# Living Scenes · A scene somebody wrote down · receipt · 13 September 2026

Chapter three of _Creation_, and the track's first merged chapter. Delivered in
the Eve edition on 13 September 2026. `living-scenes-runtime` (16 slides) and
`living-scenes-composition` (19) become one forty-three-slide chapter under the
new id `living-scenes-score`. The brief is
[authoring/living-scenes-score-chapter.md](../../authoring/living-scenes-score-chapter.md).
Machine-readable state is in [state.json](state.json).

## What was delivered

Forty-three slides: thirty-five inherited slides rewritten in place and eight
new — a cover, six section dividers and a close. Nothing merges away and every
slide id is kept, per the Phase D rule for merged chapters.

`content/46-living-scenes-composition.json` held nothing but this chapter's
slides and **retires**. `content/08-living-scenes.json` gives up thirteen of its
nineteen and keeps six governance slides;
`content/24-creation-deepdives.json` gives up three and keeps six. The decks of
both source guides are removed.

Two legacy graph entries retire with this chapter — `scene-compose-boundaries`
and `scene-compose-reorder` — which closes **two of the five** that remained in
Living Scenes guides. Three remain there (`scene-tech-layers`,
`scene-compat-pipeline`, `scene-compat-gating-order`), in the two chapters this
track has not reached. Library-wide, counted from the builder's own output,
**twelve** slides still render as legacy graphs: the other nine are in privacy,
Telegram, payments, native, Veritas, Metis, operations and V2 guides that
Phase D has not rewritten.

## Evidence

Eight probes, ninety-one verdicts, all executed at the pin on this host. Every
module in this chapter is pure: it returns a decision and performs nothing.

