Eve library

Living Scenes · What may be suggested, and what may leave · receipt · 13 September 2026

Guide

Revision 1

On this page

Chapter five of Creation, and the track's third merged chapter. Delivered in the Eve edition on 13 September 2026. living-scenes-assist (12 slides) and living-scenes-governance (10) become one thirty-slide chapter under the new id living-scenes-assist-share. The brief is authoring/living-scenes-assist-share-chapter.md. Machine-readable state is in state.json.

What was delivered#

Thirty slides: twenty-one inherited slides rewritten in place and nine new — a cover, four section dividers, a close, and three content slides the executed evidence demanded. Four sections: what assist may propose (5), what the helper will not decide (7), keep is not share (6), when an input goes away (6).

One inherited slide is deliberately not carried over. scenes-promise was listed by both living-scenes-governance and living-scenes-score, and Creation chapter three already rewrote it in the Eve edition as the step-journey this chapter's plan asked for. It stays in chapter three rather than being taught twice, and no narration or deck is orphaned by that.

Both source manuscripts retire differently. content/08-living-scenes.json held nothing but the six governance slides and is removed; content/47-living-scenes-assist.json is removed; content/24-creation-deepdives.json loses its three governance slides and keeps the three agentic-studio slides it also held. Both source decks are removed and living-scenes-consumers is repointed onto the merged id.

scenes-keep-share was the last slide in a Living Scenes guide rendered by the legacy diagrams.json graph renderer. Its entry is deleted and the slide is now this chapter's one real capture, leaving eight legacy diagram entries library-wide.

The three new content slides#

Each exists because executing the modules turned up something the inherited slides could not have said:

  • scenes-fallback-forms — the four reduced-shareability forms work by opening components up, so between them they can relieve only four of the nine components. With the running scene's source grant lowered to private-only, the offered list comes back empty.
  • scenes-cascade-execution — takedown/execution.ts exists because the nightly audit was checking a cascadedAtUnixSeconds nothing produced. Its own header says so. Executed against a mutator that throws on the second link, it reports status: failed and mutatedShareLinkIds: [] after three mutator calls, two of which had already landed.
  • scenes-empty-denominator — seven counters in this chapter handed their emptiest accepted input. Four return a pass; three refuse.

Six treatments taken past the planned target#

Fifteen of the twenty-one per-slide treatments landed exactly as planned. Six did not, each for a reason recorded here:

  • scene-assist-policy · card grid → stat-panel. The slide is four exact counts of the same twelve technique ids under four contexts, and a stat panel is the layout that carries an exact number with its unit and its source.
  • scene-assist-empty · card grid → compare-panel. It contrasts exactly two inputs across five aligned rows; a card grid would break the alignment that is the point.
  • scenes-safety · step-journey → decision-tree. Executing the detector showed one threshold question with three endings, not a sequence of steps with actors and records.
  • scenes-provenance · card grid → stat-panel. Measuring the three signals produced four numbers with units, which is a panel, not a grid.
  • scenes-release-gates · card grid → stat-panel. Same reason: four configured thresholds read from the default set.
  • scenes-keep-share · step-journey → capture-callouts. This is the chapter's one real capture and it retires the last legacy diagram entry in a Living Scenes guide. The four-step journey the plan asked for is not lost: it is living-scenes-governance-worked, which is a step-journey of five steps with an actor and a record each.

The governance half changed source#

Its nine inherited slides cited V1/features/scene-*.md and nothing else: they taught a specification. Every one now teaches the shipped module beside the feature document — shareability/matrix.ts, fallbacks.ts, per-component-revocation.ts, takedown/dispositions.ts, execution.ts, nightly-audit.ts, release-gates/eval-gates.ts, the four provenance/ modules, safety/pse-strobe-detector.ts and, where it reaches them, the BFF route.

Evidence#

Everything below was executed at the pinned revision 1ceae52c7896a59ac74a4429d410f3218f0b91c6 on this host, which is byte-identical to the working tree for every path touched. No media was rendered, watermarked in production or shared, and no service was started.

