# First 20, published for review

September 5, 2026. The user authorized adding this batch to the existing
Cloudflare Pages fieldnotes so it can be reviewed in place.

**20 draft treatments, 40 original-content renders, zero completed examples.**
Publication does not accept the fidelity findings, approve scaling, or replace
the five-treatment approved benchmark.

## Review the work

- [Browse the searchable first 20](review/): select a treatment, play its original
  and different-content adaptation, then its synchronized source comparison.
- [Consolidated review](REVIEW_FIRST_20.html): proposed coverage, reuse contract,
  source choices, concrete discrepancies and remaining decisions.
- [Runnable toolkit and original asset downloads](downloads/): the shared engine,
  typed inputs, recipes, source-mapped contracts and agent instructions.
- [Public derivative ledger](public-derivatives.json): hashes and field-level
  diagnostic redactions. [Review provenance](review-provenance.json) identifies
  the exact local freeze from which this review was built.

## What remains open

Five scoped source reviews have no open findings. Across the 20 current source
reports, nine material and 27 minor findings remain. The Nook reuse trial also
qualifies the inherited pill geometry and early cache-isolation evidence. None
of these differences is accepted by publishing this page.

The clean local toolkit run passes 136 tests and renders/verifies all 40
variants. The research suite passes 108 tests. Nine clean re-renders differ in
decoded pixels; no cause or perceptual approval is inferred. One independent
reuse pilot has a scoped pass and a second a qualified recommendation. This
does not satisfy the required 20 successful unseen briefs or the final
200-example library.

## Evidence and privacy

Original/adapted movies, source comparisons, contact sheets, render-bound code,
inputs and permitted assets retain their reviewed bytes. Full-resolution native
PNGs are transported in split archives; every frame row identifies its ZIP,
member path, source/local frame and original SHA-256. The 6,582 native images
remain available in 78 archive parts rather than thousands of loose URLs.

Public render-manifest copies remove only the machine-specific browser location.
Their binding, input, asset, implementation and output fields remain unchanged.
Independent reports may redact diagnostic paths. The viewer displays the
reviewed source-record hash separately from the public derivative hash. These
copies are not falsely described as byte-identical local reports.

Publication review caught two viewer defects: the old viewer rejected the
current comparison schema, and exact-boundary seeking could show the preceding
decoded frame. The published viewer accepts the current schema and seeks a small
amount inside a requested frame. Duration rounding no longer offers an extra
terminal frame. These are viewer repairs; movies and choreography are unchanged.
HTML-video frame seeking remains approximate. Native PNG/member mappings are
the reference for exact-frame inspection.

The live Pages response returned the full MP4 with HTTP 200 even when a byte
range was requested. In Chromium that left the native seekable range empty.
The published viewer now loads only the selected clip into a bounded in-memory
Blob, preserving every encoded byte and its cadence. It aborts unfinished
requests and releases the Blob when the tab or treatment changes. Each clip is
limited to 25 MiB; the first-20 study previews are all smaller than 4 MiB.
Playback waits for that selected clip to finish loading. No backend, external
media host, re-encoding or source-treatment change is involved.

The download page also has a bounded, horizontally scrollable command block for
narrow screens. Its extraction index updates only that HTML file's transport
hash and size, with a new companion index checksum. The original toolkit ZIP,
asset ZIPs, member hashes and render bindings are unchanged. The derivative
ledger records the source and public checksums separately.

Source excerpts and native source frames © OpenAI, included only in attributed
research comparisons. Full source films, private product material, local logs,
credentials, and research source imagery are excluded from runnable asset packs.
Original specimens are not claims about real product behavior.

## Independent reuse pilots

These consumer edits retain their inherited source limitations. They are not
additional completed source treatments.

### Considered tomorrow — two-shot passage

<video controls playsinline muted preload="metadata" src="reuse/considered-tomorrow.mp4" aria-label="Considered tomorrow independent two-shot edit"></video>

The paragraph-to-underline passage has a scoped fresh-review pass, including
the frame 82→83 reset cut. Inherited glyph differences remain open.
[Review report](reuse/brief03-review.html) · [Original MP4](reuse/considered-tomorrow.mp4)
· [Public render manifest](reuse/considered-tomorrow.manifest.json).

### Nook — different identity and exact copy

<video controls playsinline muted preload="metadata" src="reuse/nook.mp4" aria-label="Nook independent name-reel adaptation"></video>

The fresh reviewer recommends qualified reuse success. Pill-width differences
and the early npm-cache isolation uncertainty prevent an unqualified claim.
[Review report](reuse/brief04-review.html) · [Original MP4](reuse/nook.mp4)
· [Public render manifest](reuse/nook.manifest.json).

Download the [exact consumer inputs and original assets](reuse/reuse-inputs.zip)
and [member hashes](reuse/inputs-index.json). Extract beside the toolkit core so
their `project/` directories merge. From that project, after the documented
installation, render to new output paths:

```sh
npm run render -- --edit inputs/trial/considered-tomorrow.edit.json --out out/my-considered-tomorrow.mp4
npm run render -- --recipe type-masked-name-reel --content inputs/trial/nook.content.json --out out/my-nook.mp4
```

To verify the supplied original pilot movies rather than render again, place
each movie and its supplied manifest in `project/out/trial/` under its original
name, then run `npm run verify -- out/trial/considered-tomorrow.mp4` or
`npm run verify -- out/trial/nook.mp4`. The exact input and original asset files
from the ZIP must be installed. The public manifest redacts only diagnostic
locations; its render binding remains unchanged.

## Preservation

This release starts from the current live 4,962-file manifest. Existing files
retain their bytes except for one reversible navigation link on the benchmark
index. Existing videos, source maps, sections and downloadable projects are not
replaced. The full release manifest supports before/after verification.

The frozen review documents retain their historical pre-publication wording.
The user's [draft-publication authorization](APPROVAL.html) supersedes only that
publication status, not their source findings or incomplete-library status.