- **Two validators, seventeen cases, thirteen disagreements.** The Score schema
  is written twice on purpose, and the contract's comment says it _"mirrors the
  in-runtime `validateScore` semantics so callers crossing process boundaries
  get identical guarantees."_ Run side by side: the runtime accepts an empty
  `scoreId`, `tenantId`, `authorUserId` and `template` — it validates **none**
  of the four identity fields — while the contract refuses each. The contract
  adds bounds the runtime never applies (segment id 80 chars, display name 120,
  height 7,680, an integer bitrate, a positive-integer timestamp). The runtime
  refuses whitespace-only segment ids and display names that the contract
  accepts.
- **The one field that cannot be mirrored.** `channels` is an array on the wire
  because a `Set` is not JSON-serialisable, and a `Set` in the runtime.
  Executed: an array passes the contract and is refused by the runtime
  (`invalid-cue`), a Set passes the runtime and is refused by the contract, and
  `JSON.stringify` renders a Set as `{}`. Exactly one hand-written seam bridges
  them — `channels: new Set(cue.channels)` in the BFF living-scenes route, which
  accepts either shape. **This is not an unbridged gap**, and the inherited
  slide already said so; the correction is recorded because a first, too-narrow
  grep nearly put "nothing converts" on a slide.
- **One rule, three enforcements, and the unreachable message.** A lookahead
  below two is refused by the contract's literal union `2|3|4`, by the runtime's
  set membership (`invalid-envelope: renderEnvelope is malformed`), and by a
  third guard in `deepParseScore` whose message — _"renderEnvelope must require
  ≥ 2 lookahead segments"_ — **can never be emitted**, proved by enumerating
  every value the union admits.
- **Eight fields, and thirteen that do not exist.** `SegmentSpec` has eight
  fields; the older prose described thirteen, and checked against the schema on
  this host **none of the thirteen names is present**.
- **The cue vocabulary.** Nine verbs, identical and in the same order in the
  contract, the runtime type and the parser. `sanitizeForCueParse` strips
  control characters, quotes, backticks, code fences and tags, collapses
  whitespace and truncates to 240 characters before the parser runs;
  `advance <script>alert(1)</script> now` arrives as `advance alert(1) now`, and
  a control character inside `linger` splits it into `ling er`, which matches
  nothing. `low-confidence` applies below 0.6 and **only on the voice channel**:
  a voice cue at 0.59 is low-confidence, a text cue at 0.59 parses.
- **A tap `linger` beats a voice `kill`.** The tap bypass is checked before the
  nine-rung priority ladder, so the hard-stop verb does not win against any tap
  in the same window. Measured over seven cue races, and recorded plainly
  because `kill` is the verb a reader assumes always wins.
- **A mis-keyed timestamp map drops every cue silently.** `reconcileCues` keys
  its timestamps by a composite of channel, verb, args and raw text, not by cue
  id. Executed with a map keyed by id: no winner, **zero** considered, no error.
- **The conductor.** The pre-warm window is `[playhead + 1, playhead +
  minLookahead]`, so the slot about to play is never in its own window: two
  slots at a lookahead of two, three at four, and exactly **one** under
  backpressure. `completePreWarm` from `ready` or `pending` is
  `invalid-transition`; an unknown segment is `segment-not-found`;
  `advancePlayhead` into a slot still `pre-warming` is refused by name and does
  hand the outbound carry-state into the next slot's inbound. `planReconnect`
  resumes from `streaming` **and** from `ready` — but a ready slot's resume
  payload is `null`. Of six declared slot states, **five** are reachable by
  these functions; nothing here produces `aborted`.
- **Policy refuses before arithmetic.** Four gates block a pair whatever the
  numbers say, and `fade-to-black` is the only accepted recovery: a crisis pair
  bridged by it scores a composite of 1.0000 and the identical pair bridged by a
  dissolve is still `hard-incompatible`. The mechanical composite is the
  unweighted mean of seven dimensions with exact-equality floors, so the worst
  single dimension alone still yields **0.9000** and every dimension at its
  floor yields **0.3714**.
- **A crisis frame collapses the technique list to three.** Executed over all
  twelve techniques on five templates: a Tara arc admits 7 ordinarily and
  **3** one frame into a crisis; a Metis lesson on a node under assessment
  admits **0**; Arete admits 9 without the operator grant and 12 with it.
  `assertTechniqueAllowed` names the gate and the enforcement point that
  refused.
- **Two files answer the same policy question differently.** The template
  catalog's `allowedCueVerbs` and the runtime's
  `DEFAULT_WORKFLOW_CLASS_POLICIES` disagree on **both** templates that appear
  in each: for Veritas the catalog forbids `morph`, `shift-style` and
  `shape-by` where the runtime forbids the first two; for Metis the catalog
  forbids `morph` and `shift-style` where the runtime forbids only
  `shift-style`. By contrast the duplicated per-template **technique**
  allowlists (catalog and compose-assist) **agree** on all five templates once
  the Arete operator grant is supplied.
- **Promotion.** `validateCompositionDraft` returns and never throws;
  `promoteCompositionDraft` throws five different ways. The continuity floor of
  **0.85** is applied twice — 0.85 promotes, 0.84 throws, and a draft at 0.9
  with one boundary at 0.84 throws naming that boundary. The crisis lock is
  read **before** the tier, so the uncapped AAA draft is refused by it too.
- **What the boundary plan does not check.** Count and adjacency are enforced;
  **unique ids and full coverage are not**. Two boundaries sharing an id promote
  with one map entry — the later value replaces the earlier — and a plan that
  repeats one adjacent pair while omitting the other also promotes. A single
  segment returns no boundaries and ignores a supplied array.
- **The record.** Six fields — `scoreId` (the draft id with a `score_` prefix),
  `draftId`, `destinationTemplateId`, `segmentOrder`, `techniquesByBoundary`,
  `reducedMotionTechniquesByBoundary` — with no media, no render envelope and no
  persistence result. Promoting the same draft twice returns the **same** score
  id. The twelve techniques collapse onto four reduced-motion targets, one of
  which (`still-frame`) is not itself a composition technique.
- **The gold set grades refusals.** An always-suggesting assistant scores recall
  1 with `correctRefusals: 0`; an always-refusing one scores precision 1 with
  `falseRefusals: 3`. Only the refusal columns separate them.

### What this chapter's evidence tops out at

Nothing here rendered a frame, called a provider, played a scene or wrote
anything. Continuity metrics such as FVD are consumed as inputs by the gate and
computed nowhere in these modules. The five crops are the shipped component over
its own built-in rehearsal data at a pinned commit. The route from a promotion
record into the standard runtime needs an adapter, a complete runtime record and
execution evidence that does not exist here.

