# 0.7.1a1 — native runtime preflight and paired media review

This extends the qualified [0.7.1a0 video pipeline](0.7.1a0.md). The SDK model,
payload, strength, temporal selection and default preset are unchanged:
`economy_5s` remains default, `economy_2s` remains configurable, `resilient`
still computes a fresh signal on every frame. Legacy compatibility is retained.
Qualified SDK code commit: `35cf34d4b794f0ba65588ba6f8d0b49f9040a24c`.

All new runtime/media tests were performed on the Oregon m6a.large host,
Linux x86-64 / Python 3.12, with 2 vCPU and a 6 GiB container limit. Runtime
tests and media processing had no network access. The remover was provisioned
separately from upstream before offline experiments. No media tests ran on Mac.
Existing public services were not redeployed. This is an installed-wheel
qualification on a previously provisioned image, not a fresh Docker bootstrap.

## Runtime preflight

`ariadne runtime check` requires c2pa-python 0.38.3 / native 0.91.11 and the
SHA-256 in `c2pa-bootstrap.json`. It checks the actual ctypes handle's file,
the native API version, and Linux mapped-file identity; cwd/library overrides
cannot silently select another build. See [startup and recovery](../operations/c2pa-runtime.md).

Actual server probes accepted a copied qualified library, rejected native
0.91.7 under the same Python wrapper, rejected a changed binary still reporting
0.91.11, and rejected an atomic pathname replacement after library loading.
Mismatches prevent SDK/media and supplied HTTP-verifier startup. Low-level
signature inspection returns `not_checked`, not a forged/absent credential.
The 0.91.6 case is a unit fixture, not an actual Linux binary experiment.

The first preflight after wrapper import took approximately 70 ms; 100 cached
checks averaged 0.40 ms. This measures the check itself, not complete Python,
native-library or model startup. No hash is computed per video frame. This
compatibility/drift check is not a sandbox against hostile native initializers.

## Installed-package checks

| Run | Result |
|---|---|
| Focused runtime/process checks | 17 passed, zero skips |
| Full real-engine suite, network disabled | 177 passed, zero skips, 336.166 s |
| Clean dependency-free Python 3.11 / 3.12 / 3.13 / 3.14 | 121 executed per interpreter; 56 optional/media skips per interpreter |
| Requirement traceability | 15 records resolve; structure check, not legal certification |

The complete media suite covers those optional paths separately. Skipped core
checks are not counted as passes. The full suite ran against the installed
wheel from outside the source checkout, with read-only filesystem and bounded
temporary storage. Package/source parity and archive inventory are recorded
with the distributable evidence. The final source archive also contains this
post-test documentation; the tested wheel is retained unchanged. Its README
metadata predates this validation page, while package payload parity is checked
against both the installed wheel and final source.

## Short clips and public remover

The paired experiment uses two existing 1280×720 / 24 fps / 10 s stock fixtures
(`crowd`, `face`) plus three-second renditions prepared before SDK export.
It is a small scenario check on two source scenes, not population statistics.
Blind video search receives no expected ID, export profile, schedule or registry.
Expected IDs are used only to score completed search results.

The scenario matrix covers fresh exports, H.264 CRF23 identity transcodes,
1–4 s excerpts, 0.2–1.8 s window-free excerpts and a 0.2 s leading trim of
three-second exports. FFmpeg seeks are quantized to the 24 fps frame grid.

| Blind ID recovery | Economy 5s | Economy 2s | Resilient |
|---|---:|---:|---:|
| Fresh 10 s export | 2/2 | 2/2 | 2/2 |
| Fresh 3 s export | 2/2 | 2/2 | 2/2 |
| Identity CRF23 transcode, 10 s and 3 s combined | 4/4 | 4/4 | 4/4 |
| Excerpt 1–4 s of 10 s export | 0/2 | 2/2 | 2/2 |
| Excerpt 0.2–1.8 s, between Economy windows | 0/2 | 0/2 | 2/2 |
| Leading 0.2 s trim of 3 s export | 0/2 | 2/2 | 2/2 |

The failed Economy cases returned bounded `not_found`; they were not budget
exhaustion or foreign-ID results. These are scenario counts on two originals,
not estimated probabilities. Both original unmarked controls returned scoped
`not_found`. The complete paired remover result is recorded alongside the raw logs.

Mean SDK export wall / cgroup CPU seconds (two exports per cell; includes
worker startup, signing and final read-back; not throughput/SLA):

| Duration | Economy 5s | Economy 2s | Resilient |
|---|---:|---:|---:|
| 10 s | 13.92 / 22.77 | 17.29 / 28.78 | 94.69 / 166.22 |
| 3 s | 5.81 / 8.44 | 6.48 / 10.40 | 29.85 / 51.95 |

These are this host/corpus's measurements, not a before/after speed regression
against the earlier different-duration or Mac measurements.

Image remover trials use two source frames with two independently generated
IDs each: four exports, not four independent source images. Both the identity
control and remover output are RGB PNGs with C2PA metadata stripped. The video
pair uses the SDK's corrected color/timestamp/encoder pipeline for both the
identity control and public TrustMark `remove_watermark()` on every frame.
All four image identity controls recovered the expected ID. The public remover
removed the readable ID in all four image trials (two scenes × two IDs), with
45.03–49.62 dB PSNR relative to the marked image and 4.05–4.44 s wall time per
image. Both control and treatment lost embedded C2PA; the watermark difference
is therefore not attributed merely to metadata stripping.

Both video identity controls recovered their ID after two frames. After the
public remover, `crowd` recovered the correct ID only by cached-logit aggregation
after all 240 frames (zero individual ID reads); `face` exhausted 240 frames
and returned `not_found`. No foreign ID was observed. Remover processing took
1108.70 / 1110.77 s wall and 2121.15 / 2129.05 cgroup CPU seconds respectively.
Median per-frame PSNR was 45.70 / 49.51 dB relative to decoded marked RGB,
measured before final H.264 encoding. This is not output-video PSNR or evidence
that every frame survived the attack. Both runs used 240 remover calls.

Whole-job cgroup memory peak was 2,454,904,832 bytes (about 2.29 GiB), including
image and video engines, their caches and file-cache charges. It is not process
RSS, a per-stage peak or a minimum RAM requirement. The two real-child startup
probes also passed: normal preflight with cleanup, and child-only native drift
rejected before HTTP bind. [Machine-readable record](results/0.7.1a1.json).

Only Resilient video is attacked in that paired remover trial; Economy's
known temporal-window removal boundary is not repaired or hidden by this work.

Wall time and cgroup CPU time are separate fields. CPU accounting includes
resident Node and FFmpeg children. Whole-job memory peak includes loaded image
and video engines; it is not a per-export peak or a minimum RAM guarantee.
All parameters, source/output hashes, full blind-search results and official
remover model hashes remain in the JSONL evidence.

## Interpretation

A fresh short export and an excerpt of an existing export are different cases.
Economy's selected frames can all be removed by a short edit. A shorter period
reduces the gaps but cannot protect every arbitrary excerpt. Resilient is a
per-frame robustness option, not a guarantee against a public targeted remover.
For distribution workflows, prepare final cuts/aspect ratios before marking
and qualify the transformations actually used by the product.

The earlier crop/letterbox and severe-compression limits remain documented in
[0.7.1a0](0.7.1a0.md) and the [video guide](../implementation/video-profiles.md).
These tests do not establish a legal frequency minimum, certify Economy for
compliance, calibrate a population false-positive rate, or turn a recovered ID
into content authentication. All three modes remain available unchanged.
