The 200-example build and validation plan

Updated September 15, 2026. This is an execution plan, not a claim that 200 examples have passed review.

Accepted R20 checkpoint

work-200-r20 is accepted at 17:19:54 UTC on September 15, 2026. It combines Title-chunk's first dual-tool package with the same-ID Forecast R03 correction: 64 packaged candidates / 65 registered, 53 ordinary guides / 106 ordinary exports and three separate legacy completions. It adds one package, not a registration, ordinary-export count or quality approval. Both existing render pairs are retained without another encode in assembly. All 8,273 served payloads /3,431,427,852 bytes pass both the full delivery verifier and the browser's byte sweep. Four ordinary profiles each check 159 players, 53 guides and 36 source-film pages. Four legacy profiles and actual downloaded Title/Forecast ZIPs pass; the two ZIPs run 26 literal commands and 77 shipped tests. Title also checks three included movies across four profiles. No older route is removed, and the personal-project manifest remains unchanged.

The receipt is publications/work-200-r20.json, SHA 1cdc28cb9f301cdc718cd4bf74aea8e9b39187e4d87ae8cb2a1cfd7ae3b7c853. Quality approvals, moving wins, native Resolve, whole-film completion and clean corpus-pass credits remain zero. Type-chip is the one registered item without a package; its diagnostic preparations do not resolve the historical mismatch. Pulse's cause and repair isolation remain unassigned. Source intake and novelty reviews continue separately, with no new admission from the current lanes yet. This repository copy records post-deployment acceptance; the frozen plan in R20 retains its explicitly dated predeployment state. Do not redeploy accepted R20.

Earlier accepted R19 checkpoint

work-200-r19 is accepted at 14:39:29 UTC on September 15, 2026. It has 63 packaged candidates / 65 registered, 53 ordinary guides / 106 ordinary exports (98 fresh and eight retained in the selected reports), and two separate legacy completions. Task-card adds one first package, not another registration or a quality approval. The source-clock frame-step correction is also published. All 8,243 payloads / 3,422,257,480 bytes pass the second full live sweep. The actual downloaded Task ZIP passes 16 literal commands, 60 shipped tests and four movie/browser profiles. Ordinary and legacy profiles pass locally/live; the personal project remains unchanged. The first full sweep's one timeout is preserved, not rewritten into a pass. A supplemental default-client HTTP403 is also preserved; the normal verifier-client manifest check passes separately.

The exact receipt is publications/work-200-r19.json, SHA ff2350c7f6f11e694a95b0b9715677433b965b448de81fbdd43b143ace274017. Quality approvals, moving wins, native Resolve execution, whole-film completion and clean-corpus-pass credit remain zero. Title public R03, Forecast public R03, Composer R06, Type-chip recovery and Pulse diagnostics remain outside this release. Wave17 adds zero admitted leads. Two private video-only source recoveries and both fresh source-only orientations are verified. Dense observations were sealed before whole-contract comparison; all six bounded Wave18 proposals are rejected as additive. Neither film nor the corpus is exhausted. A third private source is in fresh neutral orientation, not admitted. Title's first-package integration passes independent technical review. A combined private R20 candidate with the same-ID Forecast repair proposes 64/65, with no new ordinary exports or quality credit. Composition review and full delivery checks remain. These private artifacts do not change the accepted counts. Do not redeploy R19.

Earlier accepted R18 checkpoint

work-200-r18 is accepted at 12:13:40 UTC on September 15, 2026. It retains 62 packaged candidates / 65 registered, 53 ordinary guides / 106 exports and one separately counted legacy completion. Permission and Deep are revised, not newly counted. The sampled-review accounting correction is now published. All 8,241 served payloads, both actual downloaded revised kits and local/live browser profiles pass. Quality, moving, native Resolve, whole-film and clean-pass credits remain zero. Forecast R03, Composer R06 and Task-card R03 remain private; their renders, critics and delivery boundaries are separate gates.

Earlier accepted R17 checkpoint

work-200-r17 is accepted at10:47:52UTC on September15:62packaged candidates, 65registered items,53ordinary guides/106ordinary exports andone separately counted legacy completion. It adds Forecast R01 only. Its two actual renders retain the published C2 density gap; a same-ID R02pair is rendered and undergoing fresh anonymous review. Local/live checks cover8193served payloads, four ordinary and four legacy browser profiles at each stage, and the actual downloaded new kit. The personal project remains unchanged. All quality, moving, native Resolve, whole-film and clean-corpus-pass credits remain zero.

