounded in the world — no bones nearby, no servant.

What

As a sorcerer, I want to assemble skeletal remains within 6 m into a temporary skeletal servant for 45 s so that the dungeon's dead fight for me.

Summon op requiring remains-tagged material; servant follows/attacks; crumbles back into remains (reusable) on expiry; the spell comes from a forbidden book.

Scope

• In: Animate Bones spell
• In: skeletal servant archetype (reuses bestiary skeleton data if available)
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Necromancy. Placeholder cues: vfx-spell-animate-bones-{assemble,crumble}, sfx-spell-animate-bones-{assemble,rattle,crumble}. Icon: icon-spell-animate-bones. Book cover: ui-book-cover-animate-bones.

Related: bestiary skeleton archetype from the creatures group; use a fixture archetype until it exists. Book is forbidden (e07-forbidden-books).

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given a remains-tagged object within 6 m, when cast, then a servant with caster allegiance spawns at the remains location and the remains are consumed.
• AC-3 integration Given no remains in range, when cast, then the cast fizzles with a SpellFizzled event and full refund.
• AC-4 integration Given the servant expires, when 45 s pass, then a remains-tagged object is left at its position (reusable).
• AC-5 integration Given a hostile skeleton creature, when targeted with Animate Bones, then it is not converted (only inert remains qualify).
• AC-6 integration Given a lawful NPC witness, when the player casts Animate Bones in view, then a ContrabandWitnessed event fires (forbidden school).

Why

Constitution example; transformation that solves fights, social problems and puzzles in unexpected ways.

What

As a sorcerer, I want to turn a creature into a small animal (e.g. a chicken-like critter) for 20 s so that I can neutralise threats, sneak past guards or make a heavy creature light enough to Mage Hand.

Transform op to a small-creature archetype; large/boss creatures tagged immune-transform resist; the critter form has its own properties (small, light, flammable) that other spells interact with.

Scope

• In: Polymorph spell
• In: critter archetype for the transformed form
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Alteration. Placeholder cues: vfx-spell-polymorph-{poof-in,poof-out}, sfx-spell-polymorph-{poof,critter-cluck}. Icon: icon-spell-polymorph. Book cover: ui-book-cover-polymorph.

TOO USEFUL — INTENTIONALLY ALLOWED (constitution §4). Do not nerf this spell to prevent sequence breaks or cheese; encounters and puzzles must tolerate it. Balance only via mana/cast time/cooldown, never by adding pairwise exceptions.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given a humanoid guard, when polymorphed, then it becomes the critter archetype for 20 s with its alert memory intact and restores exactly afterwards.
• AC-3 integration Given a polymorphed creature now tagged small and light, when grabbed by Mage Hand/Telekinetic Throw, then it can be lifted (properties of the new form apply).
• AC-4 integration Given a creature tagged immune-transform, when targeted, then it resists and emits SpellResisted.
• AC-5 integration Given the critter is killed and Polymorph uses onFormDeath=revert, when killed, then the original reverts with 1 HP and a stagger status.
• AC-6 unit Given Polymorph cast on the caster via a reflective surface bounce (if a beam reflection returns), when resolved, then the caster is transformed — intentionally allowed.

Why

Emergent combinations are the point of magic; a suite proves they emerge from properties and catches regressions when E03 rules change.

What

A generated integration test matrix: for every pair of stimulus types emitted by catalogue spells (heat, cold, force, light, moisture, charge, life, decay) applied in both orders to a fixture set of property-tagged generic objects and creatures, assert deterministic, rule-consistent outcomes (golden snapshots). Plus curated scenario tests (Wellspring → Arc Lash chain, Wellspring → Frostglass bridge, Gravity Well → Thunderclap scatter, Gust → Fire spread, Polymorph → Telekinetic Throw) expressed via properties. Plus a greybox "spell gallery" room exhibiting every property for e2e/manual play.

Scope

• In: combination matrix generator, golden snapshots, curated scenario tests, greybox gallery scene, e2e smoke of the gallery
• Out: E03 rule changes (Core group) — the suite reports, it does not fix

Notes

Golden updates require an explicit reviewed flag in the PR. Suite runs headless in < 60 s.

• AC-1 integration Given every ordered pair of catalogue stimulus types and the fixture object set, when run, then outcomes match golden snapshots and two runs produce identical hashes.
• AC-2 integration Given Wellspring then Arc Lash on 3 creatures in the puddle, when resolved, then all receive charge damage (conduction through wet).
• AC-3 integration Given Wellspring then Frostglass on a pit filled with water, when resolved, then a walkable ice surface exists across the pit.
• AC-4 unit Given a new spell added to the catalogue emitting an unseen stimulus type, when the suite runs, then it fails listing the missing pair coverage.
• AC-5 perf Given the full suite, when run in CI, then it completes in ≤ 60 s.
• AC-6 e2e Given the greybox spell gallery, when a scripted run casts every P1 spell at its gallery station, then no console errors occur and each station reports its expected property change.
• AC-7 manual Given a 30-minute playtest in the gallery, when testers experiment freely, then they report at least 3 combinations nobody told them about.

Why

Third step of the fire chain: area denial as a new verb — split a fight, block a corridor, burn a bridge.

What

As a sorcerer, I want to draw a 6 m line of fire on the ground that lasts 8 s so that I can block paths and ignite what crosses it.

Persistent line volume with heat stimulus; creatures with fear-of-fire disposition refuse to cross; it emits strong light.

Scope

• In: Fire Wall spell
• In: chain link Flame Jet → Fire Wall
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Fire. Placeholder cues: vfx-spell-fire-wall-{spawn,loop,end}, sfx-spell-fire-wall-{spawn,loop}. Icon: icon-spell-fire-wall. Book cover: ui-book-cover-fire-wall.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given a Fire Wall placed across a corridor, when a creature whose disposition includes fears-fire paths to the player, then it routes around or waits rather than crossing.
• AC-3 integration Given a wooden floor tagged flammable under the wall, when the wall persists 8 s, then the floor ignites and continues burning after the wall ends.
• AC-4 integration Given a moisture stimulus (rain, Wellspring) applied to the wall, when resolved, then wall segments inside the moisture area are extinguished.
• AC-5 unit Given a placement line crossing a wall, when cast, then the fire line is clipped at the geometry.

Why

Magic "can reveal secrets" — a sorcerer-flavoured alternative to the thief finding hidden passages.

What

As a sorcerer, I want a 15 s self buff that outlines objects tagged hidden, magical or trapped within 12 m so that I discover secret doors, traps and enchanted items.

Revealed status applied to matching entities; revealed traps become visible to the player only.

Scope

• In: Arcane Sight spell
• In: revealed status rendering hook (placeholder outline)
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Arcane. Placeholder cues: vfx-spell-arcane-sight-overlay, sfx-spell-arcane-sight-cast. Icon: icon-spell-arcane-sight. Book cover: ui-book-cover-arcane-sight.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given any object tagged hidden within 12 m, when Arcane Sight is active, then it gains revealed for the duration.
• AC-3 integration Given a trap tagged trapped behind a wall with no line of sight, when active, then it is revealed (sight through walls up to 12 m).
• AC-4 integration Given a hidden object at 13 m, when active, then it is not revealed until the player comes within 12 m.
• AC-5 unit Given Arcane Sight ends, when duration expires, then all revealed statuses it applied are removed.

Why

Shadow magic that serves stealth: the sorcerer can create darkness like the thief extinguishing lights.

What

As a sorcerer, I want to snuff every flame and magical light within 6 m so that I can plunge a room into darkness and use it to hide or disorient.

AoE that removes lit state from light-source objects and ends light ops; smothers small fires (not large blazes); quiet cast.

Scope

• In: Snuff spell
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Shadow. Placeholder cues: vfx-spell-snuff-pulse, sfx-spell-snuff-hush. Icon: icon-spell-snuff. Book cover: ui-book-cover-snuff.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given 5 lit candles and 1 torch within 6 m, when Snuff is cast, then all become unlit and the light field drops accordingly.
• AC-3 integration Given a burning object larger than the small-fire threshold, when Snuff is cast, then it remains burning.
• AC-4 integration Given guards in the darkened room, when they evaluate visibility of the player, then the reduced light level lowers the player's visibility score.
• AC-5 unit Given Snuff noise loudness, when cast, then emitted loudness is ≤ 20 (quiet).

Why

Time magic creates novel puzzle and combat verbs: hold a falling gate, stop a swinging blade, pause an enemy.

What

As a sorcerer, I want to lock a single target in time for 5 s so that I can freeze moving traps, suspend falling objects or stop an attacker mid-swing.

Stasis status: zero velocity, immune to damage and forces, mechanisms paused; stored impulses from hits apply on release (capped).

