# Psyche and Lilith · What happens in the moment, and what it costs · receipt · 13 September 2026

Chapter six of _Shared systems_, delivered in the Eve edition on 13 September 2026. Fifteen
inherited slides become a twenty-slide chapter under the `psyche-lilith` id. The
brief is [authoring/psyche-lilith-chapter.md](../../authoring/psyche-lilith-chapter.md).
Machine-readable state is in [state.json](state.json).

## What was delivered

Twenty slides in the guide: fourteen inherited slides rewritten in place, one
carried unchanged, and five new — a cover, three section dividers and a close.
Nothing merges away.

`member-day` already lives in the web PWA chapter, which is in the Eve edition;
the guide lists it and this chapter authors it nowhere.

The fifteen arrived across five manuscripts.
`content/06-psyche-runtime-detail.json` and `content/06-lilith-experience.json`
held nothing but this chapter's slides and retire.
`content/06-evidence-memory-runtime.json`, `content/23-shared-deepdives.json`
and `content/72-lilith-web-pwa.json` lose theirs and keep the rest.

## Evidence

No capture, and the reason is a judgement rather than a limit: `apps/psyche/admin`
ships real page components at the pin, but they are operator surfaces for the
half of the substrate the source itself calls partially built, so a screenshot
of a working dashboard would imply a working runtime — the one claim this
chapter exists to refuse. Every number below is executed against the pinned
modules on this host, or read at the pin and said to be read.

- **The turn reducer, executed here.** Idle to active on a turn-start. An
  interruption whose policy approval is false leaves the phase active; an
  approved one moves to interrupted and emits `turn-end` itself, after which an
  explicit `turn-end` is refused with "cannot apply turn-end without an active
  turn". Partial speech fades over 180 ms by default.
- **The event vocabulary, executed here.** Fifteen canonical wire events, two
  server turn events the published list omits, nineteen legacy dotted kinds —
  thirty-six in the combined union.
- **The latency targets, executed here.** Voice first audio 500/900/1500 ms,
  first token 350/700, avatar frame 80, final recognised text 150; the SLO set
  adds a round trip of 250/500/750 and a pipeline total of 800/1200.
- **Admission, executed here.** At the session cap: defer for a paying member
  and an operator, queue at 75 s for a standard member, reject preview and
  background. With one service saturated, everything but the two privileged
  tiers queues at 15 s.
- **Shedding, executed here.** Background, then preview (downgraded to text),
  then standard off the critical path (disconnected), then standard on it
  (handed off). Asking to reduce six sessions to one sheds four: a paying member
  and an operator are never candidates.
- **Provider backpressure, executed here.** Degraded, critical and unavailable
  produce the same three actions; only elapsed time changes the plan, and past
  the 2,500 ms text-fallback threshold the status becomes text-fallback. The
  thinking indicator deadline is 200 ms and both throttle ratios are 0.5.
- **The fallback cascade, executed here.** Each skipped tier carries its reason.
  **A device that cannot play audio still resolves to avatar** — the avatar tier
  does not test audio playback. State it as a gap in the tier's conditions at
  this pin, not an observed failure.
- **Recovery, executed here.** Conditions recovered gives `offer-upgrade` with
  confirmation required; still degraded gives `hold`; a second of stability
  gives `hold` with a later next-eligible time; a member asking gives the same
  offer. `attempt-upgrade` appeared in none of these cases.
- **Tone bands, executed here.** Eight bands with their ceilings; narrowing
  holds in both directions and widening is refused; `validateToneBandCatalog()`
  passes on the shipped catalog.
- **Lilith crisis detection, executed here.** Thirteen crisis types, a
  fifteen-signal taxonomy of literal phrases. Four catalogued phrases returned
  needs-escalation at critical with three resources; four ordinary paraphrases
  returned safe. `mergeLilithSafetyAnalyses` takes the stricter in either order.
- **Read at the pin.** The crisis frame computes the persona break, the
  synthesis halt and the memory-write suspension in one step with the triple
  stamped on two audit events; persona handoffs carry a memory boundary; the
  persona registry authorises per consent kind, surface, modality and
  commercial use, and has no type called a consent record.

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

Every module here is pure: it returns a decision, a plan or an envelope and
performs nothing. Nothing establishes that a session was admitted, shed,
degraded or recovered in a deployed runtime, and the operator surfaces that
would show the events and the audit trail are named as unbuilt in the source.

## Checks

| Check         | Result                                                             |
| ------------- | ------------------------------------------------------------------ |
| Build         | 1,687 slides, 702 sources; freshness check clean                    |
| Python suite  | 490 tests pass                                                     |
| Node suite    | 508 tests pass                                                     |
| Browser sweep | 20 slides × 1440, 1280 with audio chrome, 390 and print — 0 findings |
| PDF           | 20 pages, exact authored visual text; the one table clears the header floor |
| Narration     | 20 tracks, 36.3 minutes; all 1,687 library tracks verified         |
| Bindings      | 5 reviewed source units re-reviewed; 0 stale across the library     |

## Treatment deviations

- `psyche-turn-outcomes`: card grid or stat panel → `compare-panel`.
- `psyche-runtime-fallback`: kept as the plan asked, restyled `decision-tree`.
- `psyche-lilith-worked`: step-journey → `case-study`.
- `persona-avatar-voice`: flow → step-journey → `decision-tree`.
- Sections: **three of five**, not one divider per five to seven slides across
  six sections — fifteen content slides cannot carry six dividers without
  sections of two.

## Binding work

Five reviewed units cite this chapter, across seven slide bindings, and every
one went stale when the slides were rewritten. Each was re-reviewed against the
rewritten slide rather than re-fingerprinted: the event taxonomy and the
gap-free stream, server-mediated turn-taking (two slides), the latency budgets
and the thinking-indicator deadline (two slides), backpressure, and the fallback
cascade (two slides). All seven still teach their unit, so none moved.
`check-center-coverage.py` reports zero stale bindings across the library.

## Tests updated

Four, each with the reason beside it:

- `pinned-psyche-runtime.test.mjs` read `psyche-latency-targets` and
  `psyche-runtime-fallback` **by file path** from the manuscript that retired.
  It now looks a slide up wherever it lives, which is the shape the other pinned
  tests are moving towards. Both slides kept the shape it binds, deliberately.
- `test_tables.py` took both of its specimens from the same retired file, and
  one of them stopped being a table. Its quantitative specimen is now found
  wherever it lives, its prose specimen is whichever prose table the library
  ships, and the row and column counts are read off that specimen instead of
  being written into the assertion.
- `test_eve_depth_pdf.py` read the same retired file for the priority ladder;
  repointed the same way.
- The rollback guard gained `psyche-lilith` in `REWRITTEN_CHAPTER_GUIDES`, and
  the tracked-chapter count went 1,251 → 1,271.

## What this chapter does not establish

No live transport was used, no speech or avatar provider was called, and no
session was run. Every module here returns a decision, a plan or an envelope and
performs nothing; the operator surfaces that would display the events and the
audit trail are named as unbuilt in the source.
