# Agentic recovery source review · 2026-09-07

This is verified source investigation and a pending correction manuscript, **not
publication or acceptance of corrected slides**. The current `agents-lifecycle`,
`agents-checkpoints` and `agentic-studio-failure` explanations still require the
changes in [the review](../../drafts/agentic-run-recovery-review.md) and
[prepared replacements](../../drafts/agentic-run-recovery-corrections.json).

Thirty-four new probes execute five unmodified TypeScript modules directly from
the immutable source pin. They cover the selected-field checkpoint
serialization, omitted-content and zero-budget counterexamples, state and
preservation behavior, operation-specific control guards, replay limits,
subscriber/summary separation and accounting. They use controlled data and
whitelisted local module dependencies, without transport, providers, storage,
authentication or a model. They describe source behavior; they do not endorse
its limitations or demonstrate a production exploit.

The existing **25-test `runs.test.ts` suite passes** through its real Vitest
configuration with one worker and no file parallelism. The package's tracked
content matches the source pin. This is one focused suite, not all package,
service, security, durability or release tests. Maximum recorded RSS is about
754 MiB; the run completed without swap activity in about nine seconds.

Three dry authoring tests pass for the correction manuscript: current
guide/slide identities, unchanged live content, existing table/layer contracts,
text rendering, semantic speech blocks and source resolution at the pin. They do
not establish browser geometry, usability, PDF rendering or recorded speech. The
proposed checkpoint table compares independent boundaries; its visual review is
pending. No live presentation content, HTML, PDF, speech input or runtime
implementation is changed by this increment.

`state.json` binds the exact source bytes, seven still-unassigned documentation
units, current assembled slide hashes, unchanged delivery artifacts, draft,
tests and receipt logs. The verifier deliberately reports
`correctionsApplied: false`, `visualVerified: false` and
`wholeObjectiveComplete: false`. Updating the affected sources/slides/draft
makes this receipt historical and requires a new review; green source probes do
not clear the pending editorial work.

```sh
node --test tools/presentations/tests/pinned-agentic-recovery.test.mjs
python -m unittest discover -s tools/presentations/tests -p 'test_agentic_recovery_draft.py'
node_modules/.bin/vitest run --config libs/oshun/agentic-studio/vitest.config.ts src/runs/runs.test.ts --maxWorkers=1 --no-file-parallelism
python docs/presentations/presentation-center/verification/agentic-recovery-source-review-2026-09-07/verify_evidence.py
```

Use the prepared presentation Python environment and the repository's resource
checks; run expensive targets serially. After the active global narration batch
is stable, apply the corrections with their new source citations, re-render the
three changed tracks, refresh affected inventories and execute the full
applicable browser/audio/PDF/source-review gates. One corrected track was in the
original 177-current baseline: document the deliberate replacement rather than
claiming every original recording stayed unchanged.
