New: release-specific CPU measurements, first-job and resident runs ↗
SDK 0.7.2a1 · documentation review 2026-11-10.
Use this index to select the right evidence for a claim. Engineering tests, small robustness experiments and population performance estimates are different things. Historical measurements are not relabelled as new 0.7.2a1 results.
| Record | Actual scope | Supported conclusion | Boundary |
|---|---|---|---|
| 0.7.2a1 qualification | 217 installed-media tests, core matrix, clean installs, kit/HTTPS/reports/registry, external c2patool | Current packaged workflow passes those checks on Linux CPU | Development identities; no current population durability or throughput benchmark |
| 0.7.1a1 paired media | Two stock scenes, 10 s and prepared 3 s renditions, three profiles, identity transforms and remover | Measured short/excerpt/remover boundaries on that corpus | Two source scenes are not all content; repeats/IDs are dependent |
| Reliability observations | Production-Node 22 selected clean 720p clips; four positive controls, auxiliary reference analysis | No recovered ID in those scoped clean checks; controls recovered | Not a population FPR or cross-provider detector guarantee |
| 0.7.1 pipeline | Colour/FPS/BCH fixes and focused release tests | Versioned pipeline regression evidence | Earlier exports/CPU conditions must remain identified |
| 0.7.0 recovery | Signed reports and detached manifest integration | Signature/ID/binding distinctions | Not authentication of arbitrary edited copies |
| Historical profile selection | Legacy Fast/shared residual versus per-frame candidate | Earlier technology-selection rationale | Not Economy5s/2s or current installed-SDK throughput |
| Legacy VideoSeal | Older CPU encoder/stock corpus | Compatibility/history of old format | Not the current default engine |
| Spatial research | Geometry recovery feasibility | Research rationale and measured limitations | Experimental cropdetect is not silently enabled in the release verifier |
Current cadence integration observation
The 0.7.2a1 6 s/12 fps fixture has 72 frames. Economy5s marked 4 frames in two windows; Economy2s 6 frames in three; Resilient all 72. This distinguishes the implemented schedules. It is not a durability percentage or benchmark ranking.
Historical speed and resources
The paired 0.7.1a1 Linux/m6a.large record reports the following means over two exports per cell, including worker startup, signing and read-back:
| Input duration | Economy5s wall / cgroup CPU s | