# Real production proof

## v1.6.0 host-agent planning migration

- The Repo's directions, script, and scene are now authored by the host coding agent and imported through the existing shared validators. The separate Wiggly `/create` planning provider is unchanged.
- Two synthetic offline regression inputs (wooden-toy history and cookie gifting) reach a validated official scene without API keys or network calls. Invalid directions, evidence IDs, scene plans, and mismatched script selections are rejected before media generation.
- These are planning and runtime checks, not newly generated or creatively approved videos. The production references below keep their original provenance.
- An independent extracted-package check exposed a missing JSX configuration: source-tree smoke passed by inheriting the app's config, while the standalone renderer reported `React is not defined`. The kit now includes its own `tsconfig.json`, and its commands use `node --import tsx` without an IPC launcher. Archive tests guard against missing config and stale packaged runtime files.
- After that fix, a fresh extraction passed the standard `npm run smoke` using dependencies installed independently from the kit's package manifest, rather than the Wiggly workspace dependencies.

## Canonical successful proof

The FinalStraw Day 5 video is the current Style B reproducibility target.

- Script: five concise beats moving from hidden product to customer friction, mechanism, resolved use, and buyer-action CTA.
- World: one bright blue/cyan blueprint-grid explanation stage across the entire story.
- Demonstrator: one silent semi-stylized 3D man with a dark backward cap, white sunglasses on the cap, pink polo, and tan cargo shorts.
- Media: one six-frame storyboard, full-quality anchors at frames 1, 3, 4, and 6, and two 10-second Seedance clips.
- Human review: clip 2 was rejected and regenerated before approval.
- Post-production: existing voice and late clip action were retimed locally rather than buying replacement provider calls; captions preserve the approved script wording.
- Final: one 1080×1920, 20-second MP4 with one audio stream and a four-second product/CTA hold.

See `goldens/finalstraw.mp4`, `goldens/finalstraw-contact-sheet.jpg`, and `fixtures/finalstraw-reproducibility.json`.

## Supporting creative quality evidence

The Repo also includes three earlier Wiggly production references in `goldens/`, with provenance, reviewed narration, strengths, and known weaknesses in `goldens.json`.

- **Grüns:** the earlier Wiggly baseline. It proves the product can travel into an explanation world and return to a clean payoff, but its abstract hook and repeated tunnel shots remain weaknesses to beat.
- **Kiala Nutrition:** the clearest routine-compression example. A pile of separate supplements becomes one unified product system.
- **Theragun:** the clearest physical-mechanism example. A familiar customer action becomes a cutaway that shows why heat changes the result.

FinalStraw establishes the reproducibility target. The supporting references broaden the comparison set. None authorize copying brand claims, wording, or shots.

## LEGO technical proof

- Input: LEGO's documented company-origin article.
- Direction: rewind from a modern brick to the first documented 1932 wooden-toy product line.
- Planning: five directions, one selected Style B script, six storyboard frames, and two 10-second clip plans.
- Paid image attempts: one storyboard plus full-quality frame-1 and frame-4 starts. The old runner then enlarged tiny storyboard panels into frame-3 and frame-6 endpoints.
- Video: two 10-second Seedance 2.0 Mini clips were generated in one attempt from those mixed-quality inputs.
- Result: `failed-paid-media-quality`.
- Human review: the clips are not good enough to run as ads. Their ending images are visibly soft, the payoff appears too early, and clip 2 invents a different CGI person and setting.
- Root cause: planning thumbnails were used as production media, while old prompt rules forced a recurring presenter and three hard-cut subshots even when the approved story was object-led.
- Historical correction: the runner was changed to require four independently generated endpoints (frames 1, 3, 4, and 6), preserve object-only frames, and request 480p. The later FinalStraw proof showed that forcing one continuous transformation was itself too restrictive.
- Technical note: the old clips are 10.042 seconds at 496×864 and contain no audio stream. Passing those mechanical checks did not make them creatively acceptable.

See `agent-runs/lego-origin-proof/` for the inputs, scene contract, images, attempt history, and quality report.

### Corrected 480p proof

- Direction: a failed carpentry livelihood turns heavy wooden work into the first easy-to-sell toys.
- Planning: the locked GLM 5.2 path produced a 48-word script, six-frame plan, and two continuous clip contracts.
- Paid image attempts: one Nano Banana 2 Lite storyboard plus four separate full-quality endpoints at frames 1, 3, 4, and 6.
- Video: two 10-second Seedance 2.0 Mini clips at 480p, generated and inspected one at a time.
- Result: `continuity-passed-style-b-failed`.
- Human review: the ladder, carpenter, workshop, wood, and toy car remain coherent; both clips use continuous physical motion; the final car is sharp; no unrelated person appears. However, both clips stay in the warm workshop and omit Style B's recognizable blue explanation stage.
- Technical review: both clips are 10.042 seconds at 496×864, the duration gate passes, and the quality report contains no problems.
- Historical correction: this run introduced a lifestyle-to-blue-to-lifestyle world sequence. The successful FinalStraw proof later replaced that rule with one coherent blue-grid world across all six frames.
- Boundary: no voice or final 20-second composition was generated.

See `agent-runs/lego-origin-quality-proof/` for the scene, four production endpoints, both clips, contact sheet, attempt history, and quality report.

### Style B world-arc technical proof

- Direction: LEGO survives the 1930s downturn by turning difficult-to-sell carpentry into simple wooden toys.
- Planning: the approved scene moves from the warm workshop into the blue/cyan breakdown world, then returns to the workshop for the payoff.
- Paid image attempts: one Nano Banana 2 Lite storyboard, four accepted full-quality endpoints, and one replacement for a rejected final endpoint that contained fake article text.
- Video: two separately approved 10-second Seedance 2.0 Mini clips at 480p, with no retry or duplicate generation.
- Result: `style-b-clips-passed`.
- Human review: clip 1 keeps the carpenter and wooden objects coherent while moving from the workshop into the blue grid. Clip 2 assembles the wooden car in the blue world, returns to the same workshop, and ends with a finished wooden-toy collection. Neither clip invents visible text or an unrelated person.
- Technical review: both clips are 10.042 seconds, the duration gate passes, and the Repo inspector reports `clips-ready`.
- Provider finding: the first Replicate prediction took 319 seconds and the second took 1,725 seconds. The foreground runner incorrectly stopped polling before both healthy predictions completed. Their exact prediction outputs were recovered without retries. A future runner must persist and resume provider prediction IDs instead of extending a foreground timer.
- Voice: one Fish S2.1 Pro Free narration, 19.193 seconds.
- Final: one 1080×1920, 20-second MP4 with audio, rendered locally through `AdRenderSurface`.
- Final result: `technical-passed-marketing-failed`. The automated report found no mechanical problems and the contact sheet preserves the approved world arc, but the hook does not quickly establish LEGO or a strong viewer problem. The finished file is evidence that the pipeline works, not evidence that the ad is good enough to run.
- Spend note: the final proof reused the accepted clips and made no new Replicate call.

See `agent-runs/lego-origin-world-arc-proof/` for the accepted endpoints, rejected endpoint evidence, scene, clips, narration, final MP4, contact sheets, attempt history, and quality report.

## David's Cookies stopped attempt

- Input: a selected blueberry-pie product from a real saved site-research run.
- Result: stopped before selection or image generation.
- Reason: the proposed directions relied on unsupported product claims because the available evidence did not specifically support the chosen product.
- Lesson: product-led stories must validate that their evidence belongs to the selected product before approving a direction.

See `agent-runs/davids-blueberry-pie-proof/` for the research, proposed directions, and stopped state.