Scope

• In: Stasis spell
• In: stasis status with stored impulse
• In: spell JSON + book item JSON, unit tests for effects, integration tests for world interactions, placeholder cue ids
• Out: final VFX/SFX/icon/book cover (asset prompt beads)
• Out: world placement of the book (E24–E26)

Notes

School: Time. Placeholder cues: vfx-spell-stasis-{lock,release}, sfx-spell-stasis-{lock,release}. Icon: icon-spell-stasis. Book cover: ui-book-cover-stasis.

TOO USEFUL — INTENTIONALLY ALLOWED (constitution §4). Do not nerf this spell to prevent sequence breaks or cheese; encounters and puzzles must tolerate it. Balance only via mana/cast time/cooldown, never by adding pairwise exceptions.

• AC-1 content Given the spell data file, when validated and loaded, then it has a school, book item, vfxCue and sfxCue and casts in the headless sim.
• AC-2 integration Given a moving mechanism (swinging blade, falling portcullis), when stasis is applied, then it halts for 5 s and resumes its prior motion afterwards.
• AC-3 integration Given a creature in stasis struck by forces totalling 300 N·s, when stasis ends, then the stored impulse (capped at 200 N·s) applies at once.
• AC-4 unit Given a creature in stasis, when damage is applied, then health is unchanged during stasis.
• AC-5 unit Given a creature tagged boss or immune-time, when targeted, then duration is reduced to 1 s.

Why

Stealth feel must be tunable by designers without code, and defaults pinned by tests.

What

A StealthTuning schema and default file: light-term curve, stance multipliers, motion bands, distance falloff, contrast bonus, gait base dB, surface table link, door-state and wall/floor transmission defaults, global audibility floor, hiding-place concealment defaults. Hot reload in dev; difficulty presets can override a whitelisted subset.

Scope

• In: Schema, defaults, hot reload, pinned default tests
• In: Difficulty override whitelist (consumed by e09-stealth-assists)
• Out: AI thresholds and timers (e11-ai-tuning-data)
• Out: Sense ranges (e12-sense-profiles)

Notes

Curves are piecewise-linear point lists for easy editing.

• AC-1 content Given the default file, when validated, then all curves are monotonic in the documented direction.
• AC-2 unit Given a difficulty preset overriding a non-whitelisted key, when loaded, then loading fails naming the key.
• AC-3 integration Given a dev build, when the stance multiplier is edited, then the visibility HUD reflects the change within 1 s.
• AC-4 unit Given the shipped defaults, when the pinned-default test runs, then documented values match.

Why

Verticality is a stealth input: rafters, balconies and rooftops should be real thief highways because guards rarely look up.

What

As a thief, I want to be harder to notice above or below a guard’s eye line so that rafters and ledges become routes. Vertical sight: per-profile vertical half-angles up/down (humanoids look up poorly by default, e.g. +20° up, −45° down) with an extra falloff; creatures with looks-up tag (e.g. spiders, fliers) ignore it; Searching agents perform occasional look-up scans. Height difference affects noise via floor transmission (already in propagation) and adds a visibility penalty when the target is above a lit area but itself in dark rafters.

Scope

• In: Vertical cone limits + falloff
• In: Look-up scan in Searching/Investigating when LKP is elevated
• In: Tuning data
• Out: Climbing movement (e10-climbing)
• Out: Flying creature navigation (post-MVP)

Notes

LKP elevation > 2 m above agent eye triggers look-up scan behaviour.

• AC-1 unit Given a humanoid guard with vertical up half-angle 20°, when the target is 30° above horizontal at 8 m, then no sight stimulus is produced while the guard is Unaware.
• AC-2 unit Given a creature tagged looks-up, when the same target is evaluated, then a sight stimulus is produced.
• AC-3 unit Given a Searching guard whose LKP is 4 m above its eye height, when it reaches the LKP area, then it performs a look-up scan that widens its up half-angle to 60° for 3 s.
• AC-4 integration Given a scenario where the player crosses rafters above 2 patrolling guards in a lit hall, when simulated, then neither guard exceeds Suspicious.
• AC-5 unit Given the target directly overhead (90°), when evaluated, then no division-by-zero or NaN occurs and visibility follows the vertical rule.

Why

Hiding places are a constitution stealth input and give failed stealth a recovery move rather than a reload.

What

As a thief, I want to duck into a wardrobe, crouch in tall grass or press into an alcove so that I can break pursuit if I am quick and unseen. Two kinds: enter-able containers (wardrobe, barrel, under-bed — interact to enter/exit, player hidden, limited view) and concealment volumes (tall grass, curtains, alcoves — visibility multiplier while inside and stationary/crouched). Data: concealment factor, entry noise dB, entry time, capacity, searchable flag. AI that saw you enter knows; AI searching may check (e11-search-behaviour). Being found inside = instant Combat at close range with a brief stagger window to flee (failure creates a story).

Scope

• In: Hiding place component + data
• In: Enter/exit via interaction with timing and noise
• In: Concealment volumes feeding visibility model
• In: Being-discovered outcome
• Out: Hiding bodies in containers (e10-body-carry-hide)
• Out: Hiding prop art (placeholder boxes)

Notes

Entering while observed must be tracked as a memory saw-enter:<hidingPlaceId>.

• AC-1 unit Given the player inside a wardrobe, when any observer computes visibility, then the result is 0 unless the wardrobe is opened by that observer.
• AC-2 unit Given the player crouched still in tall grass (concealment 0.3), when visibility is computed, then it is multiplied by 0.3; moving at run speed removes the bonus.
• AC-3 unit Given a guard observes the player entering a barrel, when the guard searches, then the barrel is checked first (via memory saw-enter).
• AC-4 unit Given an occupied single-capacity hiding place, when a second entity attempts to enter, then entry is refused with a prompt reason.
• AC-5 integration Given a guard opens the wardrobe the player is in, when discovered, then the guard enters Combat and the player gets a 0.6 s stagger window during which exiting is possible.
• AC-6 e2e Given the stealth grey-box course, when the player enters and exits a wardrobe via the interact action, then the camera switches to the hiding view and back without errors.

Why

A thief’s actions leave traces; guards noticing a doused torch or a missing colleague creates tension and new situations.

What

As a thief, I want guards to notice what I changed so that covering my tracks is part of the craft. Anomalies are world facts compared against an agent’s expected state (routine knowledge): unconscious/dead bodies, doors open that are normally closed, lights extinguished that are normally lit, missing valuables from a watched spot, blood/broken objects, a colleague missing from post for > N s. Each anomaly type has data: detection via sight (uses visibility of the anomaly), suspicion weight, resulting state (Suspicious for an open door, Alerted for a body).

Scope

• In: Anomaly registry + expected-state baseline per area/agent
• In: Perception integration (seen-anomaly stimulus)
• In: Data-driven anomaly types + weights
• Out: Relighting/closing on return (e11-return-to-routine)
• Out: Body carrying (e10-body-carry-hide)

Notes

Expected state captured at level load and updated when guards themselves change things. Hidden bodies (in containers or darkness below visibility threshold) are not perceived.

• AC-1 unit Given a door normally closed is open, when a guard perceives it, then it becomes Suspicious and investigates the door.
• AC-2 unit Given a guard perceives an unconscious body, when evaluated, then it enters Alerted, investigates the body and may wake it (unconscious ally wakes after 3 s interaction).
• AC-3 unit Given a body placed in a container or in light ≤ 0.1 beyond 6 m, when a guard walks its route, then the body is not perceived.
• AC-4 unit Given a torch normally lit is extinguished, when a guard perceives the dark torch, then it becomes Suspicious (not Alerted) with cause light-out.
• AC-5 unit Given a guard itself opened a door on its route, when it later perceives that door open, then no anomaly is raised.
• AC-6 integration Given a scenario where the player knocks out guard A and leaves the body in a lit corridor on guard B’s route, when guard B passes, then guard B enters Alerted within 2 s of the body entering its view.

Why

Stealth cannot be an eye icon: the player needs to read the inputs — how lit they are and how loud they are — to make decisions.

What

As a thief, I want to see how exposed and how loud I am so that I can make informed choices. Elements: light-exposure gem (continuous, 5 visual bands, driven by self-visibility), noise ring around the character footprint that pulses at each emitted noise scaled to its audible radius for a typical guard, stance/gait indicator, and optional diegetic cue (character material darkens in shadow). All toggleable and scalable via settings.

Scope

• In: HUD components in src/ui, data binding from sim
• In: Placeholder art (CSS/SVG shapes)
• In: Settings toggles
• Out: Per-enemy awareness (e09-awareness-indicators)
• Out: Final UI art (asset beads)

Notes

Noise ring radius = distance at which the emitted noise drops to the baseline humanoid hearing threshold in open space.

