Proposed coverage for the consolidated review
These are production targets, not completed examples or a claim that 210
distinct treatments have already been measured. A source-discovery candidate
does not enter the completed catalog until it satisfies CONTRACT.md.
| Primary family | Proposed distinct examples | What must differ |
|---|---|---|
| Openings | 16 | Construction, emphasis, hold and handoff into the next shot |
| Typography | 24 | Grouping, anchoring, reflow, replacement and departure |
| Camera moves | 22 | Path, focal target, blur, settle and camera/content separation |
| Interface interactions | 24 | Intent, state construction, timing and readable outcome |
| Masks | 12 | Reveal geometry, clipping relationship and boundary motion |
| Selection/highlighting | 16 | Region, drag order, persistent selection and emphasis transfer |
| Overlays | 12 | Screen/scene anchoring, construction, stacking and exit |
| Lower thirds | 10 | Spatial fields, onset relationship and removal mechanism |
| Layered compositions | 14 | Density, depth, trajectories, occlusion and foreground handoff |
| Montage | 10 | Rhythm, image relationships, field occupancy and pacing |
| Transitions | 14 | Source-visible mechanism and exact entry/exit relationship |
| Shot continuity | 12 | Preserved anchors, scale, object state and attention across edits |
| Logo reveals | 12 | Original identity construction, order, material and settling |
| Endings | 12 | Exit cadence, retained information, mark and final hold |
| Primary-family total | 210 | Ten above the required 200; no duplicate counting |
At least 30 of these should be multi-shot passages. That is a cross-cutting tag, not 30 additional entries. A passage qualifies only when its coordination is materially distinct from the isolated recipes already counted.
The first 20 are a breadth-and-contract review batch, selected from actual downloaded OpenAI footage. Research scouts are re-inspecting source frames rather than promoting sampled historical notes into measurements. The five approved benchmark treatments remain a separate preserved reference; they are not automatically credited toward the new completion test.
Source discovery starts with the 100-film corpus and its shot/motion/opening, logo and overlay studies, then expands through the official-channel inventory where coverage is thin. Multiple URLs for the same creative execution do not yield additional treatments. Repeated stock wordmarks/lower thirds with only different copy are consolidated. A family target can be revised at the review if the actual source evidence does not support that many distinct mechanisms.
Scalability choices to review
One typed layer/timeline engine plus source-specific recipes keeps edits reusable. The alternative, one bespoke component per example, duplicates timing and asset handling and makes independent adaptation harder to audit. The opposite extreme, one generic animation preset with 200 cosmetic variations, fails the distinctness requirement. Each recipe therefore retains measured native timing and geometry while sharing the mechanisms that render them.
The site loads a compact searchable index, then only the selected recipe and
video. Full-resolution consecutive frames remain in downloadable evidence
packages; the browsing UI uses contact sheets and a single synchronized video.
Package splitting and a file-budget check will precede publication. Cloudflare's
documented Pages ceiling is 25 MiB per asset and 20,000 files on the free plan
(official developers.cloudflare.com/pages/platform/limits/, checked September
5, 2026). Do not assume a paid-plan upgrade or move files to another host.
The measured first-batch sizing and explicitly hypothetical 200-treatment
projection are in PUBLICATION_PLAN.md. They support testing split archives,
not a claim that the unbuilt 200-treatment release fits or is deployable.
Evidence-backed assumptions
| Claim | Basis | Status / consequence |
|---|---|---|
| The existing benchmark is approved, not a blank slate | ../remotion-fidelity/publication/APPROVAL.md and the live benchmark |
Preserve bytes and disclosed limitations |
| The local public build is stale relative to live | Local manifest has 3,886 entries; the fetched live manifest has 4,962 at task start | Rebase against the latest live manifest before approved deployment |
| The current toolkit dependency versions can install | evidence/clean-install.log |
Pinned to benchmark versions; new code still requires its own checks |
| The corpus can support every proposed family budget | Existing research inventory, not yet a 210-treatment census | Unverified; never backfill this gap with cosmetic duplicates |
| A declarative engine can express the first 20 faithfully | Source inspection and prototype rendering underway | Reject unsupported source mechanisms rather than simplify them |
Independent reuse trials will include at least 20 unseen briefs, cover all major families, and include several sequences. Trial agents receive the frozen toolkit and original asset packs, not builders' notes. They choose recipes themselves; implementation edits and undocumented help invalidate the trial. Render validity, fidelity, timing, continuity, readability and substitution are separate checks.