PermissionR03 is not included. Its exact ZIP passes, but its old packet falsely uses270/540 clip extents instead of111/228 sampled review frames. A typed accounting correction is implemented privately and under independent integration review; no corrected packet or later release is accepted yet. Earlier live Permission content stays preserved with its existing limitations.

Wave14 has zero admissions and Wave15 is active. Task-card and ComposerR05 have actual paired renders with verified native clocks, but their new work remains outside the accepted count pending review and separate legacy delivery boundaries. The R16 paragraph below records the earlier checkpoint, not the current count.

work-200-r16 is accepted at09:18:13 UTC on September15: 61 packaged candidates,64 registered items,52 ordinary guides and104 ordinary exports. Composer is the first separately recorded legacy completion. It adds one package, no registration and no ordinary export or frame-coverage credit. All8,160 served payloads pass local/live checks. Four ordinary and four legacy browser profiles pass at both stages. The actual local/live downloaded Composer ZIPs each pass55 tests in the main reports plus18 in separate extra-core annexes. The original reports are not relabeled as73-test runs. All moving wins, source-quality approvals, native Resolve executions and whole-video or clean corpus-pass credits remain zero. The personal site remains unchanged.

Wave12 now has zero admissions; Wave13 is in progress. Forecast's two actual renders await completion of anonymous review. Task-card's recovered original code now renders both contents, with294 native frames and25 file bindings each; it is not packaged or published yet. DeepR04 and PermissionR03 remain private review/integration work. Pulse's separately reviewed repeat capture shows within-window variation but does not identify its cause or isolate the historical default repair. No further Pulse pair is authorized. These lanes add no current delivered count. R16 must not be prepared or deployed again.

Earlier R15 checkpoint

work-200-r15 is accepted at 07:03:51 UTC on September 15: 60 packaged candidates, 52 additional guides, 104 content exports, 64 registered items. All 8,158 served payloads and four local/live browser profiles pass. All three new or revised actual downloaded ZIPs pass, including eight additional helper checks across three kits. This checkpoint adds Deep Research18 and revises Sora Recut03 and Permission without counting those repairs as new examples. There are 49 changes-required and three moving-review-outstanding statuses. No quality, moving, native Resolve, whole-video or clean-corpus-pass credit is added.

All three new/revised public packets have bounded independent technical closure. Deep retains a close-framing gap. Permission improves sampled closing opacity while entry foreground isolation remains unresolved. SoraR02 has no established material contradiction under bounded static criteria, not moving-quality approval. The earlier Recording reconstruction snapshot repair and release-tooling corrections remain closed. Type has qualified empty-aperture improvement, Recording restores a stop-event sequence, and Gallery removes sampled note collisions. Their other source-relative gaps and motion limits remain. R14 and R15 use the fail-closed deployment wrapper. R13's separate HTTP403/shell-fallthrough exception remains in its immutable receipt; it is not retroactively erased. Wave05's single admitted lead is packaged. Wave06's single admitted lead has actual R02 and R03 renders and separate source-first correction reviews. The reviewed R02 still has a final-panel framing gap; the R03 render is not approval. Waves07 and08 yield no admissions. Wave09's one source-bound question is now independently nonadditive, with zero allocation. Wave10 now has one qualified admission: a forecast parameter changes a prospective graph segment and linked aggregate while historical data stays fixed. Three nearby source contracts were compared; this is not an exhaustive novelty claim. Its first two renders await fresh source-first comparison. Wave11 has zero admissions; Wave12 is active. The existing, already-registered composer-token project now has a recovered public kit with 73 delivered tests and bounded independent technical closure. A separate legacy-completion ledger records packaging without rewriting the historical 12 registrations and eight packages. It cannot add another registration. Its lower-menu-row visibility gap remains open. Native Resolve and moving review are still outstanding. Repairs stay same-ID. Later release totals come from the live workbench, not these dated accepted-baseline figures. The later narrative below retains its dated earlier checkpoints and limitations.

What counts

One example is one independently useful motion or editing treatment, tied to an official OpenAI source and a specific source-relative interval. Alternate copy, another color, a revised render and individual layers do not create extra examples. Related occurrences remain browsable without multiplying the count.

The starting baseline for this plan is 18 registered work items, including six additional dual-tool guide kits and twelve fresh Remotion movies. The earlier workbench has eight packaged candidates. None of these numbers means 200 quality-approved examples. The research queue contains 211 provisional clusters, only 111 of which have named treatments. The other 100 require actual mechanism definition and distinctness review before implementation.