• AC-1 unit Given self-visibility 0.05, when the HUD model is computed, then the gem shows band 1 of 5; at 0.95 band 5.
• AC-2 unit Given a 40 dB footstep and a 30 dB baseline threshold, when the noise ring is computed, then its radius is ≈ 3.2 m (10^(10/20)).
• AC-3 integration Given the settings toggle hud.stealth=false, when the HUD renders, then no stealth elements are present.
• AC-4 e2e Given the stealth grey-box course, when the player walks from a lit area into shadow, then the gem band decreases within 250 ms.
• AC-5 manual Given a playtest with 3 players, when asked to explain the gem and ring without instructions after 5 min, then at least 2 describe both correctly.

Why

Players must see who is noticing them, from where, how close to detection, and why — without the omniscient radar that kills tension.

What

As a thief, I want directional indicators for enemies becoming aware of me so that I can react before being caught. For each agent with awareness > 0 of the player: an indicator near the agent (on-screen) or screen-edge chevron (off-screen) that fills with awareness and changes shape per alert state (6 distinct shapes, not colour alone), plus a cause glyph (eye = saw, ear = heard, hand = anomaly). Indicators only appear for agents the player could plausibly perceive (on-screen or heard), preserving uncertainty.

Scope

• In: Indicator model from awareness + state + dominant cause
• In: On-screen and off-screen rendering
• In: Colour-blind-safe shapes
• Out: Text barks (e09-alert-barks)
• Out: Last-known-position ghost (e09-lkp-ghost)

Notes

Indicators for agents the player has not perceived (behind walls, silent) are hidden unless an accessibility assist enables them (e09-stealth-assists).

• AC-1 unit Given an agent at awareness 0.45 in Suspicious caused by hearing, when the indicator model is built, then fill = 0.45, shape = suspicious, glyph = ear.
• AC-2 unit Given an agent behind a wall that has not been seen or heard by the player, when the model is built with assists off, then no indicator is produced.
• AC-3 unit Given each of the six states, when shapes are resolved, then six distinct shape ids are returned.
• AC-4 e2e Given the stealth grey-box course, when a guard off-screen to the left becomes Suspicious from a noise, then a left-edge chevron appears within 1 frame of the state change.
• AC-5 unit Given 12 aware agents, when the model is built, then at most 6 indicators are shown, prioritised by awareness × proximity.

Why

Players read enemy minds through barks ("Who’s there?", "Must be rats…"). With no voice in MVP, text barks plus audio stingers carry this.

What

As a thief, I want guards to mutter and call out as their state changes so that I understand what they think without a UI meter. Bark tables per behaviour profile map (transition, cause) → weighted lines, shown as floating subtitle text near the speaker within hearing range; audio stinger per escalation (placeholder tones) via the audio engine; cooldowns prevent spam. Barks also enable stealth-layer music hooks.

Scope

• In: Bark table schema + selection (seeded)
• In: Floating text renderer with distance fade
• In: Stinger cue ids + cooldowns
• Out: Voice acting (no voice in MVP)
• Out: Adaptive music system (Related: E28/E38 music layering)
• Out: Final bark writing (narrative pass, Related: e39-story-bible tone)

Notes

Stingers: sfx-stealth-suspicious-01, sfx-stealth-alerted-01, sfx-stealth-combat-01, sfx-stealth-lost-01 — placeholder beeps until asset beads land.

• AC-1 unit Given a Unaware→Suspicious transition with cause heard, when a bark is selected with a fixed seed, then it comes from the suspicious.heard table and is deterministic.
• AC-2 unit Given the same agent barked 2 s ago with cooldown 5 s, when a new transition occurs, then no bark is shown.
• AC-3 unit Given a speaker 40 m away beyond hearing range, when a bark would show, then it is suppressed for the player.
• AC-4 integration Given an escalation to Alerted, when it occurs, then exactly one stinger cue is played via the audio engine.
• AC-5 content Given a bark table referencing an unknown transition key, when validated, then validation fails.

Why

Failure creates stories. Being spotted should cause lockdowns, reinforcements, changed patrols or a chase — situations the player can still play through.

What

As a thief, I want getting caught to change the level rather than end my run so that I can improvise an escape. Area alarm levels (0 calm, 1 wary, 2 alarm, 3 lockdown) per zone driven by alarms, bodies found and combat. Data-defined responses per zone and level: lock specified doors, light extra lights, reinforce posts from a reserve pool (pre-placed, activated — not spawned from nowhere), release hounds, shift patrols; decay back after N minutes of calm to a stable “wary” level, never fully reset. Responses recorded in world state.

Scope

• In: Zone alarm model + response data + decay
• In: Integration with doors, lights, reserve creatures
• In: World state persistence
• Out: Crime/bounty system across settlements (Related: E22/E27 crime; not in this epic)
• Out: Specific dungeon zone authoring (E26)

Notes

Reserve pool creatures are placed in level data with reserve: <zone>; activation paths them in from their barracks so the player can see them come.

• AC-1 unit Given zone Z at level 0, when an alarm device activates in Z, then Z becomes level 2 and all level-2 responses fire once.
• AC-2 unit Given a level-3 response lock-doors: [d1, d2], when triggered, then d1 and d2 are locked via the door system and pathfinding reflects it.
• AC-3 unit Given no alarm triggers for 5 sim minutes at level 2, when the decay rule runs, then Z drops to level 1 (never to 0 once an alarm has occurred).
• AC-4 unit Given a reserve creature already killed or knocked out by the player, when reinforcements are called, then it is not activated and no replacement is spawned.
• AC-5 integration Given a save while Z is at level 2, when loaded, then Z is at level 2 with its responses applied.
• AC-6 integration Given the stealth grey-box course, when the player is spotted and the alarm rung, then the course remains completable via at least one alternate route (scripted e2e path succeeds).

Why

Stealth systems need one playable space that exercises light, sound, surfaces, doors, verticality and hiding for e2e tests and tuning playtests.

What

A grey-box level greybox/stealth-course with labelled stations: light/dark gradient hall, surface strip (carpet/stone/gravel/creaky wood/water), door maze (open/closed/locked), vertical hall with rafters and ledges, hiding place gallery, distraction room, patrol yard with 3 fixture guards, alarm bell zone with reserve guards. Includes a baked navmesh, room/portal graph and patrol routes.

Scope

• In: Level data, placeholder props, patrol routes, spawn points
• In: Console tp station <n>
• Out: Final dungeon content (E26)
• Out: Art

Notes

Used by e09 e2e ACs, e09-infiltration-scenarios and e10 thief challenge.

• AC-1 e2e Given the course, when loaded in the browser, then all 8 stations are reachable and no console errors appear.
• AC-2 content Given the course data, when validated, then navmesh, portal graph and patrol routes pass validation.
• AC-3 integration Given tp station 3, when entered, then the player is at the door maze start.
• AC-4 perf Given the course with all guards active, when measured on the reference machine (High), then frame time p95 ≤ 16.7 ms.

Why

Releases must be trusted without reading code. Scripted infiltrations with asserted guard states prove the stealth systems work together and stay working.

What

A suite of ≥ 10 headless scenarios using the AI scenario harness on the stealth course layouts, each with a scripted player path and asserted guard states over time. Scenarios: shadow sneak (never above Suspicious), sprint on gravel (Suspicious ≤ 1 s), closed-door muffling, thrown distraction lure, break LOS and hide (guard ends Searching then postAlert Unaware), body discovery → Alerted, alarm → zone level 2, rafters crossing, hiding-place seen-enter, ally communication chain. Each records a state-hash sequence as a regression baseline.

Scope

• In: Scenario files, baselines, CI job
• In: Readable failure timelines
• Out: Browser e2e (course e2e ACs)

Notes

Scenarios reference fixture creatures only so bestiary tuning never breaks them.

• AC-1 integration Given the shadow-sneak scenario, when run, then no guard exceeds Suspicious for the full 60 s.
• AC-2 integration Given the break-LOS scenario, when run, then the guard sequence is Combat → Searching → Unaware(postAlert) with the Searching phase starting within 5 s of LOS break.
• AC-3 integration Given the closed-door scenario, when the same noise is made with the door open vs closed, then the guard enters Suspicious only in the open case.
• AC-4 integration Given every scenario, when run twice, then state-hash sequences are identical to each other and to the committed baseline.
• AC-5 integration Given a deliberately broken rule (test-only flag doubling light weight), when the suite runs, then at least 3 scenarios fail (the suite actually detects regressions).
• AC-6 perf Given the full suite, when run in CI, then it completes in ≤ 30 s.

