Source revision: 1ceae52c7896a59ac74a4429d410f3218f0b91c6. The
full checklist remains active. This is a verified Iris source,
narration and presentation increment, not final library, memory-safety or
deployed-product acceptance. The bound receipt records its exact
artifacts and deliberately leaves global completion claims false.
Source fidelity and teaching#
The learning map records six reviewed source units, the remaining Iris decomposition and implementation boundaries. The guide adds focused ranker and inspection slides, corrects seven existing Iris explanations and corrects the shared Eve memory explanation in both Iris and Eve context.
- The write-conflict helper can protect a confirmed preference while recall
independently selects a newer inferred entry. The write helper's
most-recent-winslabel does not itself skip the response requirement. - Recency uses a 30-day e-folding time, not a 30-day half-life. Scope, gate, budget, origin-score and Jaccard behavior follow the pinned code.
- Recall consumes caller grants and candidates, not the consent ledger, entry revocation/expiry or current revision storage. Metadata redaction leaves body text intact; dedupe/budget omissions are not gate-suppression counts.
- The in-memory store links revisions, changes the prior lifecycle, retains tombstone content and cannot invalidate an old caller snapshot. A mutator can replace provenance; summarization does not supersede source entries.
- Ledger validation is separate from append. Selected-field FNV-1a fingerprints are not authenticated signatures or automatic live-terms drift detection. Active views are type-wide and ignore future effective times.
- Inspection gates supplied request fields but does not authenticate consent references or validate the supplied snapshot's identity/sensitivity. Local audit objects and notice queues are not proof of persistence or delivery.
Thirty-one new Node checks execute original pinned modules and retain the teaching boundaries in authored slides. The contracts use the pinned primitives and the locally installed contracts package's Zod 4.4.3 dependency. Write-plan audit/redaction dependencies are guarded unused replacements; audit ingestion uses a controlled acknowledgement. No runtime implementation was changed. These are not HTTP, authentication, real storage, live privacy leakage, consumer web/mobile UI, provider or deployed-service acceptance results.
All 54 Python tests and 82 Node tests pass in the Python log and the Node log. The teaching gate correctly exits 1: 21 reviewed units, 103,305 unassigned meaningful units. The freshness check has no invalid or stale assignment errors. This is an expected rejection of global teaching completeness, not a passing complete gate.
Design and verification scope#
The frontend skill guided two focused additions: a four-factor scoring table and a request-to-close inspection flow. The worked comparison uses two plain columns because write planning and recall are different paths, not successive steps of one guaranteed correction. Existing entry layers and comparison tables retain the cream, ink and terracotta language; detailed counterexamples stay in the explanation. Existing navigation, disclosure and reduced-motion behavior remain unchanged.
Repository Playwright, axe and Poppler are the available browser/PDF workflow; no dedicated presentation/PDF skill or normal-speed audio-listening tool is available. Automated decoding, silence measurement and accelerated browser playback cannot establish human listening quality.
Narration and standalone results#
Two sequential, bounded CPU batches rendered ten tracks: seven previously
pending Iris explanations, the two new Iris explanations and a replacement for
the corrected shared Eve memory explanation. The unchanged substrate track was
already current. See batch one and
batch two. Stock Kokoro af_heart, speed 0.95, 24 kHz and
the established sentence-pause policy were retained; no voice cloning or
provider requests were used.
The current-bank audit and
decoded-audio log verify all 162 current tracks:
13,973.959 seconds and 2,161 sentence segments. Fingerprints, file bytes,
SHA-256, actual decoded duration and encoded pauses pass. 221 tracks remain
pending: 201 stale and 20 missing. The current manifest hash is
9fd436d56e5ce218db8c5c3ce7f9d95ca968f1e4083084394c0bae2e2ef8357c.
Both full non-draft standalone suites pass: Iris, 11 slides and Eve context, nine slides. They exercise source and transcript alignment, numeric/stable links and history, keyboard navigation, search, empty states, source expansion, dialog focus, every slide's desktop, compact and mobile layout/accessibility, print sequence, HTTP audio recovery, missing/corrupt files, playback races, actual decode and playback controls. Their Iris log and Eve-context log record no browser runtime errors and unchanged tested HTML. Iris has 1,388.471 seconds of current narration; Eve context has 784.354 seconds.
Center, PDF and visual results#
The complete center run passes all 14 checks across
38 guides, 383 unique slides and 409 guide-slide instances. It validates all
38 PDF page sequences and captures every embedded slide at desktop and mobile
sizes: 818 hash-bound initial-viewport images. Actual audio boundaries pass
at both sizes for all 17 ready guides (34 boundaries); the report explicitly
skips audio for 21 incomplete guides. --draft --ready-audio is not full-bank
narration acceptance. Navigation, search, history, hierarchy, measured
durations, focus, drawer playback interruption, accessibility, local interaction
budget and unchanged HTML all pass. See the
center log.
The visual ledger retains 69 inspected images across ten unique changed slides and eleven guide contexts. Each context includes standalone desktop, compact, full-height mobile, embedded desktop, embedded mobile and the final PDF page; the three tables also have keyboard-focused mobile endpoint captures. The shared Eve slide is reviewed in both guides. Embedded mobile captures show the initial viewport, not the complete scrolling page. Lower content is covered by the full-height standalone and PDF review. No material defect remains in these inspected states. A suspected first-row PDF collision was disproved by native-resolution review and the PDF's word coordinates; the retained Iris failure-page render is 2400 pixels wide. No layout modification was needed.
The static log records seven targeted checks: generated artifact freshness, coverage freshness, Ruff, scoped ESLint, presentation Prettier, explicit evidence README/ledger formatting, and Git whitespace checks. No runtime TypeScript source changed; the executable pinned-module tests are not a project typecheck. The receipt revalidates all 818 temporary capture bytes, 45 center-bound files, 38 final PDFs, 162 current MP3s, reviewed slide hashes, six new assignments, the 69 retained images and the input/evidence file hashes. Raw reports are retained without rewriting their historical claims.
Verification ran one expensive owned job at a time on the 15 GiB host with 31 GiB swap. The narrator, audio audit and browser runs used the memory supervisor; center resource observations show the competing processes and available memory. No broad Nx/build matrix or unrelated process termination was performed.
Reproduction#
Use the configured Python environment and Chromium installation. Run these checks sequentially with the repository's resource controls:
python -m unittest discover -s tools/presentations/tests -p 'test_*.py'
node --test tools/presentations/tests/*.test.mjs
python tools/presentations/build-eve-oshun.py --check --preview
python tools/presentations/check-center-coverage.py --check
python tools/presentations/check-center-coverage.py --require-complete
CHROME_CHANNEL=chromium python tools/presentations/run-supervised.py node tools/presentations/tests/eve-oshun-deck.mjs --guide iris
CHROME_CHANNEL=chromium python tools/presentations/run-supervised.py node tools/presentations/tests/eve-oshun-deck.mjs --guide eve-context
CHROME_CHANNEL=chromium python tools/presentations/run-supervised.py node tools/presentations/tests/presentation-center.mjs --draft --ready-audio
The complete-coverage command deliberately fails while the global checklist is unfinished. Normal-speed listening, full narration inventory, full source/code teaching, real Eve integration and final Domains/visual acceptance remain open.