  • Two callers, and a route that reaches part of the rest. Enumerated over all fifteen runtime exports of the Compose Assist module, exactly two non-test files reference any of them — ComposeClient.tsx and the mobile compose.tsx, both components — and ten of the fifteen are referenced by neither. apps/oshun/bff/src/routes/living-scenes.ts exposes eighteen distinct paths and imports the keep, share, revoke and abuse-report helpers, all seven takedown dispositions, executeTakedownDisposition and createEd25519C2paSigner. It does not import the shareability matrix directly, but createLivingSceneShare does, so the matrix runs on every share. Called by nothing at all, in the library or out: the fallback list, the revocation cascade, executeTakedownCascade, the nightly audit, the release gate, the channel evaluator, the strobe detector, the visible-mark helper and the watermark DSP. Two corrections landed here — the first draft claimed the whole half was wired, and the second claimed nothing outside the library imported the matrix, which was true of imports and false of what runs.
  • The matrix is wired, and handed its own answer. The share route builds its grants with defaultShareabilityGrants(reach), which returns nine components all set to the requested reach. Executed at all four levels: the nine reaches collapse to one distinct value, the resolved reach equals the request every time, and all nine components come back as deciding. The only input that can change the outcome is the crisis label.
  • The validator checks three things. Seventeen records through validateSuggestion: nine accepted, eight refused with four distinct codes. A record with no requestId, no candidateSegment and an empty rankedTechniques array passes. A rationale of seven characters padded to eleven with spaces fails, because the length is measured after trimming.
  • Policy, counted over every template. All twelve technique ids against all five templates: 7, 6, 6, 9 and 9 admitted in an ordinary frame, 0 for an unknown template id, and 3 — hard cut, dissolve, fade to black — on every template one frame into a crisis. The Metis assessment lock is template-scoped (four templates keep exactly their prior counts; metis-lesson-visualizer goes to 0) and the strictest Lilith tone band narrows only the three operator-curated Arete techniques, and only on arete-living-offering.
  • The budget Boolean conflates two situations. Every cap met exactly returns false; one token, cent or millisecond over returns true; and a zero cap, an infinite cap, a negative observation and a NaN observation all return the same true.
  • The gold set, and what a perfect score can mean. The five-case ledger gives precision 0.666…, recall 0.666…, refusal rate 0.4, one correct and one false refusal. An empty item list returns zeros. Twenty cases that all expect a refusal and all receive one return precision 1, recall 1, refusal rate 1. A suggestion whose rationale is two characters and whose ranked list is empty — a record the module's own validator throws on — is counted as a true positive.
  • The matrix names the component that decided. The running grant set resolves to link-with-password, deciding component music; a tie returns both names. Seven grants throw missing-grants listing the two absent ones, a repeat throws duplicate-component, an unreadable reach throws invalid-grant, and minReach returns private-only for an empty list and for an unrecognised value. A non-safe crisis label returns crisis-private-only before any validation: a set missing seven of nine components, and a set carrying a duplicate, both resolve without an error.
  • Four fallback forms, four relievable components. Asking for public reach on the running scene offers the muted still frame and the transcript, both resolving to public; full-render and audio-with-still do not clear it. The four forms drop only voice, music, likeness and persona, so with the source grant at private-only the offered list is empty, as it is under a crisis label.
  • Seven dispositions, five banner-only. All seven executed: five leave the shared link live with a banner, one invalidates it, one tombstones it; three leave playback unchanged. A terminated persona contract is a distinct kind from a retired voice profile, with its own banner and its own re-render option — the inherited table showed six rows and merged those two.
  • A failed cascade is honest and incomplete. The re-detection sweep applied against a recording mutator produces three mutations and a stamped record; against a mutator that throws on the second link it returns status: failed, cascadedAtUnixSeconds: null and mutatedShareLinkIds: [] — after three mutator calls, two of which stood. The whole seven-disposition cascade against that mutator returns seven failures and no aborts.
  • The nightly audit counts, and passes on nothing. Applied exactly at the window edge counts as met; one second past it does not; 999 met of 1,000 passes a 0.999 floor exactly. With no impacts at all it reports slaMetRate: 1 and raises nothing.
  • The release gate reports every comparison. A worst-case report returns 24 named failures across six eval kinds — five continuity, three determinism, two watermark, one cue policy, three crisis, ten accessibility. The diverging-frame fraction never fails on its own, because it is defined as zero when no frames were compared; the separate framesCompared minimum of one catches that case. A channel with no fixtures returns passes: true, and only evaluateAllChannels refuses, by listing the channel as missing.
  • Provenance, measured. The audio watermark round-trips at a bit error rate of 0 with the right key and 0.53 with a seed one integer away; an all-erasure read scores 1 and fails rather than dividing by zero. The C2PA sidecar signs with a real Ed25519 key and distinguishes three tampers by name — media hash mismatch, canonical hash mismatch, signature mismatch. The visible short code is a 32-bit rolling hash: over 200,000 sequential ids of the form scene-<base36> it produced 97,807 distinct codes and 102,193 collisions, where a uniform draw over its 729-million-code space would produce about 27. The first colliding pair is scene-1n and scene-30, both MG4HG6.
  • The safety gate is hard, and can be starved. More than three flashes in any one-second window refuses: a 12 Hz strobe reports 24 per second, and four per second is the smallest refusal. A strobe between 0.85 and 0.98 reports zero, because the darker side of every pair is above the 0.8 threshold — the criterion working as written. A NaN luminance clamps to 0 and is counted as a flash. An empty frame list returns safe: true.