Why

Showing where guards believe you are turns last-known position into a tool — the player can see the gap between belief and reality and exploit it.

What

As a thief, I want to see a faint silhouette where the enemies last saw me so that I can plan to double back or set up an ambush. When at least one hunting agent (Searching/Alerted) holds a player LKP and the player is out of its perception, render a translucent silhouette at the LKP in the stance last observed. Disappears when all agents lose the memory or re-acquire. Toggleable.

Scope

• In: Ghost model from aggregated agent memories
• In: Placeholder translucent capsule silhouette
• In: Settings toggle
• Out: Accurate player mesh silhouette (asset integration later)

Notes

Aggregate: show the most recent LKP among hunting agents; if two diverge by > 10 m, show both.

• AC-1 unit Given one Searching agent with an LKP and no perception of the player, when the ghost model is built, then one ghost at that LKP is returned.
• AC-2 unit Given the agent re-acquires the player, when the model is built, then no ghost is returned.
• AC-3 unit Given two hunting agents with LKPs 15 m apart, when built, then two ghosts are returned; if 5 m apart, one.
• AC-4 e2e Given the stealth grey-box course, when the player breaks LOS from an Alerted guard, then a ghost appears at the break point within 250 ms.

Why

mw-e09.1 gives line of sight its own per-collider Occlusion (opaque | transparent | partial:<factor>) supplied through LineOfSight's occlusionOf callback; unregistered colliders are opaque. The world-property vocabulary (v2, mw-e03.31) has opaque/transparent flags but no partial-cover factor, and colliders do not yet map to entities.

What

Map physics colliders to their entities and derive occlusionOf from world properties: opaque -> opaque, transparent -> transparent, neither -> partial with a factor (add a sightTransmittance-style property or material default). Level geometry without properties stays opaque.

Notes

Adding a property is a schema change: bump WORLD_PROPERTIES_VERSION.

Why

The visibility model (mw-e09.2) has a profile term (equipment brightness × cloak × disguise), and selfVisibilityOf reads a stealth.visibilityProfile component, but nothing writes it: items have no brightness and there is no cloak or disguise flag, so every character reads neutral.

What

Add an optional item stealth field (brightness 0–2, conceals: cloak) and have equipment derive each actor's VisibilityProfileComponent (product of worn brightness, cloaked when a concealing item is worn); disguise from the disguise system when it exists.

Scope

• In: Item schema field, equipment-derived profile, register the component in the game world

• Out: Disguise gameplay (its own bead)

• AC-1 unit Given polished plate (brightness 1.3) equipped, when the profile is derived, then brightness is 1.3 and self-visibility rises by that factor.

• AC-2 unit Given a cloak equipped, when derived, then cloaked is true and the tuning's cloak modifier applies.

• AC-3 content Given an item brightness outside 0–2, when validated, then loading fails.

Why

Two tables now describe how visible movement is: the controller's stealth.profiles[stance][gait].visibility (mw-e02.10, published on MovementProfile) and the stealth tuning's stance × motion bands (mw-e09.2), which the visibility model uses. The shipped values agree today (stance × motion reproduces the controller table) but can drift.

What

Pick one source: drop visibility from the controller profiles (MovementProfile.visibility derived from the stealth tuning) or document why both exist. Move the AI stand-in senses onto the visibility model.

Scope

• In: Remove or derive the duplicate, update the stand-in senses, docs

• Out: New tuning values (mw-e09.6)

• AC-1 unit Given the shipped content, when a movement profile and the visibility model are computed for every stance × gait, then they read from one table.

• AC-2 content Given the controller file, when validated, then no second visibility table exists (or a check pins them equal).

Why

Creature footstep noise (mw-e09.5) has no collider to read, so it uses the scene's tagged surface regions and otherwise stone. A skeleton walking on untagged gravel steps as stone (0 dB instead of +8), so it is quieter than the player on the same ground.

What

Give creatures the ground probe the player has: a downward query (or the nav mesh's surface material) that yields the ground material, wet and frozen state for a creature's feet, fed to installFootstepNoise's creatureSurface in startSceneNoise (src/game/creatures/ai.ts).

Acceptance criteria

• AC-1 integration Given a creature walking on an untagged gravel collider, when it steps, then its footstep noise carries the gravel offset (+8 dB).
• AC-2 unit Given a creature on a tagged carpet region over gravel, when it steps, then the region wins, as for the player.
• AC-3 [bench] Per-tick cost with 100 walking creatures stays within the footstep-noise bench budget.

Why

Stealth must be approachable for players with different reaction speeds, hearing or vision without removing its systemic nature.

What

As a player, I want stealth assist options so that I can enjoy infiltration at my own pace. Options: detection speed (50–150%), show indicators for unperceived agents, visual captions for noise events (“Footsteps — left”), stronger visibility gem contrast, extended hiding stagger window. Applied via the whitelisted stealth/AI tuning overrides.

Scope

• In: Settings entries, override application, captions for noise events
• Out: Global accessibility framework (E31)

Notes

Assists alter tuning inputs only; rules stay identical so scenario tests still hold at 100%.

• AC-1 unit Given detection speed 50%, when a constant stimulus accumulates, then time to Suspicious doubles versus 100%.
• AC-2 integration Given noise captions on, when a 40 dB noise is perceived by the player from the left, then a caption with direction left appears within 100 ms.
• AC-3 unit Given an assist value outside the allowed range in settings storage, when loaded, then it clamps to the nearest bound.
• AC-4 integration Given all assists off, when the infiltration suite runs, then baselines are unchanged.

Why

A perfectly executed infiltration should feel as satisfying as a hard fight; only playtesting confirms that.

What

Run 5 playtest sessions on the stealth course and first dungeon grey-box, record telemetry (detections, causes, time in each state, restarts), adjust tuning data, document changes in docs/design/stealth-tuning-log.md.

Scope

• In: Sessions, telemetry capture, tuning edits, log
• Out: New mechanics (file new beads)

Notes

Tuning changes that alter scenario baselines must update them in the same PR with rationale.

• AC-1 manual Given 5 sessions, when players complete the course, then ≥ 4 of 5 rate stealth ≥ 4/5 for “I understood why I was seen”.
• AC-2 integration Given telemetry events for alert transitions, when exported, then each includes cause, state, and time-in-state fields.
• AC-3 manual Given the tuning log, when reviewed, then each change lists the metric that motivated it.

Why

Thief tools are data-driven items so shops, loot and progression can hand them out without code.

What

Item definitions (in the inventory schema) for: lockpick set (durability), coin/pebble/bottle throwables (impact noise, breakable), smoke bomb, flash powder, noisemaker trap, tripwire kit, caltrops, rope, water flask (extinguish), sap/blackjack (non-lethal), disguise items referencing faction. Each carries world properties (weight, fragile, noise) and tags consumed by E10 features.

Scope

• In: Item data files + tests
• In: Tags: tool:lockpick, throwable, trap:*, disguise:<faction>
• Out: Loot/shop placement (E18/E20)
• Out: Icons (asset beads)

Notes

Placeholder icon ids: icon-tool-lockpick, icon-throw-coin, icon-throw-bottle, icon-smoke-bomb, icon-trap-noisemaker, icon-trap-tripwire, icon-caltrops, icon-water-flask, icon-sap.

• AC-1 content Given each thief item file, when validated, then it passes the inventory item schema and carries its required tags.
• AC-2 content Given a throwable without an impact noise value, when validated, then validation fails.
• AC-3 unit Given a lockpick set with durability 10, when resolved, then durability is 10 and it stacks to max 5 sets.
• AC-4 content Given every thief item, when the content harness runs, then each has at least one automated test (spawn into inventory + use).

Why

Thrown distractions are the thief’s first verb in the progression ladder (Sneak → distractions → …) and a classic stealth tool.

What

As a thief, I want to throw a coin or bottle so that guards look away or walk over to investigate. Any throwable item or physics object can be aimed (trajectory preview arc) and thrown; on impact it emits a noise event = item impact dB + surface offset; fragile items shatter (louder, leave shards as glass-shards surface). Guards perceive the propagated noise and investigate the perceived position. Thrown object remains in the world (anomaly if found near a post repeatedly — hook).

Scope

• In: Throw aim + arc preview + release
• In: Impact noise via sound propagation + surface
• In: Pickup-able thrown objects (non-fragile)
• Out: Combat throwing damage (E04)
• Out: Arrows (E05)

Notes

Uses physics objects for the throw; the noise rule is property-based (throwable, fragile), never item-specific.

