The composition guide follows one three-segment rehearsal through source eligibility, arranging, picking, text preview, accept/cap behavior, validation, promotion, reduced motion, direction and web/native differences. Its nineteen slides lead into twelve Compose Assist slides explaining requests, candidate validation, filtering, budget signals, refusal, gold-set arithmetic and challenger decisions.
Five actual component crops, two adjacency diagrams and six decision diagrams replace a shallow feature inventory with observable behavior and explicit relationships. Source-shaped records and an executed five-case evaluation keep the precise fields and arithmetic inspectable. Peer rows have equal weight and their headings describe the corresponding values.
The package adds 31 slides, with current narration and complete PDFs. It preserves all 1,028 earlier slides, scripts and recordings, including the entire 25-guide, 217-slide Eve course. The library now has 81 guides, 1,059 unique slides, 1,082 guide-slide instances and 502 cited source paths.
| Evidence | Scope |
|---|---|
source/source-cases.json |
24 executed source cases; 25 unmodified helper tests and one native host test. Complete pinned modules and dependencies, with source hash continuity. |
source/web-owning-tests.json and the component asset provenance |
Three unmodified owning web flows, 12 further component observations, six reproducible crops and zero violations in the two scoped axe scans. |
browser/report.json |
Every new slide at four standalone widths and three embedded layouts; seven enlarged-text cases, image panning, semantic content, graph geometry, accessibility and four rejected negative controls. |
center/result.json |
Both new guides through navigation, source access, audio boundaries, embedded capture and full PDF export; preservation of the 79 earlier guide HTML files. |
pdf-text.json, pdf-bounds.json, optical-review.json |
Ordered visible text and page bounds for all 31 PDF pages; agent inspection of final contact sheets and selected full-size browser/PDF images. |
commands.json, logs/ |
Serial resource-supervised narration, builder, browser, center, PDF, focused presentation/assignment tests, freshness, coverage and scoped static checks. |
state.json, verify-receipt.py |
Exact artifact/evidence hashes, all 1,059 recording hashes and preservation of all 1,028 earlier canonical slides, scripts and recordings. |
The source pin is 1ceae52c7896a59ac74a4429d410f3218f0b91c6; the implementation
baseline is 836abf58de0346dbfe7bd1b318400f128b1dbbc7. The complete customer
composition, AI Compose Assist and template-editor assist sections have
reviewed assignments.
The per-slide ledger
records teaching jobs, sequence decisions, treatments and final content hashes.
The actual web component runs in a declared isolated host, with its complete helper and the three root styles in the owning layout's order. Its built-in rehearsal data is preserved. No invented utility stylesheet is added. The owning Playwright suite uses an explicit fixture facade to reach this component host; this does not establish a Next shell, authentication, a deployed service or a destination write. Native testing retains the existing React Native host mocks. The inspected Maestro flow was not run on a device. Transpilation is not typechecking; application implementation files are unchanged by this package.
The slides qualify the handbook's stronger integration wording. Filtering does not compute a pair-specific ranking; preview does not render overlap media; budget checking returns a Boolean; promotion returns six metadata fields rather than a complete saved runtime Score; challenger comparison does not deploy a model. The composer does not dispatch the described Isis assist run. These limits are taught, not silently repaired in product code.
The first optical pass prompted the two adjacency diagrams and shorter, more
useful captions. A tightly wrapped graph label was corrected before final
verification. The 390-pixel check caught an overflowing record name and map
identifier; the 320-pixel check caught a long function name in a decision node.
Those visible labels were rewritten in readable language, preserving the spoken
explanations and recordings. The interrupted browser reports are retained under
attempts/. The 131-case presentation run had one failure in a new synthetic PDF
fixture: it placed the visual body after the takeaway. Actual PDFs already
passed. The fixture was corrected and its seven-case module rerun; the other
124 presentation tests were retained from the passing portion of the original
run. Retained discovery logs include an initial duplicate local
variable in the new source harness, an undeclared URL import in the capture
harness, and an invalid temporary preview-config path; the applicable corrected
checks passed. Earlier passing app/source tests were not repeatedly rerun for
unrelated presentation edits. Compressed logs retain their original hashes.
This delivery does not finish the full V1 or whole-library design program. The complete technique/primitive catalog, full Score Editor, Tara venue editor, templates, rendering and sharing remain separate work, along with the other V1 workstreams. Human comprehension sessions and normal-speed human listening were not performed or inferred from automation.
At the package's owning revision, reproduce its artifact audit with:
python3 docs/presentations/presentation-center/verification/v1-living-scenes-composition-2026-09-09/verify-receipt.py
The runnable source and component harnesses are
tools/presentations/tests/pinned-living-scenes-composition.mjs and
tools/presentations/tests/living-scenes-composition-capture.mjs. Guide review
scope is in authoring/v1-living-scenes-composition-review-config.json.