# Initial composition treatments and Psyche/Lilith source review

This is a verified implementation increment, **not a completed pilot, accepted
benchmark or finished library**. The complete contract remains in
[TODOS.md](../../TODOS.md). Source claims retain the immutable revision
`1ceae52c7896a59ac74a4429d410f3218f0b91c6`; the implementation base is
`ed2780cfb7245bd587a999850bcab6b3b8b89d98`.

## What changed

The supplied web/PWA opening now pairs four independent questions with their
evidence, emphasizing the completed outcome without a funnel. Shared
`member-day` now shows optional activities, a no-activity path and an explicit
return labeled **Library, history and permitted memory**. Its caption, notes and
replacement narration preserve release, consent, interruption and genuine
retention boundaries. Both web/PWA and Psyche/Lilith reuse the same authored
relationship.

The new semantic renderer validates explicit layouts, briefs, node reading
order, emphasis, labels and endpoints. Unknown layouts no longer become columns.
Recurrence rejects compulsory chains among optional peers; the evidence
comparison rejects directed edges. Relationship data is separate from CSS/SVG
geometry. The initial two treatments do not constitute the required repertoire.
Legacy numbering, identical flow/sequence rendering and generic graph boxes are
still deliberate migration targets.

Psyche/Lilith expands from eleven to fifteen slides, separating turn outcomes,
configured latency metrics, runtime fallback selection and recovery. Ten
existing explanations were corrected or enriched before the shared journey's
redesign. Five architecture units have new reviewed teaching assignments. The
[learning map](../../drafts/psyche-lilith-learning-map.md) retains detailed
source findings and untaught depth. In particular, the Lilith adapter's raw
`analyzeSafety` is distinct from wrapped `assessSafety`; explicit analyzer
replacement and rejected dependencies remain caller obligations.

The frontend skill guided cardless relationship-first composition, concise
visible labels and retention of useful source/detail controls. Canonical
[per-guide direction](../../authoring/composition-pilot.json) covers a complete
eleven-slide web/PWA sequence plus four additional pilot candidates. The
[reader protocol](../../authoring/reader-review.md) contains neutral tasks, not
participant observations.

## Evidence and limits

| Evidence                                                  | Current result                                                                   | Boundary                                                                                                                                         |
| --------------------------------------------------------- | -------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------ |
| [Python log](python-tests.log)                            | 76 passing tests                                                                 | Includes 19 composition/audit and 3 supervisor tests; not a visual acceptance score                                                              |
| [Node log](node-tests.log)                                | 134 passing tests                                                                | Includes 47 pinned Psyche module and 5 isolated Lilith adapter tests; not product/media/crisis-detector acceptance or a TypeScript project check |
| [Static checks](static-checks.json)                       | Seven scoped gates pass                                                          | Builder/fixture freshness, coverage freshness, Ruff, ESLint, normal and evidence-author formatting, diff whitespace                              |
| [Web/PWA standalone](v1-web-pwa-browser.json)             | All eleven slides, non-draft                                                     | Current actual audio, controls/failures/races, desktop/compact/mobile, axe, print/source checks                                                  |
| [Psyche standalone](psyche-lilith-browser.json)           | All fifteen slides, draft                                                        | Same applicable layout/source/navigation checks; nine pending tracks prevent full audio acceptance                                               |
| [Complete center](center-ready-audio.json)                | 38 guides, 413 instances, 38 PDFs                                                | Draft/ready-audio mode, not the fully narrated release gate                                                                                      |
| Center actual audio                                       | 34 desktop/mobile boundaries across 17 ready guides                              | 21 incomplete guides explicitly excluded from audio, not from layout/PDF coverage                                                                |
| Center review inputs                                      | 826 ordered, hash-verified embedded initial-viewport captures                    | Only the targeted ledger's images are claimed visually reviewed                                                                                  |
| [Current audio audit](current-audio-check.json)           | 167 tracks, 14,773.101 seconds, 2,276 sentence segments                          | Actual MP3 decode, fingerprint, bytes, hash, duration and encoded pauses; not normal-speed listening                                             |
| [Focused pilot](after/after-result.json)                  | 33 current captures; no runtime/axe/semantic errors                              | Provisional text floors, normal/fallback/enlarged fonts and scroll reachability; not reader acceptance                                           |
| [Model visual ledger](visual-review.json)                 | 165 inspected images, 16 unique slides, 17 guide contexts                        | 108 guide images + 24 before + 33 after; all eight craft areas explicitly scoped                                                                 |
| [Teaching report](../../inventory/teaching-coverage.json) | 26 reviewed, 10,018 structural, 103,300 unassigned; no invalid/stale assignments | `--require-complete` correctly exits 1; complete teaching is still missing                                                                       |

The catalog has 387 unique slides, 413 guide instances and 235 cited files. All
38 delivered PDFs match the current center run's hashes and complete page
sequences. The [before](before-audit.json) and [after](after-audit.json)
structural audits keep unique/instance counts separate; four-column openings
change from 33 to 32. This locates editorial work, not proof of quality.