• AC-1 unit Given a coin (impact 45 dB) landing on stone 10 m from a guard with threshold 30 dB, when propagated, then the guard receives a heard-noise stimulus and enters Suspicious.
• AC-2 unit Given a bottle tagged fragile landing, when it hits any surface, then it shatters, emits the fragile-break loudness and adds a glass-shards surface patch.
• AC-3 unit Given a coin landing on carpet, when propagated to the same guard, then perceived loudness is reduced by the carpet offset and the guard stays Unaware.
• AC-4 integration Given the distraction-room scenario, when the player throws a bottle behind a guard, then the guard turns and investigates the impact location within 2 s.
• AC-5 unit Given a throw into water, when it lands, then a splash noise kind is emitted and the object sinks (not retrievable unless shallow).
• AC-6 e2e Given the stealth grey-box course, when the player aims, then a trajectory arc renders and the landing marker matches the actual landing point within 0.5 m.

Why

Extinguishing lights lets thieves create their own darkness, and any system that affects light (water, wind, frost magic) should work too.

What

As a thief, I want to snuff candles and douse torches so that I can create shadows where I need them. Light sources with property extinguishable go out from: interact (pinch/snuff, quiet), any wet/water effect via element propagation (thrown water flask, rain), strong wind/gust effects, frost. ignitable light sources relight from any fire source. Light-field updates immediately; extinguishing emits a small noise and creates an anomaly (e09-anomaly-detection).

Scope

• In: Light source state + property-based rules
• In: Interact snuff action
• In: Light-field update + anomaly record
• Out: Light-field implementation (e03-light-field)
• Out: Water arrows (E05) — works automatically via property

Notes

Magical lights may be tagged non-extinguishable; that is data.

• AC-1 unit Given any extinguishable light within a water/wet effect volume, when the effect applies, then the light turns off and the light-field level at 2 m drops to the ambient value within 1 tick.
• AC-2 unit Given an ignitable unlit torch in contact with any fire source, when 0.5 s passes, then it relights.
• AC-3 unit Given a light tagged non-extinguishable, when water hits it, then it stays lit.
• AC-4 integration Given the player snuffs a candle via interact, when complete, then a ≤ 15 dB noise is emitted and an anomaly light-out is registered for that light.
• AC-5 integration Given a guard walking past a torch the player extinguished, when perceived, then the guard enters Suspicious with cause light-out.

Why

Opportunistic combat is the thief’s combat fantasy: ambushes and takedowns reward patience and positioning.

What

As a thief, I want to backstab or knock out an unaware enemy from behind so that a clean infiltration can thin the guard without a fight. Conditions: target alert state ∈ {Unaware, Suspicious}, player within rear arc (±60°) and 1.5 m, target does not perceive the player this think. Lethal backstab: damage multiplier from damage model (lethal on humanoids by default). Non-lethal (sap/choke): target becomes unconscious for a data-defined duration (default 10 min sim) unless woken. Creature schema flags (immune-to-knockout for undead/constructs, no-backstab for huge creatures) gate the verbs. Takedown emits noise (choke 20 dB, backstab 35 dB). Failed attempts (target turns) cause a scuffle: target enters Combat staggered.

Scope

• In: Prompt availability, execution, outcomes, noise
• In: Creature flags gating
• In: Failure scuffle outcome
• Out: Environmental takedowns (e10-environmental-takedowns)
• Out: Animations (placeholder snaps)

Notes

Unconscious bodies are carriable (e10-body-carry-hide) and anomalies (e09-anomaly-detection).

• AC-1 unit Given an Unaware humanoid and the player 1 m behind within 60°, when takedown is evaluated, then both backstab and knockout verbs are available.
• AC-2 unit Given the same target in Investigating or higher, when evaluated, then no takedown verb is available.
• AC-3 unit Given a creature flagged immune-to-knockout, when evaluated, then only backstab is available.
• AC-4 unit Given a knockout, when executed, then the target is unconscious for 600 s sim and a 20 dB noise is emitted at its position.
• AC-5 unit Given the target perceives the player during the 0.5 s wind-up, when the takedown resolves, then it fails, the target enters Combat and is staggered for 0.8 s.
• AC-6 integration Given the patrol-yard scenario, when the player knocks out a guard out of sight of others, then the other guards remain Unaware.

Why

Hiding bodies is how thieves cover their tracks; a body left in the light should cause the alarm the player earned.

What

As a thief, I want to pick up and hide bodies so that patrols do not discover what I did. Pick up unconscious/dead humanoid-size creatures (size class limit), carry at reduced speed (−40%) and no sneak-sprint, higher visibility (+profile), drop or place into hiding places with body-capacity (chests, wardrobes, deep shadow, water). Placed bodies inside containers are invisible to anomaly detection; bodies in darkness depend on visibility. Unconscious creatures carried keep their wake timer.

Scope

• In: Carry state, movement modifiers, drop/place
• In: Body containers
• In: Integration with anomaly detection
• Out: Ragdoll physics beyond placeholder (Related: physics/animation)
• Out: Looting bodies (e17 / E18)

Notes

Dropping a body emits a thud noise scaled by surface (e.g. 35 dB on stone).

• AC-1 unit Given an unconscious guard, when the player picks it up, then player speed is ×0.5 (ADR-0003, matches mw-e02.17) and the visibility profile term increases by the configured amount.
• AC-2 unit Given the player carrying a body, when placed into a container with body capacity 1, then the body is hidden and the container is full.
• AC-3 unit Given a body placed in a container, when a guard perceives the container, then no body anomaly is raised unless the guard opens it.
• AC-4 unit Given a creature larger than the carry size limit, when pickup is attempted, then it is refused with reason too-heavy.
• AC-5 unit Given the unconscious carried guard’s wake timer expires, when it wakes while carried, then it is dropped, enters Alerted and the player is staggered 0.5 s.

Why

The constitution prefers world systems over disconnected minigames. Lockpicking must keep the player in the world, exposed to guards and noise, while still involving skill.

What

Time-box 1 day. Question: how can lockpicking be an in-world, interruptible action with a light skill layer (e.g. timing tension clicks while the camera stays third-person) instead of a full-screen minigame? Prototype 2 variants in the grey-box (A: timed hold with pin-click timing windows; B: pure duration by lock quality vs tool/skill with noise risk), evaluate readability, tension with nearby guards, accessibility (a no-skill option), and progression hooks. ADR with recommendation.

Scope

• In: Two prototypes in spikes/lockpicking/
• In: ADR with lock data shape (quality, pins, noise per failure, jam chance)
• Out: Production lockpicking (e10-lockpicking)

Notes

Must remain compatible with sorcerer/knight alternatives (Unlock spell, bashing) via the same lock component from e03-doors-locks-switches.

• AC-1 manual Given the ADR, when reviewed, then it records the choice, rejected alternatives, and a data shape for lock quality.
• AC-2 manual Given 3 internal testers, when trying both prototypes near a patrolling guard, then preferences and time-to-open are recorded in the ADR.
• AC-3 manual Given the ADR, when read, then it specifies an accessibility mode that removes timing requirements.

Why

Locked doors and chests are the thief’s signature alternate routes; picking them in the world while guards patrol creates tension.

What

As a thief, I want to pick locks while staying aware of my surroundings so that opening a door is a tense, skilful choice rather than a menu. Implements the ADR: locks carry quality (1–5), pin count, jam chance; picking is an in-world action (player stays visible and hearable, camera third-person) consuming lockpick durability; skill layer per ADR with an accessibility no-timing mode. Failures emit small noises and may break a pick; jamming a lock makes it unpickable (bash/magic still work). Progression hooks raise max quality.

Scope

• In: Lock component extension, pick action, noise, durability, jam
• In: In-world UI overlay (placeholder)
• In: Accessibility mode
• Out: Door/lock base mechanics (e03-doors-locks-switches)
• Out: Unlock spells/bashing (E06/E04, same component)

Notes

Default noise per failed pin 18 dB; broken pick 25 dB.

• AC-1 unit Given a quality-2 lock and one lockpick set, when the pick action completes successfully, then the lock is unlocked and durability decreases by 1.
• AC-2 unit Given a failed pin attempt, when resolved, then an 18 dB noise event is emitted at the lock.
• AC-3 unit Given a lock jams (seeded chance), when evaluated, then further pick attempts are refused with reason jammed while other unlock methods still work.
• AC-4 unit Given the player is attacked or moves > 0.5 m while picking, when resolved, then the action is interrupted and progress is lost.
• AC-5 unit Given lock quality above the player’s unlocked max, when the prompt is evaluated, then picking is not offered and the prompt says why.
• AC-6 integration Given accessibility no-timing mode, when picking a quality-3 lock, then it completes after the documented duration without input timing.
• AC-7 e2e Given the stealth grey-box door maze, when the player picks a locked door, then the door opens and the navmesh link becomes traversable.