## Checks

| Check         | Result                                                                  |
| ------------- | ----------------------------------------------------------------------- |
| Build         | 1,703 slides, 713 sources; freshness check clean                         |
| Python suite  | 490 tests pass                                                          |
| Node suite    | 508 tests pass                                                          |
| Browser sweep | 43 slides × 1440, 1280 with audio chrome, 390 and print — 0 findings     |
| PDF           | 43 pages, exact authored visual text; the chapter has no table           |
| Captures      | 5 existing crops reused byte-for-byte; provenance unchanged              |
| Narration     | 43 tracks, 76.1 min; library-wide `--check` verifies 1,703 tracks        |
| Bindings      | 18 reviewed evidence entries re-reviewed and re-stamped; 0 stale         |

Four sweep fixes: a four-step case study 6 px over at 1280 became three steps,
a second one likewise, and two callout or record strings ran past the 390 px
viewport and were shortened.

## Bindings re-reviewed, not re-stamped

`assignments/v1-living-scenes-composition.json` is hand-authored and
hand-reviewed; no generator produces it. The rewrite staled all **18** evidence
entries that cite this chapter, across two of its three units. Each was
re-reviewed against the rewritten slide before its fingerprint moved:

- Unit `4090dbe213c560dab964549d` — _Customer-Tier Composition Surface_ — 15
  slides. The rewritten slides still teach the surface, its source eligibility,
  its tier and crisis validation, promotion, live direction and every
  reduced-motion mapping.
- Unit `21ef39b149ef5a40912be5fa` — _Compose Assist — budgets and gold-set
  eval_ — 3 slides. `scene-compose-tiers`, `-promotion` and `-record` still
  carry the caps, the continuity-belongs-to-promotion distinction and the
  returned-maps-are-metadata point the unit's rationale names.

No evidence entry was added or dropped: a re-review confirms the existing
bindings rather than widening them. `reviewedAt` moves to 2026-09-13 and the
`guideId` of each entry moves to the merged id.

**A trap worth recording:** the fingerprint the coverage tool compares is
computed over the slide as `build-eve-oshun.py`'s `load_content()` returns it,
not over the raw manuscript — the builder folds the diagram, the inherited
release and the chapter into the hash. Re-stamping from the manuscript left all
18 entries still stale.

## Tests updated

Seven, each with the reason beside it:

- `test_build.py`'s tracked-chapter count went 1,300 → 1,343.
- `test_build.py`'s legacy-diagram wrap specimen was `scene-compose-reorder`,
  whose entry retired here. It had already been repointed once for the same
  reason, so it now takes whichever legacy `graph` slide with declared geometry
  the library still delivers.
- `test_eve_design.py` named `scenes-editors` for the comparison-focus and
  heading-intent contracts and `scenes-cue-priority` for the gated-flow
  contract; all three now pick the first delivered slide of the layout under
  test.
- `test_v1_living_scenes_composition_assignment.py`'s sequence test moves to the
  merged id, 43 slides, and the `atelier-studio` prerequisite — and now asserts
  that the Eve chapter's slides **do** carry an emphasis while the un-rewritten
  assist guide's do not.
- Its product-crop test follows the crop rather than the `product-view` layout,
  since the five slides are `capture-callouts` now.
- Its reorder-diagram test restates the property over the new `graph-diagram`
  shape: a seam is entered by exactly one segment and leaves to exactly one.
- Its PDF graph-text test now includes the required fields past the third, which
  a composition has and a legacy graph did not.
- `test_v1_living_scenes_techniques_assignment.py` and
  `test_v2_rollback_session.py` both learn the merged id.

## Treatment deviations

- **Five `product-view` slides → `capture-callouts`**, reusing the same crops.
- **Two legacy `graph` slides → `graph-diagram`**, retiring their entries.
- **Five tables → stat panels, compare panels and card grids**; no table
  remains.
- **`scene-compose-opening` is a three-step case study**, not four, for the
  1280 px overflow reason above.

## What this chapter does not establish

No frame was rendered, no provider was called, no scene was played and no
composition was promoted into a runtime.
