Measured 18 November 2026 on the unchanged qualified Linux CPU image.
Ordinary JPEG/H.264 recompression and the listed resizes retained the expected ID in all four scenes tested for each applicable profile. A three-second excerpt lost Economy5s in all four scenes; Economy2s recovered two, had one unconfirmed observation and one completed negative; Resilient recovered four. Cropping already-marked landscape video to 9:16 lost the ID in all twelve cases. These are observed counts, not estimated probabilities or a legal pass threshold.
Paired observations
Each cell is confirmed expected IDs / four source scenes. An unconfirmed observation is not a recovery. A completed negative applies to the declared current schemes; it does not mean human-created content.
| Operation | Image | Economy5s | Economy2s | Resilient |
|---|---|---|---|---|
| Unchanged signed export | 4/4 | 4/4 | 4/4 | 4/4 |
| Metadata removed without pixel/video-essence change | 4/4 | 4/4 | 4/4 | 4/4 |
| JPEG quality 90 | 4/4 | — | — | — |
| JPEG quality 75 | 4/4 | — | — | — |
| Image fit within 768px + JPEG quality 85 | 4/4 | — | — | — |
| Image center crop, retain 80% of each dimension | 4/4 | — | — | — |
| H.264 CRF28, same geometry and FPS | — | 4/4 | 4/4 | 4/4 |
| Video resize to 480px high, H.264 CRF23 | — | 4/4 | 4/4 | 4/4 |
| Keep 3 seconds starting at 0.2 seconds, H.264 CRF23 | — | 0/4 | 2/4 + 1 unconfirmed | 4/4 |
| Center crop landscape to 404×720 (9:16), H.264 CRF23 | — | 0/4 | 0/4 | 0/4 |
All sixteen original exports passed the unmodified export gate. Their independent checks recovered the matching ID and validated the C2PA signature and file binding, using development credentials. An edited copy can recover the ID without retaining a valid C2PA manifest for those edited bytes. ID recovery does not authenticate the edited file or demonstrate whole-video coverage.
Protocol and scope
- Four retained stock scenes: face, crowd, mixkit-1164 and mixkit-5016. Four extracted 1024×576 PNG frames and four 1280×720 SDR H.264 clips at 24 fps, 8.96–10 seconds. Source and prepared-input hashes, stream metadata and preparation commands are retained.
- One new ID per scene/profile, no repeated trials: four image exports and twelve video exports, then 96 verification observations. Rows and profiles share source scenes and are dependent; do not pool them into a durability percentage.
- m6a.large, 2 vCPU / 8 GiB host; container limited to 2 CPU / 4 GiB, two inference threads. Runtime source
49578688710a8bae2373aaf9d47ff9521f566d4c, immutable imagesha256:1a4dd358bcf698df654daaac631a26d3bc874104de5108bc94ce941c33a49980. - Encoder and verifier ran in separate sessions. Verification used a read-only recovery registry,
verification_scope="current", no signing key and no expected-ID/profile hint. The harness compared the report to the receipt only after verification. - Ordinary transforms use Pillow or FFmpeg. Exact codec, color, scaling and seek commands are in the archive. This is not a test of a named social platform, HEVC/HDR delivery or an adversarial remover.
- Engineering fixtures, not actual AI-generated outputs. Fixture provenance claims identify a test system and must not be reused as production generation claims. The purpose is to observe mechanics of the current exporter/verifier.
- Container CPU counters include Python, Node and FFmpeg; per-observation wall time and memory are retained. The first operation includes startup and later checks reuse a worker. This campaign is not a replacement for the isolated cold/warm performance benchmark.
Reproduce and inspect
Protocol · Corpus hashes and probes · Per-file CSV · Full results · Export receipts summary · Independent consistency audit · Raw reports, logs and executed scripts · File hashes.
The archive contains the full reports/receipts, container measurements, timeline logs, transformation commands and exact executed harness; it excludes source/derived media and private signing keys. Masters and transformed files remain on the experimental server's persistent volume. Source media are not redistributed here. Reproduction with different media is a new cohort, even when the settings match.
For an operator with authorized source files, inspect scripts/campaign.py and scripts/observe.py in the data directory before running them on an isolated Linux host with the pinned image and at least 5 GiB free. The campaign root and retained input locations are explicit in the harness. It creates an isolated development identity and registry; do not point it at production keys or an existing data root. Audit extracted records using python3 audit_delivery.py /path/to/extracted-raw-archive; the audit re-derives all outcomes from 96 raw reports and binds expected IDs to sixteen export receipts.
Practical use
Keep Economy5s for workflows that can retain its marking windows; do not describe it as reliable for arbitrary short excerpts. Economy2s adds opportunities but did not recover every tested excerpt. Resilient improved excerpt survival in this small cohort and still failed the landscape-to-portrait crop. Prepare the final aspect ratio before marking, and test the application's actual export/CDN path. Existing modes and defaults are unchanged.
There were no wrong confirmed IDs and no incomplete searches in this cohort. That is not a population false-positive estimate: this campaign has no held-out negative corpus. Broader footage, durations, frame rates, crop locations and randomized excerpt offsets, current remover tests and independent production CA/TSA trust remain separate work. Historical cohorts remain labelled with their original versions.