# Verification record — r02 iteration-01

Recorded September 13, 2026. These are builder mechanical results, not an
independent aesthetic verdict, source-fidelity pass or native Resolve run.

## Actual renders and clocks

Both delivered variants were actually rendered using the retained shared core,
exact pinned dependencies read externally, and full Chrome for Testing
152.0.7977.42. Its actual executable SHA-256 is
`9a1fec86405041cdf16c9977d022130c4caf96da6d1a9a20b1767b4ae370e8f6`.
It was not Chrome Headless Shell. No dependency installation, dependency copy
or dependency symlink was used. Encoders ran under the shared lock.

| Variant | Movie SHA-256 | r01 movie comparison |
| --- | --- | --- |
| default | `640290630fe7329bedc1cd29485bcadf9751371da9943f02229af9d0776ee0a0` | one bounded corner repair; new independent candidate |
| adaptation | `0d24b8f94f6edcafeaad6e70e69814a217c02e86f4d208d1b8ac8fed5e61698f` | byte-identical |

All 76 decoded RGB frames and timestamps were checked. Adaptation's movie and
all 38 lossless PNGs match r01 byte-for-byte. Default's lossless frames0–10
match r01; its remaining27 frames change by at most115 pixels per full canvas,
all inside the changed corner supports. Default's decoded movie frames0–9
match r01; frame10 also changes under inter-frame H.264 encoding although its
lossless PNG is unchanged. No locality claim is made for codec residuals.

Every timestamp matches r01:38 frames each,1920×1080,exact30/1 CFR, observed mux
time base1/15360 (512ticks/frame), silent H.264/yuv420p. The retained toolchain's
requested timescale30 did not force a1/30 mux time base; no normalization or
location-only rerender was performed. Source-window PTS and the960×540 CFR
source excerpt with its different1/30 time base are documented in `SOURCE.md`.
`REVISION.md` records the sole hypothesis, geometry change, all-frame locality
checks, negative geometry tests and25→26/30→31 boundary checks.

## Real SVG preparation, including discrepancies

Both preparation commands completed. Each generated 38 full-pose SVGs, 111
individual leaf SVGs, 38 opaque composite PNGs and 152 ordered RGBA PNGs. Every
preparation-manifest output hash was checked. All 76 full-canvas SVG composites
were compared with the corresponding lossless shared-core render frame; all
304 ordered RGBA frames were inspected for their expected alpha presence and
recomposited over white using straight-alpha Pillow composition.

**The SVG raster path is not pixel-identical to the Remotion raster path.**
Both paths use the same original coordinates, path data, fills and strokes.
The following measured residuals are disclosed, not turned into a new pass
threshold after seeing the output:

| Metric, maximum over 38 frames | default | adaptation |
| --- | --- | --- |
| Full-canvas changed pixels, SVG vs lossless Remotion | 351 | 347 |
| Maximum single-channel difference, 8-bit | 90 | 89 |
| Ink-mask XOR/union, luminance below 128 | 0.206165% | 0.168397% |
| Every corresponding ink bounding box equal | yes | yes |
| Ordered RGBA recomposition vs complete SVG: max channel difference | 2 | 2 |
| Ordered RGBA recomposition: max ink-mask XOR/union | 0.009975% | 0.009367% |

These are sparse raster differences at matching geometry, not proof that an
unexecuted Resolve import/export will match. The prepared composite is the
reference for the layered-plate workflow; the new Remotion movie is the sealed
candidate. Source-relative contour judgment stays separate.

`EVIDENCE.json` provides the portable summary, hashes and all-frame measurements.
Full generated binaries, detailed command logs, native source inspection crops,
and per-frame render verification live only in the separately supplied
off-volume evidence handoff. Every r02 binary was generated off-volume from the
outset. All r01 code, source evidence, movies and frozen review were preserved.

## Contract and packaging checks

- 24 automated Node tests cover exact regeneration of both examples, 111 path slots,
  all frame/occlusion/stroke controls, bounded design inputs and rejection cases,
  all 298 packaged SVGs, the guide's complete 38-row table, public asset
  boundaries and explicitly external output commands. Six join regressions
  reproduce the original defect, restrict the change to32 corner commands,
  freeze173 files plus choreography, test nine width/curvature combinations,
  and reject restored/misaligned/reversed corner handles.
- TypeScript 5.9.3 reports zero diagnostics with unchanged strict compiler
  settings and external dependency resolution, without stubs or copied types.
- Ten additional verifier regressions reject forged browser/hash/font fields,
  duplicate JSON inventory keys, duplicate-position lists, missing/extra PNGs,
  substituted frame hashes and forged counts. The receipt-bound verifier checks
  all four actual artifact directories, not just their self-reported counts.
- Four real preparation-command negatives (missing browser, wrong binary,
  invalid variant and existing output) fail with the expected diagnostic and
  leave existing evidence unchanged. The two positive preparation runs remain
  preserved separately. One adaptation preparation invocation ended with signal
  15 before allocating output; its logs/guard were retained, and a subsequent
  invocation completed. Neither actual movie was rerendered.
- The complete verifier binds the frozen manifest SHA-256, exact required
  code/font fields and unique path sets to `EVIDENCE.json`, then checks the real
  files. This is unsigned integrity under the separately supplied ZIP checksum,
  not historical machine attestation or protection against replacing the whole
  kit and receipt. The older low-level `verify.py` is not the full delivery gate.
- The public kit/ZIP contains code, original SVGs, controls, documentation and
  the permitted font. It contains no generated MP4/PNG, source clip, native
  research image, dependency payload or dependency symlink.
- Source evidence is not an implementation input. The SVGs contain no embedded
  image, text, script or source footage. Per-frame coordinates are generated
  from the existing original factory, not copied source identity paths.

## Open gates

The sealed r02 default is awaiting independent judgment of the single corner
repair. Adaptation retains the preceding static tie and is not rerolled.
Other source differences, including the early cover, remain unchanged. All38
native source/r01/default states were inspected as still crops, not moving
playback. No native Resolve execution, native SVG-import verification,
blind-review win or full source-fidelity completion is claimed. One existing
mechanism is extended; **new mechanism credit is zero**.