# 0.7.1a0 — video pipeline and decoder hardening candidate

This candidate extends the qualified 0.7.0a0 wheel. All new runtime tests,
benchmarks and media experiments are performed on the Oregon EC2 host. The
existing public preview services are unchanged. No model, payload, strength,
watermark cadence or native dependency pin changes.

- **Color:** explicitly preserve SDR matrix/range through YUV→RGB→YUV, use
  accurate rounding/full chroma interpolation, and tag the encoder. Economy
  keeps its unselected-YUV bypass. Read-back uses the same conversion policy.
- **Fractional FPS:** a deterministic observed-phase bank stays within the
  existing 512-hypothesis limit at NTSC rates. 64 phase bins/second and 93.75ms
  recovery windows are reported explicitly. Media PTS are preserved.
- **BCH:** Python and Node acceptance guards re-encode a recovered BCH5 payload
  and require a canonical codeword within five errors of the received 96-bit
  data/parity word. Headers are outside that code. This is defensive validation
  of external decoder output, not provenance authentication or FPR calibration.

## Qualification

Final installed-wheel Linux qualification: **165/165, zero skips, 339.665s**,
CPU2/6GiB, UID10001, network disabled. Python3.11/3.12/3.13/3.14 clean-core
installs each passed 109 executed checks (56 optional/media checks skipped).
Those skips are covered by the separate full-media run, not counted as passes.
All tests and media experiments ran on EC2; no GitHub-hosted runtime CI was
started in this work. Qualified SDK code commit: `4933202`.

A fresh isolated Python3.12 virtualenv exported and verified all three video
presets without importable Torch, NumPy or Python TrustMark. C2PA setup used the
pinned bootstrap; Node/ONNX models were the preprovisioned checksum-verified
cache. This is not a claim of a new clean Docker build or fresh model download.

All **55** packaged resources matched both the installed wheel and qualified
source. Archive inspection found no private keys, media, model binaries or
unsafe archive paths. Own SDK wheel/source are separate from combined binaries.

Preliminary color/FPS build: 162 tests, 161 passed; the remaining example-queue
test lacked its `/examples` fixture mount. It passed after the mount was fixed.
An additional matrix/range export test passed separately. This preliminary build
does not include the later canonical BCH guard; the complete final run above
is the qualification result.

BCH tests exercise actual codec correction at 0–5 bit errors, deterministic
random controls and a deliberately incorrect upstream success outside the
correction radius. A small uniform random sample did not reproduce a
noncanonical success; the tests do not claim it did. All early failed test
attempts and their correction are retained in the evidence logs.

## Measured behavior of the final decoder

Two 720p stock fixtures were converted to 24000/1001 and 30000/1001, six seconds
each. The final build exported all three presets and recovered the expected ID in all 36 checks
(fresh file, H.264 CRF28 and 90%-dimension center crop plus resize). Four clean
negatives exhausted their bounded bank without a recovered ID; the NTSC banks
used 210 and 273 hypotheses. Separate ten-second controls reproduced the old bug: 0.7.0 decoded 240/300
frames but exhausted 512 hypotheses and returned `not_checked`; 0.7.1 completed
254/334 hypotheses and returned scoped `not_found` for both source clips at
both rates. This small corpus is not population calibration.

The heavier 640×360/CRF32 transform recovered IDs in 0/4 cases for Economy5s,
0/4 for Economy2s and 3/4 for Resilient. These failures remain visible; the
ordinary 36/36 result must not be generalized to severe compression. No wrong
recovered ID was observed in the test corpus.

For six-second 720p clips (two originals × two rates, one export per combination,
a mix of cold/warm per-profile worker states), SDK export means were 9.17s for
Economy5s, 10.74s for Economy2s and 61.98s for Resilient. These are measurements,
not SLAs. Approximate sampled whole-container peak was 1.928GiB, not a minimum
RAM guarantee or per-export peak. Native portrait renditions prepared before
marking passed six exports and 12 fresh/delivery checks after square-pixel
normalization; see the [integration guide](../implementation/final-renditions.md).

Identity-marker exports isolate conversion error: for the face clip's Resilient
path, average raw-Y error changed from approximately -0.94 to +0.03 levels out
of 255. Real watermark exports on these fixtures also had small mean luma error.
This is not a guarantee of no local/model/compression artifact on arbitrary
content. Matrix/range reference and export tests cover BT.601/BT.709 in both full
and limited range. Source primaries/transfer tags are preserved, not converted.

## Spatial scope and remaining limits

[Ready crop/letterbox approaches](../research/spatial-recovery-2026-11-06.md)
are documented separately. A server-only FFmpeg cropdetect probe is outside the
SDK's qualified verifier. Geometry detection, watermark recovery at reduced
resolution and whole-composite attribution are separate questions. Do not
advertise severe 16:9→9:16 crop or letterbox support based on a single example.

Known temporal removal/copy attacks, limited held-out false-positive calibration,
partial legacy-engine installations, operator trust/certificate/TSA requirements,
provisional C2PA algorithm registration and native redistribution qualification
remain explicit. These fixes do not certify Economy, Resilient or a product as
legally compliant. Signed reports/recovered manifests retain 0.7's trust limits.

Exact artifacts, source/resource digests, scripts, failed preliminary attempts and
results are recorded in [the validation record](results/0.7.1a0.json) and the accompanying release-evidence
bundle. Raw test media and generated private identities remain on EC2.
