First 20 — consolidated review

September 5, 2026. Local review only. 20 authored treatments; 0 completed examples; no publication approval. The approved benchmark is unchanged.

This review brings together the proposed coverage, reusable API, first source choices, actual paired renders, independent findings and reuse pilots. It does not ask you to accept an unresolved difference by approving the general plan. Further production beyond this batch waits for your response.

The selected-generation index now links one hash-matched current source review for each of the 20 treatments. review-provenance.json identifies the review documents, selected artifacts and verification records. Historical renders and reports remain separate; they cannot supply a current verdict.

Open the work

From the repository root, run:

python3 benchmarks/remotion-library/scripts/serve_review.py

Open the loopback URL printed by the server. The searchable review has original and different-content movies, synchronized research comparisons, source mapping, native-frame evidence, editable inputs and hash-bound findings. The index is a checked snapshot, not a browser-side live verification service.

The runnable project and agent guide are the consumer entry points. An agent retrieves a subset instead of reading all recipes or this review history.

Decisions in this review

1. Coverage: approve or revise the family budget below. It targets 210 distinct treatments, with at least 30 multi-shot passages included in that total. It is a production target, not a claim that 210 qualifying sources have already been found.
2. Reuse contract: approve or revise the native-timing, typed-input and rejection rules. They preserve the selected treatment rather than silently shrinking text, slowing movement or replacing its mechanism.
3. First 20 source choices: review the actual treatments and their content-substitution boundaries. Technical defects and material source differences remain uncounted until fixed or explicitly accepted by ID.

Approval of coverage or the contract is not approval of every example. No material departure, inherited font difference or inspection limitation is waived by default. Publication requires a separate approval of a frozen release.

Coverage proposal

Multi-shot is a cross-cutting tag, not an extra count. A passage must demonstrate distinct coordination; merely concatenating two existing recipes does not create a third treatment. The original five approved benchmark examples are preserved separately and are not automatically credited toward this new completion test.

Discovery starts with the existing 100-film research corpus and continues with actual OpenAI films where coverage is thin. Every counted treatment needs its own inspected interval and consecutive evidence. Family quotas may change if the sources do not support enough distinct mechanisms. See COVERAGE.md.

Reuse contract

The toolkit uses one typed layer/timeline renderer with source-specific recipes. Shared mechanisms cover text, masks, media, layered groups, camera transforms, directional exposure, pointers, identity paths and explicit shot joins. The book, wordmark and reverse-selection examples also have high-level content builders; consumers do not have to author hundreds of pose paths.

The current batch is 1920×1080, 16:9. Native rates are kept as exact fractions, including 24000/1001 and 60000/1001. Other aspect ratios require reviewed layouts, not an automatic crop. Mixed-rate passages map elapsed time onto an explicit output clock; native single-shot renders remain available for source review.

Unsupported content throws. A character count does not prove fit: the loaded licensed font is measured before rendering. Sketch, for example, does not fit the reel's 181px first-row box and was rejected in an independent trial. The brief was not rewritten, and the type was not made smaller.

Source footage, OpenAI identity and demo data stay in attributed research comparisons. Reusable outputs use original or permitted assets. The illustrative interfaces are not claims about real Poe behavior; real product captures require separate Poe2 inspection and privacy checks. No all-hands presentation is being produced here. Full rules: CONTRACT.md.

Find and render a subset

From project/:

npm ci
npm --silent run catalog -- search "emphasize a word" --duration '<=3'
npm --silent run catalog -- show study-keyword-double-underline --contract
npm run render -- --recipe study-keyword-double-underline --variant adaptation --out out/underline-adaptation.mp4
npm run verify -- out/underline-adaptation.mp4

For new copy, use the selected record's typed Content contract or high-level builder and pass its JSON with --content. Do not edit recipe tracks to force a brief to fit. Search accepts intent, family, source, duration, input kind and neighbor ID; it pages at 20 results by default. Neighbor matches are discovery aids, not approved transitions. Exact setup and portable-package instructions are in the project guide.

First 20 source treatments

Frame intervals below are zero-based, with exclusive ends. Each treatment link opens its selected original/adaptation pair and full source URL, evidence and findings. Durations follow those frames, not rounded timing guesses. None of these rows grants completion credit.

Findings that affect the decision

Fresh reviewers inspect actual moving source/output pairs and native frames, not just metadata or successful exports. All 20 current reports are bound to the selected movies and inputs. They report nine material and 27 minor open findings within their stated scopes. The later Nook trial adds a separate pill-geometry qualification; the source-report totals are not an exhaustive count of every possible difference.