For delivery progress, the starting package count is 14: eight earlier packages plus six new guide kits. The other four registered work items are not counted as delivered packages. The next five-kit release therefore moves the package count from 14 to 19, not from 18 to 23. Registration and source-quality acceptance remain separate counters.

Verified checkpoint history

work-200-r01: 14 packaged candidates, including the six-guide pilot.
work-200-r03: 18 packages. Four additional kits ship after a mobile-wrap correction. The failed private r02 candidate is retained, never deployed.
work-200-r04: 19 packages. The fifth kit ships after an early-camera repair and actual public-ZIP handoff/control-export checks.
work-200-r05: 24 packages. Five more kits and exact-source film backlinks.
work-200-r06: 29 packages. Five more kits after source-attribution and guide wording corrections. All 7,493 served payloads and four live browser profiles pass. Static review is 5,051 source / 10,092 candidate frames, not moving review.
work-200-r07: 29 packages. One reviewed cursor-entry repair, zero additions. All 7,498 served payloads and four live browser profiles pass. No new unique inspection or quality-completion credit is added.
work-200-r08: 34 packages. Five more kits pass independent packaging review and actual downloaded helpers. All 7,578 served payloads and four live browser profiles pass, including 26 guides, 78 players and eighteen source-film pages. Static review is 6,138 source / 12,266 candidate frames; quality counters stay zero.

These are historical delivery counts, not quality approvals. The live workbench's progress JSON records later releases. Both workbenches retain earlier artifacts; revisions do not add examples. Source-first reviewers and the parent are now checking further five-kit groups while existing kits receive same-ID repairs.

The public progress page keeps these separate:

Research leads: references worth investigating, not delivered examples.
Registered treatments: distinct work units with an implementation owner.
Delivered dual-tool kits: source-bound SVG/code, runnable Remotion, two rendered content examples and a treatment-specific Resolve guide.
Quality accepted: the delivered kit passes source and independent output review, including the requested moving comparison. Packaging cannot confer this status. The final target is 200 distinct examples, not 200 files.

Per-example acceptance record

State Python 3.11+ for the handoff helper. The core extractor's separate Python 3.10+ minimum does not cover this helper. Export control tables from raw helper output and check bytes, not just parsed field values.
Official film link, source publication date and cutoff eligibility; source hash, half-open frame interval, rational cadence and playable silent excerpt.
An explicit choreography ledger covering entry, action, secondary motion, overlap, holds, exits and the neighboring edit. Separate observation from measured values, estimates and adaptation choices.
Editable code and original/permitted SVGs. No flattened film frames, source branding or raster/video dependencies in the adaptation inputs.
Default and changed-content Remotion exports, full input hashes, decoded frame counts, timing checks and disclosed comparison differences.
A Resolve/Fusion construction guide with layer hierarchy, coordinate spaces, node order, numeric controls, easing, masks, alpha, text and export settings. Per-frame control tables aid reconstruction. They are not native Resolve project files or proof of Resolve pixel equivalence.
A fresh critic freezes a source-first record before receiving builder findings. Candidate labels are removed for an actual blind comparison when moving-view capability is available. The critic names the single largest gap; the builder repairs it, rerenders and returns the changed revision.
Every discrepancy remains public until retested. Static native-frame review, browser playback tests and numeric comparisons never become invented moving reviews, blind wins, sound checks or native Resolve renders.

Build loop

Read Poteto principles and the hillclimb/orchestration playbooks. Done.
Freeze the unit definition, acceptance gates and baseline. Do not weaken the definition to reach the count. Keep an append-only attempt/evidence log.
Pilot a complete kit through source binding, guide, render, critique and publication. The six-kit pilot is delivered with explicit repair requests.
Run a rolling window of at most three builders with disjoint unit folders. Independent critics receive sources first. One parent owns counts and releases; one encoder owns the render lock.
Improve the largest per-piece gap, measure the changed artifact and retain improvements only after regression checks. Unresolved items stay candidates.
Publish every five additional delivered kits, and publish material repairs promptly. First merge the existing six-kit preview into full production.
Reconcile distinctness, source eligibility, artifacts and all live links before the 200 checkpoint. Report every failed or unavailable gate. The original whole-channel/two-clean-pass objective remains separate and unmet.
Commits, PRs and final all-hands production are skipped: not authorized here.

Storage and publication

Use cloudflare-work, the Quora account and the existing work Pages project. The personal-account site is preserved. The user authorizes 100 GiB and repeated publication; no further permission is needed for each staging snapshot.

