# R01 — unblinded output and GUIDE critique

September 13, 2026 UTC. **Source quality remains failed; 0 wins, 0 accepted additions.** This record concerns R01 only, not the pending repaired timeline. Source and label freezes are unchanged. Unblinding was authorized after the label freeze; preferences below are not revised after seeing identities.

## Coverage and limits

Static per-native-frame inspection: **1,104 source frames first, then 2,208 new candidate frames**. The label pass inspected 3,312 frame instances because it also repeated the same 1,104 source frames. All consecutive triplet sheets were inspected, not sparse contact sheets. Crops and rechecks add no unique frames. Comparisons used 960×540 tiles; candidates are natively 1920×1080, sources 960×540. JPEG evidence and viewer fitting further limit fine detail. **No moving review, moving blind win, or source-blind certification.** Source identity was already known.

## Frozen preferences and largest gaps

Frames are zero-based, local/native, inclusive. Guide paths below are toolkit-relative. Evidence filenames are **private**, relative to the critic worker; they are not public assets.

| Piece | Unblinded filenames | Frozen new-output preference | Single largest gap |
|---|---|---|---|
| `chatgpt-traveling-image-trail` | A = adaptation; B = source; C = default | A = C | At f53–81 and f118–138, similar pale card fields and centered motifs merge into a stepped pale mass, weakening the readable history. The issue is tonal/content separation, not a demand for source photographs. |
| `codex-word-tiles-collage` | A = default; B = source; C = adaptation | A, weak static preference; choreography tied | At f39–60 and f96–113, terrain foreground/background relief compresses into a distant range fading into fog. Materiality is uncertain; this is **not an established blocking choreography defect**. |
| `prism-translucent-contours` | A = adaptation; B = source; C = default | A, weak static preference; choreography tied | At f48–177, tight contour-band lighting reads as a recessed ring/funnel rather than broad overlapping translucent faces. Filled bands and independent angular changes exist; “outline-only” or “one global zoom” would be false. |
| `sora-moving-result-pullback` | A = default; B = adaptation; C = source | A = B | Similar input strips at f0–126 lead to a fixed-hub rotor at f127–299; especially f256–299, no legible evolution between distinct forms/materials delivers the blend payoff. This is not a frozen-media defect. |
| `sora-timeline-insertion` | A = adaptation; B = source; C = default | A = C | Both candidates break opening camera continuity at f4 and f7–11, returning to the normal pan at f12. This is an observed R01 release blocker; no code cause is assigned here. |

The known source remains the static composition preference in each triplet (weak/uncertain for Codex), **not a blind win**. New-output preferences do not certify source parity. The irregular trail insertions/rests, Codex insert/removal sequence, Prism layer development, result hard cut/independent inner motion, and later timeline hover/insertion/reframe remain present. Those observations narrow the gaps; they do not erase them.

## GUIDE and actual-code audit

**No surviving portable-instruction error was established in this bounded pass.** This is not GUIDE approval or exhaustive code verification. All five pinned-guide checks match the actual R01 recipe and SVG assets. Read-only `prepareRender` accepts both documented variants for all five recipes (10/10). No fresh encode or Resolve session was run by this critic. Mechanical success proves neither composition nor fidelity.

- **Trail:** `guides/collage/chatgpt-traveling-image-trail/README.md:59` explicitly uses the installed recipe's complete content objects, not old examples. The separate persistent planes and native gates survive. The 17-slot substitution still produces the frozen tonal-history gap; correct gating does not fix that art direction.
- **Codex:** `guides/collage/codex-word-tiles-collage/README.md:59` pins the active override and retains separate terrain details, chart, and miniature. Local motion exists. Neither a missing-micro-motion allegation nor a mandatory photographic-subject claim survives. The weak terrain criticism is visual, not a demonstrated portable API error.
- **Prism:** `guides/collage/prism-translucent-contours/README.md:61` specifies registered layers, independent transforms, filled translucent faces and a luminance rim mask. Those are materially different from an outline-only implementation. The narrow-band/recessed reading remains a rendered material/composition gap; Fusion mask, filter and color behavior is not verified here.
- **Result:** `guides/sora/sora-moving-result-pullback/README.md:269` explicitly specifies the rotor, native inner clock, palette fields and associated thumbnails. Actual code also changes blade scale and vertical position: the frozen visual shorthand “rotation-only” does **not** mean rotation is its only animated property. The more precise gap is continued rotor-like motion without a distinct form/material transformation. The guide acknowledges unresolved subject deformation at line 306; it does not promise a reproduced morph.
- **Timeline:** The pinned R01 instructions reproduce an asset/recipe binding, not a continuity guarantee. The decoded opening defect affects both variants. Parent owns the camera-sample investigation and repair; this pass deliberately does not duplicate root-cause coding or inspect the repaired guide.

**Rejected instruction-error allegation:** old trail example JSON fails on missing legacy JPEGs, and old result example JSON fails the active content-key contract, in both variants. These read-only failures reproduce, but the actual guides expressly prohibit those stale example routes and prescribe installed recipe variants/custom content instead. They are legacy-example hazards, **not retained failures of the documented workflow**. No double-isolated refutation agents were available; no code bug receives a “Confirmed” designation.

## R01 timeline native evidence — PRIVATE

Both **A.mp4 (adaptation)** and **C.mp4 (default)** show the defect. Native extraction retains original 1920×1080 pixels; lower-right crops use **x=960, y=540, w=960, h=540**, without scaling. Evidence covers every f3–13 for both labels: 44 still files, not new unique frame coverage.

- `timeline-native-r01/A-f004-full.jpg`: high crop at f4.
- `timeline-native-r01/A-f007-lower-right-960x540.jpg` and `timeline-native-r01/C-f007-lower-right-960x540.jpg`: tiny whole UI at f7 in both candidates.
- `timeline-native-r01/C-f012-full.jpg`: normal framing resumes at f12.
- Exact companion filenames are `{A,C}-f003-full.jpg` through `{A,C}-f013-full.jpg` and `{A,C}-f003-lower-right-960x540.jpg` through `{A,C}-f013-lower-right-960x540.jpg`, under `timeline-native-r01/`. `timeline-native-r01/index.json` binds every still's SHA and crop coordinates.

Those four highlighted native stills were additionally viewed; all corresponding frames were already inspected in the complete R01 label pass. Source B is spatially continuous through f3–13. f5–6 return to the card view, f7–9 miniaturize the UI, f10–11 use an intermediate scale, and f12 restores the pan. Neither the native resolution mismatch nor ordinary source blur explains those large layout changes.

## Disposition

R01 remains **changes required**; source quality failed, no accepted additions. This report is complete as a bounded first-pass record, not a release approval. The repaired timeline requires its separate label-only R02 freeze before further unblinding. Exact input hashes, coverage, guide bindings and freeze hashes are in `REPORT-R01.json`; the original `source-freeze.*` and `LABEL_REVIEW.*` remain authoritative and untouched.