Why

Pickpocketing creates emergent solutions: steal a key instead of picking a lock, plant evidence, lift a guard’s disguise pass.

What

As a thief, I want to lift items from, or plant items on, unaware creatures so that I can solve problems without violence. Available when target is Unaware/Suspicious, not perceiving the player, within 1.2 m behind/side arc. Opens a compact in-world list of visible pockets (creature inventory items flagged pocketable). Each attempt takes time by item weight; success chance falls with target awareness and item weight; a failed attempt raises target awareness sharply (Alerted if it turns and sees you). Theft is recorded (disposition/faction events for witnesses).

Scope

• In: Pickpocket and plant actions, success model, witness events
• In: Creature inventory integration
• In: Compact UI (placeholder)
• Out: Creature loot tables (e18-loot-tables)
• Out: Crime/bounty (Related: crime system)

Notes

Success probability formula documented in tuning data; seeded RNG.

• AC-1 unit Given an Unaware guard with a pocketable key, when the player attempts from behind at 1 m, then the item transfers on success and the attempt takes the weight-scaled duration.
• AC-2 unit Given a failed attempt, when resolved, then the target’s awareness increases by the configured spike and it turns toward the player.
• AC-3 unit Given a witness guard perceives the theft, when resolved, then a witnessed-theft event fires applying disposition and faction modifiers.
• AC-4 unit Given a target with no pocketable items, when the prompt is evaluated, then the verb is offered but the list shows “Nothing to take” (still allows planting).
• AC-5 unit Given the target walks out of range mid-attempt, when resolved, then the attempt cancels without a transfer or awareness spike.
• AC-6 integration Given the player plants an item on a guard, when saved and loaded, then the guard’s inventory contains the planted item.

Why

Triggered traps should create new situations rather than reloads; spotting and disarming them is thief expertise.

What

As a thief, I want to notice and disarm pressure plates, tripwires and dart traps so that I can open safe routes — and accept the chaos when I fail. Traps have a concealment value; the player perceives them via a proximity/light-based check (thieves have a bonus; sneaking slower improves it) which reveals a subtle highlight. Disarm is an in-world timed action (like lockpicking) with a failure chance by trap quality; failure triggers the trap. Disarmed traps can be picked up (become placeable items where data allows).

Scope

• In: Trap detection check, reveal highlight, disarm action, fail-trigger
• In: Salvage to item
• Out: Trap mechanics/effects themselves (e03-triggers / e03-doors-locks-switches)
• Out: Dungeon trap placement (E26)

Notes

Detection uses light-field at trap position: dark traps are harder to spot.

• AC-1 unit Given a trap with concealment 0.5 in light 0.8, when a thief sneaks within 3 m, then it is detected within 1 s.
• AC-2 unit Given the same trap in light 0.1, when a knight walks within 3 m, then it is not detected.
• AC-3 unit Given a disarm attempt that fails (seeded), when resolved, then the trap triggers its normal effect.
• AC-4 unit Given a successfully disarmed pressure plate, when a creature steps on it, then nothing triggers.
• AC-5 integration Given a disarmed tripwire, when salvaged, then a tripwire kit item is added to inventory.

Why

Escape tools turn failed stealth into a chase instead of a death — the thief’s disengagement fantasy.

What

As a thief, I want to throw a smoke bomb so that I can break line of sight and slip away. Smoke bomb creates a gas volume (via element propagation, drifts and dissipates over 12 s) tagged occlusion: partial:0.1 for LOS; creatures inside have sight disabled and cough (noise). Flash powder: instant stun for creatures facing it within 5 m (sight-dependent — blind/life-sense creatures unaffected). Losing LOS feeds LKP rules, so guards search where you vanished.

Scope

• In: Smoke gas volume + LOS occlusion + dissipation
• In: Flash stun rule by sense profile
• In: Integration with memory/LKP
• Out: Gas propagation engine (e03-element-propagation)
• Out: VFX art (placeholders)

Notes

Smoke is a gas: a Gust effect should clear it via the same gas rules (property-based).

• AC-1 unit Given a smoke volume between guard and player, when LOS is queried, then visible fraction ≤ 0.1.
• AC-2 unit Given a smoke volume, when 12 s pass without wind, then it has dissipated and LOS is restored.
• AC-3 unit Given flash powder within 5 m and a guard facing it, when detonated, then the guard is stunned 3 s; a creature without sight senses is not stunned.
• AC-4 integration Given a Combat guard chasing the player, when the player throws smoke and runs 10 m sideways, then the guard enters Searching with LKP at the smoke cloud, not the player’s position.
• AC-5 unit Given any gas-clearing wind effect through the smoke volume, when applied, then the volume moves/dissipates per gas rules.

Why

Milestone m2 requires the thief fantasy proven in grey-box; a perfect infiltration must feel as good as a hard fight.

What

A grey-box vault infiltration greybox/thief-vault: guarded courtyard, lit halls, a locked vault with a key carried by a patrolling guard, climbable exterior wall, alarm bell, hiding places. At least 3 distinct solution routes (pickpocket key; climb + pick window lock; distraction + extinguish + takedown), each covered by a scripted e2e/scenario test. Records a “ghost” (never detected) flag on completion.

Scope

• In: Level data, routes, scripted solution tests, completion flag
• Out: Final dungeon (E26)
• Out: Art

Notes

Fixture creatures only.

• AC-1 integration Given route A (pickpocket), when the scripted scenario runs, then the vault opens and no guard exceeds Suspicious.
• AC-2 integration Given route B (climb + lockpick window), when run, then the vault is reached and the completion flag is set.
• AC-3 integration Given route C (distract + extinguish + takedown), when run, then the vault opens with one unconscious guard hidden and no alarm.
• AC-4 integration Given a scripted failure (player spotted in the courtyard), when run, then the alarm escalates the zone and the scripted escape via smoke + hiding place still completes the level.
• AC-5 manual Given 3 playtesters, when completing the vault undetected, then ≥ 2 rate satisfaction ≥ 4/5.

Why

Disguises are a later thief verb in the progression ladder and create social infiltration moments.

What

As a thief, I want to wear a faction’s uniform so that I can walk among guards if I behave. Equipping a disguise:<faction> item makes that faction’s members treat the player as disguised instead of intruder: sight stimuli accumulate a separate suspicion meter driven by proximity (< 3 m), suspicious behaviour (running, crouching, climbing, being in restricted zones, being seen committing crimes, carrying bodies), and observer traits (officers and dogs see through faster; sees-through-disguise flag). Suspicion full → disguise blown for that zone (remembered).

Scope

• In: Disguise state, per-observer suspicion rules, restricted zone data, blown-state memory
• Out: Dialogue with disguised player (E22)
• Out: Uniform art (placeholder recolour)

Notes

Suspicion rules are tuning data; restricted zones are level regions.

• AC-1 unit Given the player wears a disguise of faction A and walks at normal speed 8 m from an A guard in a public zone, when 30 s pass, then the guard remains Unaware.
• AC-2 unit Given the same player sprints within 3 m, when evaluated, then disguise suspicion accumulates at the running rate and the guard enters Suspicious within 3 s.
• AC-3 unit Given an observer flagged sees-through-disguise, when it perceives the player, then the disguise is ignored.
• AC-4 unit Given the disguise is blown in zone Z, when the player re-enters Z still disguised, then Z’s guards treat the player as an intruder.
• AC-5 unit Given a faction B guard, when perceiving the player in faction A disguise, then normal intruder rules apply.

Why

Traps let the thief shape the battlefield ahead of time — dirty tricks and disengagement are the thief combat fantasy.

What

As a thief, I want to place traps so that I can lure, slow or warn about enemies. Place trap items on valid surfaces with a placement preview: noisemaker (timed or triggered noise event), tripwire (trigger volume between two anchors → alarm noise for player, or trip/stagger), caltrops (slow + damage zone, noise when stepped). Traps are world objects with properties (flammable tripwire burns, etc.), perceivable by AI (visibility check); perceiving an armed trap makes AI avoid it and become Suspicious.

Scope

• In: Placement preview + validity, trap objects, trigger logic via e03-triggers
• In: AI avoidance of perceived traps (nav cost)
• Out: World/dungeon traps and disarming (e10-trap-disarm)
• Out: Trap art (placeholders)

Notes

Creatures can trigger player traps; the player can too (friendly fire).

• AC-1 unit Given a caltrops item and a flat valid surface, when placed, then a slow zone (×0.5 speed) with damage 2/s exists at the location.
• AC-2 unit Given a placement on a slope > 30° or in water, when previewed, then placement is invalid with reason.
• AC-3 unit Given a tripwire between two anchors, when any creature crosses it, then its trigger fires once and it becomes spent.
• AC-4 unit Given a guard perceives an armed caltrops zone, when pathing, then its path avoids the zone if an alternative exists and it enters Suspicious.
• AC-5 unit Given a tripwire tagged flammable in a fire volume, when 1 s passes, then it burns and is removed without triggering.