Two approaches were considered. Keeping a fresh three-gigabyte full tree for every five examples wastes the allowance on unchanged files. Instead, keep one bounded reusable publication workspace and immutable release manifests plus changed-payload archives. The frozen original stage and the ordered deltas can reconstruct every release. Verify preservation before and after each upload. Never edit the frozen historical stages or delete sources or evidence.

Pages remains the simplest current host. Use R2 only when an artifact exceeds Pages limits or the cumulative artifact workload warrants it. If needed, create an ordinary work-account bucket such as g-jones-work-logs, record object sizes against the allowance and expose only authorized silent excerpts and original editable assets. Full source downloads and private logs stay private.

Release gate

Each release verifies source/recipe/SVG bindings, render provenance, silent playable media, downloads, navigation, search, frame stepping, looping, slow motion, reduced-motion behavior, no-JavaScript fallback and narrow layouts. Hash-check every served changed payload and preserve every old route. Recheck the work account and previous production manifest before deployment. Verify the new live workbench and record its deployment ID and manifest afterward.

Throughput checkpoint

Blocking: establish the new allowance, reusable release path and full production merge. The parent does this work rather than waiting on builders.
Independent: collage/trail/contours guide builder; two Sora guide builder; source-only critic; later builders take non-overlapping five-unit queues.
Shared state: only the parent edits release manifests, counts and common tools. Encoders serialize. Each revision has one owner and one review record.
Smallest unit: one complete, independently judgeable treatment. Do not create a new example for every keyframe, asset, render or alternate wording.

Known limits

Frozen source queue wave03 admits 11 leads, not 30 or 211 ready units. Its 211 clusters comprise 61 excluded, 11 selected, 25 rejected and 114 deferred. The deferred set includes 100 occurrence-only entries without a bound native window and 14 named entries with unresolved native cadence. Even packaging every admitted lead does not establish capacity for 200 distinct examples. Further source binding and distinctness review remain necessary; repairs, variants and renamed occurrences cannot fill this shortfall.

Wave04 separately binds seven of those deferred leads, excludes one equivalent and leaves 106 deferred. Its retained master-header/PTS joins resolve the fourteen named cadence questions without rounding excerpt FPS or claiming fresh master access. Independent source-first novelty review admits four for bounded builders: detached working preview, gallery-to-device, image cutout/recombination and simulator inspection-to-instruction. The other three remain deferred. They remain part of the same 211 clusters, not seven new clusters or delivered examples. Wave04's initial 32 checks pass; later rechecks retain a stale PUBLICATION pin failure after R09. The additive allocation receipt binds R09 and the independent delta comparison instead of rewriting the old queue or claiming all old checks pass. Two wave03 rechecks also report shared-input drift; its queue/seal stay intact. The four eligible kits are built and combined as Batch10, with eight fresh content renders. Source-first review is sealed; anonymous output and public-kit checks remain separate. Wave05 binds twelve more leads but provisionally admits only one, sora-recut-03; eleven remain deferred. A fresh source-first novelty critic is comparing that one against existing contracts. None earns output credit.

R11 delivers 49 packaged candidates after full live checks, adding Batch09's five independently checked kits. R10 added Batch06's five kits as candidates, not quality completions. Exact deployment and delivery receipts are in PUBLICATION.json. The type comparison has a split result: the revised adaptation clears the sampled overlap, while the default's reveal spacing regresses. That gap is returned for a separate R03 correction. Batch09's five source-relative structural gaps remain explicit; recording-to-artifact is returned for a bounded R02 correction. Batch08 preserves the interrupted review and failed fresh-context qualification. A new critic completed source-first, key-withheld full-index chronology, then authorized reconciliation. Four gaps remain; skill disclosure has no established material structural discriminator, not approval. Earlier sampled evidence is not retroactively promoted to full-index coverage. All reviews remain static only.

Pulse R04's unchanged recorded default inputs do not establish unchanged pixels. A parent-only pair of identical default-input executions shows actual pre-encode RGBA differences at four frames. A codec-only explanation is insufficient for that instrumented pair, but its initiating cause and the historical mismatch's cause remain unresolved. No default-isolation or source-quality pass follows.

The six-kit pilot has two nonidentical reproduction outputs and six outstanding visual repair requests. There are zero new native Resolve renders, perceived moving-source blind wins or complete whole-video reviews in this extension. The environment's unavailable review capabilities must stay visible; a passing static check cannot replace them. These limitations do not block publishing useful, clearly labeled candidate kits, but they block a quality-complete claim.