The capture#

tools/presentations/tests/share-settings-capture.mjs is new. It renders the shipped ShareSettingsPanel from the pin over a declared fixture whose short code is computed by the shipped generateShortCode, and asserts four characterisations off the rendered DOM: the code's shape and alphabet, three separately copyable surfaces each with its own control, the iframe snippet's referrerpolicy, loading and separate /embed path, and one finding the capture makes visible — the panel discards the attestationUrl origin it is handed and rebuilds every surface from window.location.origin on mount.

Limits are recorded in capture-provenance.json: Tailwind utilities are generated at build time and are not reproduced, and no utility stylesheet was invented to stand in for them, so the image shows structure and text rather than deployed styling. The copy buttons are given the box size their own class names ask for (h-9 w-9 and a border) and the icons are mocked as inline-blocks of the size the component asks for. No route, session or share record was involved.

The four callout pins were re-measured against the rendered slide rather than the sweep: the first placement put pin three over loading="lazy", which every viewport check passed.

Corrections to inherited claims#

  • The takedown table showed six dispositions; the module defines seven. persona-contract-terminated has its own banner and its own re-render option.
  • "The Metis assessment lock is evaluated as an additional restriction" — it is template-scoped, and so is the strictest tone band, which narrows exactly three techniques on one template.
  • "A non-safe crisis label takes precedence before normal grant evaluation" — stronger than stated: the crisis branch returns before any validation, so a malformed grant set is never reported while the label is set.
  • "Four signals preserve attribution" — three modules and a page. Two were measured here rather than described, and the visible short code is not an identifier.

Verification run#

  • build-eve-oshun.py then --check: clean.
  • eve-edition-review.mjs --guide living-scenes-assist-share: 30 slides at 1440×900, 1280×720 with audio chrome, 390×844 and print — 120 measurements, zero findings, zero page errors. Two findings were fixed before this run: the cascade sequence carried eight messages and pushed the narration chrome off screen at 1440 and 1280, and is now six.
  • check-center-coverage.py with and without --check: no stale reviewed bindings anywhere in the library.
  • Python suite, Node suite, Ruff and Prettier: see state.json.

What this chapter does not establish#

No media was rendered, watermarked in production or shared. The audio watermark was embedded into a synthetic sine buffer, not into a scene's audio. No BFF route was started: the endpoint list is read from the route module and the modules under it were executed directly. The gold-set arithmetic is an executed illustration of the evaluator over the ledger the source module's own characterisation describes — never a measurement of an assistant, because at the pin no assistant runs.

Not checked#

  • The eighteen endpoints were read from the route source, not called. Nothing here establishes that the route behaves as its handlers read.
  • The ShareSettingsPanel capture shows the component, not the deployed scene page, which is a server component that fetches from the BFF.
  • The short-code collision measurement used 200,000 sequential ids of one shape. It is a statement about that population, not about every id scheme the product might use.