Why

Anti-grind rule: progression unlocks verbs, not +12%. Silent movement should grow into new capabilities.

What

As a thief, I want to unlock ways to move more silently so that I can take bolder routes. Capability unlocks via the progression model: soft-step (sneak gait ignores creaky-board creaks and gravel offset), muffled-gear (equipment noise halved while sneaking), shadow-dash (short burst of sneak-speed movement at walk noise while in light ≤ 0.2). Each is a data-defined modifier applied to the footstep/visibility rules.

Scope

• In: Three capability modifiers as data + rule hooks
• Out: Progression UI and unlock sources (E19)

Notes

Modifiers alter inputs to e09-footstep-noise and e09-visibility-model only; rules stay pure.

• AC-1 unit Given soft-step unlocked, when sneaking on wood-creaky, then no creak event is ever emitted over 1,000 seeded steps.
• AC-2 unit Given muffled-gear and equipment noise +10 dB, when sneaking, then the footstep includes +5 dB equipment noise.
• AC-3 unit Given shadow-dash in light 0.5, when activated, then it is refused because light exceeds 0.2.
• AC-4 unit Given no unlocks, when evaluated, then footstep results equal the base rule exactly.

Why

Information is the thief’s currency; peeking lets players scout routines before committing.

What

As a thief, I want to lean around corners and peek through keyholes so that I can scout without exposing myself. Lean shifts camera and exposes only the head sample point to LOS (reduced visibility). Keyhole peek on doors with a keyhole flag gives a narrow view of the next room with the player fully hidden from that room.

Scope

• In: Lean input + camera offset + LOS sample override
• In: Keyhole peek interaction
• Out: Camera system (e02-camera)

Notes

During lean, only the head sample point counts for LOS (1 of 4).

• AC-1 unit Given the player leaning, when LOS is queried by an observer, then only the head sample point is considered.
• AC-2 unit Given a door without a keyhole, when interaction options are built, then peek is not offered.
• AC-3 integration Given a keyhole peek, when active, then the camera renders from the keyhole position and visibility to observers in the other room is 0.
• AC-4 unit Given the player is attacked while peeking, when damage applies, then peek ends immediately.

Why

The progression ladder ends at environmental takedowns; the world itself becomes the thief’s weapon, and it must be systemic.

What

As a thief, I want to drop a chandelier, cut a hanging load or push a crate from a ledge onto an enemy so that the world does my work. Any object with suspended property released (cut rope, shot chain, lever) falls under physics; impact damage = f(mass, fall height) via the damage model. Deaths/knockouts caused by environment are tagged accident so witnesses raise lower suspicion (Suspicious/Investigating instead of Alerted) unless they perceived the player causing it.

Scope

• In: Falling-object damage rule by mass/height
• In: Accident attribution + AI response weights
• Out: Specific set-pieces (E26 content)
• Out: Rope cutting by arrows (E05, works via property)

Notes

Property-based: never "chandelier" in rules.

• AC-1 unit Given any object tagged suspended with mass 80 kg released from 5 m onto a creature, when impact resolves, then damage equals the documented mass×height curve.
• AC-2 unit Given a creature killed by a falling object with no witness perceiving the player within 3 s, when attributed, then the event is tagged accident.
• AC-3 unit Given a witness to an accident, when evaluated, then it enters Investigating (not Alerted).
• AC-4 unit Given a witness perceived the player releasing the object, when evaluated, then it enters Alerted with the player LKP.

Why

mw-e11.4 (PR #162) shipped a pure-TypeScript navmesh bake and A* and recorded the choice as ADR-0006, status Proposed.

What

Owner reads docs/adr/0006-navmesh-and-pathfinding.md and accepts (flip status to Accepted) or rejects (file a migration bead to recast-navigation-js).

AC-1 manual ADR-0006 status is Accepted or Rejected with the owner's decision recorded.

Why

Briar Wolves (0.4 m) and Brother Horn are wider than the humanoid radius (0.35 m) every scene is baked for (mw-e11.4, ADR-0006), so they can clip corners and squeeze through gaps they should not fit; goblins (0.3 m) path more conservatively than they need to.

What

Bake each navigated scene for a small set of agent size classes (e.g. 0.3 / 0.35 / 0.6 m radius, matching heights) and pick the navmesh by a creature's NavAgent radius at query time.

Scope

• In: bake settings per size class in the navmesh data or a sibling file, query-time selection, pnpm nav:check coverage
• Out: crowd simulation

Notes

Related: src/sim/nav/bake.ts, src/tools/navmesh/bake.ts, ADR-0006 consequence 'One agent size per navmesh'.

• AC-1 unit Given a 0.6 m gap, when a 0.35 m agent and a 0.8 m-radius agent request a path through it, then only the smaller one's path uses the gap.
• AC-2 integration Given the testbed, when the bake runs for every size class, then all navmeshes are produced in <= 10 s and content validation passes.

Why

Sense profiles can give creatures smell and a magic-sense channel, but perception (mw-e11.5) has no handler for either, so a wolf cannot smell the thief and a warded creature cannot feel a spell: those senses silently perceive nothing.

What

Register a smell handler (range, wind sensitivity, scent trails from the smell profile) and a magic-sense handler (active spells and enchanted items as sources) in the special-sense channel registry, producing typed percepts. Needs scent sources and magic sources in the sim.

Scope

• In: Smell and magic-sense handlers; their percept kinds or reuse of sensed-life

• Out: Wind simulation; scent trail simulation beyond a hook

Notes

Channel registry is src/sim/perception/channels.ts; content channels are SPECIAL_SENSE_CHANNELS.

• AC-1 unit Given a creature with smell range 10 m downwind of the player 8 m away, when evaluated, then a percept from smell is produced; upwind it is not.

• AC-2 unit Given a creature with magic-sense and a lit spell 5 m away behind a wall, when evaluated, then a percept is produced when requiresLineOfSight is false.

Why

mw-e11.17's AI debug overlay lists claimed search points per agent, but no search system claims points yet (search patterns are mw-e11.11), so the snapshot has nothing to draw.

What

When mw-e11.11 lands, add each agent's claimed search points to the AI introspection snapshot (src/sim/debug/ai-snapshot.ts) and draw them in src/render/debug/ai-overlay.ts (markers tinted by alert state, with the claiming agent in the label).

Acceptance criteria

• AC-1 unit Given an agent searching with claimed points, when the snapshot is built, then it lists those points.
• AC-2 e2e Given a searching guard and ai.debug on, then its claimed points render with no console errors.

Why

mw-e12.14 made a loaded world continue exactly as one never saved (creature components, brains and perception bookkeeping are all saved). One gap remains, accepted in mw-e11.4: the navmesh path request queue (src/sim/nav/queue.ts) lives in the navigation port, not the world. A save taken while an agent's path request waits or is half-searched loads with its nav.route 'lost': the agent resubmits it (same start and goal), under a new request id, and the search starts over, so the resumed run's state hash differs from the uninterrupted run's for a few ticks onward. Nothing breaks for the player, but the resume-determinism property (and any golden that saves mid-chase) does not hold with navmesh travel.

What

Make the queue's state part of the world so saves carry it: its next request id, the waiting requests in order (request + the budgets each search step was given, so a search can be rebuilt by replaying its steps) and finished results not yet taken. Load rebuilds the queue from that state, and the 'lost route resubmits' path stays only for saves from before.

Acceptance criteria

• AC-1 integration Given navmesh AI travel with a tight per-tick budget and a save taken while a request is half-searched, when loaded into a fresh world and run on, then the state hash at the end equals the uninterrupted run's.
• AC-2 unit Given a save from before this change with a pending route, when loaded, then the agent resubmits and reaches its goal (no regression).

Why

Distractions only work if guards go and look. Investigation is the core lever for thrown objects, noises and glimpses.

What

As a thief, I want a guard to walk over and check a noise or glimpse so that I can lure him away from his post. In Suspicious: stop, turn toward stimulus, text bark cue. In Investigating: path to stimulus position at walk (or cautious) gait, look around on arrival with a scan pattern, inspect the source object if perceived (e.g. a thrown bottle), then decide: nothing found → de-escalate; more stimuli → escalate. Multiple stimuli: most recent strong one wins; weaker stimuli queue up to 3.

Scope

• In: Investigate behaviour, scan pattern, source inspection, stimulus queue
• In: Cautious gait when postAlert
• Out: Group investigation roles (e11-communication)
• Out: Search around LKP (e11-search-behaviour)

Notes

Investigation position is the perceived position (may be offset by sound routing), not the true source.