Audio remains **220 pending: 196 stale and 24 missing**. Psyche has six current
tracks and nine pending; web/PWA remains fully narrated. Seven unique delivered
MP3 files changed across the source and pilot work, for a net increase of five
current tracks from the Iris checkpoint. The preserved narration log covers the
final three replacements, not every earlier render. Stock Kokoro/Misaki 0.9.4,
`af_heart`, speed 0.95, one CPU thread and the requested oh-shuhn/pacing policy
remain unchanged. No provider credential or voice clone was used.

## Visual findings and viewing conditions

Before/after comparisons use the same contexts and viewports; the older web/PWA
duration label is 15 minutes and the current one is 17. Both content and
composition changed. Baseline renders are weaker alternatives, not accepted
references. There is no before-PDF comparison.

The focused matrix uses standalone 1440×900, 1280×720, 390×844 and 844×390;
center expanded/collapsed 1440×900, compact 1280×720 and mobile 390×844. Device
pixel ratio is 1 and browser zoom 100%. Reports retain the actual frame size:
the compact frame is 988×617 and mobile 390×741. All new semantic labels use at
least the provisional 16px floor. The separate missing-font test uses generic
fallback and 32px CSS text; it is not browser-wide 200% zoom or full offline
font portability evidence. Reduced motion is enabled in the focused run.

All fifteen Psyche slides and the two web/PWA contexts were inspected at
desktop, compact, full-height mobile, embedded desktop/mobile initial viewport
and 2000px PDF scale. Three table-end states and three pilot-context PDF pages
at 1280px page-fit were also inspected. Compact/mobile center content may extend
below the initial viewport: the six new end captures verify and show the return,
qualification, footer and navigation remain reachable. No fit-all claim is made
for these reflowing views.

The review keeps these material issues open:

- Three legacy Psyche tables allow horizontal scrolling but lose row identity
  while later columns are read.
- Configured latency targets need a visible explanation of the dash, not only a
  note; the values are not measured production performance.
- Runtime fallback needs actual fixed-priority branching and the independent
  safety floor, not just the current linear overview plus detailed notes.
- Psyche's generic opening, unordered legacy ordinals, worked-trace double
  numbering and repeated silhouettes need the remaining content-led migration.
- Full guide briefs/rhythm/contact sheets, typography/font specification,
  quantitative and image contracts, suitable artifact/image examples and real
  willing-reader review remain prerequisites to benchmark promotion.

The initial compact overflow was fixed by removing redundant pilot padding,
without shrinking labels. A harness defect previously carried 32px stress styles
through same-document navigation; fresh-document navigation and normal-size
assertions now prevent that contamination. Those failed/contaminated temporary
outputs are excluded. The final reachability run's 27 common captures are
byte-identical to the preceding reviewed final run; its six new end captures
were reviewed separately.

## Reproduction and supervision

Use the repository's resource checks and prepared requirements before running
these separately, with `CHROME_CHANNEL=chromium` on this host:

```sh
python tools/presentations/build-eve-oshun.py --check --preview
python -m unittest discover -s tools/presentations/tests -p 'test_*.py'
node --test tools/presentations/tests/*.test.mjs
python tools/presentations/check-center-coverage.py --check
python tools/presentations/check-center-coverage.py --require-complete
python tools/presentations/audit-compositions.py
OSHUN_COMPOSITION_QA_DIR=/path/to/new-evidence python tools/presentations/run-supervised.py node tools/presentations/tests/composition-pilot.mjs
python tools/presentations/run-supervised.py node tools/presentations/tests/eve-oshun-deck.mjs --guide v1-web-pwa
python tools/presentations/run-supervised.py node tools/presentations/tests/eve-oshun-deck.mjs --guide psyche-lilith --draft
python tools/presentations/run-supervised.py node tools/presentations/tests/presentation-center.mjs --draft --ready-audio
```

The completeness command is an expected rejection until remaining teaching is
done. Regenerate audio with the existing renderer, then use its full `--check`
only as a full inventory gate when all pending files are current. The current
subset report does not convert that incomplete global gate into a pass.

This host has about 15 GiB RAM and 31 GiB swap. Checks ran narrowly and
sequentially; no full Nx/build matrix, Expo/Tauri/Unreal packaging or concurrent
narrator/browser batch was started. Resource checks were repeated manually. Some
earlier launcher handles terminated or disappeared while their actual sessioned
children continued to complete; child state and final reports were inspected
instead of restarting duplicate jobs. The external cause is unknown. The new
supervisor tests prove cleanup for a controlled SIGTERM, including a stubborn
descendant, and signal-to-cleanup handling for SIGTERM/SIGHUP/SIGINT. They do
not establish protection against uncatchable termination. The separate narration
resource log is partial; the empty center resource log is excluded.

Available repository Playwright/axe and Poppler plus original-size model image
inspection supplied browser/PDF evidence; no dedicated presentation/PDF skill or
normal-speed listening tool was available. There are no claimed participants,
reader quotes, full listening results, live provider/media acceptance or Eve
integration. No underlying application/service code changed in this increment.

[checks.json](checks.json) binds current inputs and evidence to exact hashes;
historical before artifacts remain explicitly historical. Rerun affected
checks/reviews when these inputs change. Keep the full goal and all unresolved
pilot/coverage/listening requirements open.