Five scoped reviews report no open findings: dock subtraction, chip retract, double underline, word-tile collage and translucent Prism contours. That is not user approval, an exhaustive pixel-equivalence claim, or completion of the library's remaining reuse and publication requirements.

The other nine source reports retain minor findings: task card (3), composer (2), reverse selection (3), click-release camera (3), modal relay (3), timeline insertion (1), masked reel (2), paragraph cuts (1) and staggered identity (1). Their exact frames, geometry/timing observations, methods and limits are in the linked records. In particular, a targeted retest is not a full-source waiver.

Several differences are substantive, not polish requests. The book still has different leaf/hinge contours. The wordmark's four-/three-lobe original identities do not reproduce the source's six-lobe density and internal finish. Photographic or recurring-subject source passages cannot be treated as faithful merely because their original procedural substitutes move on the same path. These remain open and uncounted; no acceptance is requested by implication.

The licensed Geist substitute is not an identified source font. Specific glyph/sidebearing differences remain disclosed even when timing and outer bounds match. Reviewers' numeric envelopes and pixel counts support particular findings; they are not universal fidelity thresholds.

The fresh Nook reuse review also measures fixed reel-pill spans that differ from source: 436/384px source/output at local15 and 404/362px at59, for example. Those are inherited recipe differences, not consumer edits. Earlier targeted retests of fade, suffix and ending motion do not close this broader geometry question. The trial remains qualified; the source example remains uncounted.

What the verification establishes

The current review03 clean snapshot installs pinned dependencies, passes both type checks and all 136 tests with no skips, and renders/verifies all 40 variants. Thirty-one rerenders are byte-identical to the selected movies; nine are not. All nine differing pairs have been decoded and quantified in evidence/clean-review03-pixel-summary.json. Shipped snapshot files and the generation registry remain unchanged. Rendering is not claimed to be bitwise reproducible; its variation does not establish a cause or perceptual acceptability. The earlier review02 pixel comparisons remain a separately identified record, not a substitute for the current run.

Every comparison binds its actual code, inputs, assets, source movie, native frame mapping and outputs. Important boundaries include consecutive frames; the gallery distinguishes them from sparse contact sheets. Reviewers disclose re-encoding, reduced-size playback, sampled visual observation and dropped-frame limits. Playback counters are not continuous human viewing.

Two consumer pilots produced real edits using only frozen delivered packages: a mixed-native-rate paragraph/underline passage and an exact-copy Nook name reel. Their mechanical audits pass with unchanged embedded recipes. The paragraph/underline passage has a scoped fresh-review pass, with no new adaptation or continuity defect, including its frame 82→83 reset cut. Inherited source glyph differences and inspection limits remain open; this does not accept the source example. The Nook pilot receives a qualified reuse-success recommendation: choreography and exact content survive, while source-pill geometry and an early npm-cache isolation uncertainty remain unresolved. Two earlier attempts failed and remain recorded: a missing packaged config, then an unsupported exact-copy width. The packager and discoverable fit limits were corrected; those failures were not relabelled as successes. A separate known-overwide paragraph brief is not counted as an executed trial.

The eventual library still requires at least 20 successful fresh-agent briefs across every major family and several multi-shot sequences. See trial status and validation history.

Preservation and proposed publication

The local baseline contains 18,679 pre-existing files. Its last check finds no changed or missing file. This does not prove live preservation. The task-start live release has 4,962 manifest entries, while the local public build is stale; any approved release must rebase on the latest live manifest.

The proposed addition is /videos/editing/benchmark/library/ in the existing openai-film-fieldnotes-100 Cloudflare Pages project. The old benchmark, sections and downloads remain intact. Native evidence archives, a compact index and per-recipe asset packs are proposed to keep the enlarged site usable. The local native-evidence prototype now verifies 6,582 PNGs in 78 split archives, each below 24 MiB, with attributed frame/member indexes. The complete 200-case staging budget, runnable/asset downloads, privacy-safe public derivatives and post-publication checks are not complete. No Gsites, replacement project or alternate public host is authorized. See PUBLICATION_PLAN.md.

Approval record

No coverage, first-20 exception or publication approval is recorded. A later acceptance must name the affected treatment, exact frozen generation and specific discrepancy. Anything unresolved stays out of the completed count.