• AC-1 unit Given an Investigating guard with target position 12 m away, when simulated, then it arrives within (12 / walkSpeed + 1) s and performs a scan of ≥ 3 look directions.
• AC-2 unit Given a second, louder stimulus while investigating, when perceived, then the guard retargets to the new position and the first is queued.
• AC-3 unit Given an unreachable stimulus position (on a roof), when investigating, then the guard moves to the closest reachable point, looks toward the position for the scan duration, then de-escalates.
• AC-4 integration Given a scenario where a noise occurs behind a closed door, when the guard investigates, then it opens the door (if unlocked) before scanning the room.
• AC-5 unit Given a queue already holding 3 stimuli, when a 4th weaker stimulus arrives, then the weakest is dropped.

Why

A convincing search — fanning out, checking corners and wardrobes — makes escaping feel earned, and gives up plausibly so failure does not force a reload.

What

As a thief, I want searching guards to check plausible places near where they lost me so that hiding well matters and hiding badly gets me caught. Searching generates candidate points around LKP (+ bounded prediction): nav-reachable points ranked by distance, cover/darkness, and whether the point is out of sight from LKP; plus hiding spots within radius. Agents claim points (no two agents search the same point), check hiding spots with a probability based on whether they saw the target enter it, expand radius over time, then give up.

Scope

• In: Candidate generation, ranking, claiming, radius expansion
• In: Hiding-spot checks with probability rules
• In: Diligence trait scaling of duration
• Out: Hiding place mechanics (e09-hiding-places)
• Out: Alarm raising (e11-communication)

Notes

Probability of checking a hiding spot: 1.0 if target was seen entering; otherwise base 0.25 scaled by proximity to LKP. Seeded RNG.

• AC-1 unit Given a Searching guard with LKP L, when candidates are generated, then all lie within the current radius, are nav-reachable and the first 3 are sorted by the ranking function.
• AC-2 unit Given two guards searching the same LKP, when they claim points, then no point is claimed by both.
• AC-3 unit Given the guard saw the target enter a hiding spot, when searching, then it checks that spot first with probability 1.0.
• AC-4 unit Given a hiding spot not seen being entered, 10 m from LKP, when evaluated with a fixed seed, then the check decision matches the documented probability formula.
• AC-5 unit Given no reachable candidates (LKP on an isolated ledge), when searching, then the guard looks toward LKP from the nearest point, then de-escalates at the search timeout without error.
• AC-6 integration Given a scenario where the player hides in darkness 8 m from LKP and stays still, when a single guard searches for 60 s, then the guard ends Searching without entering Combat in the scripted seed.

Why

Enemies that communicate make a single mistake escalate believably, and give players targets — silence the shouter, cut the bell rope.

What

As a thief, I want guards to call for help through the world rather than telepathy so that I can prevent or exploit alarms. Channels: shout (emits a loud noise event via sound propagation carrying a report: LKP, state), visual signal (allies in LOS within 25 m receive the report), alarm device (agents path to a nearest alarm object and activate it via the interaction/mechanism system, broadcasting Alerted to a faction zone). Only allies (faction relation ally/friendly) accept reports. Group roles in Alerted/Searching: one caller, others flank/search.

Scope

• In: Report message type, shout/visual/alarm channels
• In: Faction filtering
• In: Basic role assignment (caller, searcher, guard-post-holder)
• Out: Reinforcement spawning (e09-failed-stealth-outcomes)
• Out: Speech/voice (text barks only — e09-alert-barks)

Notes

Alarm objects are world mechanisms with an alarm property; if destroyed or jammed, the AI must fall back to shouting. Related: alarm bell mechanism may be authored with E03 switches.

• AC-1 unit Given a guard entering Alerted, when it shouts, then a noise event of the configured loudness is emitted and allies whose perceived level exceeds their hearing threshold receive the report.
• AC-2 unit Given a guard behind 2 closed doors whose shout is attenuated below an ally’s threshold, when shouting, then that ally receives nothing.
• AC-3 unit Given a creature of a hostile faction within hearing, when a shout occurs, then it does not receive the report (it may still perceive the noise as a stimulus).
• AC-4 unit Given an alarm object that has been disabled (jammed or destroyed), when a guard attempts to use it, then it falls back to shouting within 2 s.
• AC-5 integration Given a scenario with 3 guards and one alarm bell, when the first guard spots the player and reaches the bell, then all guards in the bell’s zone enter Alerted within 1 tick of activation.
• AC-6 unit Given a guard is knocked out mid-path to the alarm, when the tick runs, then no alarm activation occurs.

Why

Enemies that retreat, regroup or flee make fights and scares feel alive and allow non-lethal victories.

What

As a player, I want frightened or wounded creatures to disengage so that encounters end in more ways than death. Consumes Flee/Cower intents from e12-fears-morale and health thresholds: flee away from the fear source toward allies, safe zones or a lair point; cower when cornered; regroup and re-enter Alerted/Searching when morale recovers. Fleeing creatures may shout (communication) and drop carried items (hook).

Scope

• In: Flee destination scoring, cower, regroup
• In: Surrender intent event (hook for E14 dialogue)
• Out: Morale model (e12-fears-morale)
• Out: Surrender dialogue (E14/E22)

Notes

Flee destination scoring: distance from threat, ally proximity, darkness preference per creature.

• AC-1 unit Given a creature with a Flee intent and threat at T, when fleeing, then after 3 s its distance from T has increased and its path does not pass within 2 m of T.
• AC-2 unit Given a fleeing creature with no escape path (dead end), when cornered, then it enters Cower within 1 s.
• AC-3 unit Given morale recovers above the regroup threshold, when the intent clears, then the creature enters Searching with the threat’s last-known position.
• AC-4 unit Given a creature tagged fearless (morale disabled), when at 1 health, then it never flees.

Why

Failure creates stories: after an alert guards should go back to work, but the world should remember — lit torches, doubled posts, grumbling barks.

What

As a thief, I want guards to settle back down after an alert but stay changed so that my mistakes reshape the level instead of resetting it. On de-escalation: path back to routine, apply post-alert adjustments from data (e.g. relight extinguished lights on route, close opened doors, patrol in pairs, add a checkpoint post, carry a torch), keep heightened awareness for a duration. Adjustments are data-driven per behaviour profile and recorded as world facts.

Scope

• In: Return behaviour, post-alert adjustment list, persistence via world state
• In: Barks cue on return
• Out: Level lockdowns/reinforcements (e09-failed-stealth-outcomes)

Notes

Adjustments list example: relight-seen-extinguished-lights, close-seen-open-doors, pair-up, carry-torch.

• AC-1 unit Given a guard de-escalates with postAlert=true, when simulated, then it resumes its route within 5 s and awareness rate multiplier is 1.5 for 300 s, then 1.0.
• AC-2 unit Given a guard with the relight-seen-extinguished-lights adjustment that perceived an extinguished torch during the alert, when returning, then it paths to the torch and relights it before resuming.
• AC-3 unit Given the torch was destroyed (not merely extinguished), when the adjustment runs, then the guard skips it and switches to carry-torch if declared.
• AC-4 integration Given a save after an alert with pair-up applied, when loaded, then the paired patrol persists.

Why

Stealth feel is tuned constantly; all numbers must be data so designers can iterate without code and tests can pin defaults.

What

A AiTuning schema and default files per behaviour profile: awareness thresholds per state, accumulation rates per sense, grace and decay timers, state timeouts (Suspicious 6 s, Investigating 20 s, Searching 60 s, Alerted 120 s defaults), post-alert multiplier and duration, search radius growth, token budgets, think rates. Profiles inherit (guard-default → guard-lazy). Hot-reload in dev.

Scope

• In: Schema, inheritance, defaults, hot reload in dev builds
• In: Pinned default-value tests
• Out: Sense ranges (e12-sense-profiles)
• Out: Stealth emitter tuning (e09-stealth-tuning-data)

Notes

Changing defaults requires updating pinned tests deliberately — that is the point.

• AC-1 content Given the default tuning files, when validated, then every state has a positive timeout and thresholds are strictly increasing Suspicious < Investigating < Detected.
• AC-2 unit Given guard-lazy inheriting from guard-default overriding only Searching timeout 30 s, when resolved, then all other values equal the parent.
• AC-3 unit Given an inheritance cycle, when loaded, then loading fails naming the cycle.
• AC-4 integration Given a dev build, when a tuning file is edited, then running agents use the new values within 1 s without reload.
• AC-5 unit Given the shipped defaults, when the pinned-default test runs, then each documented default matches.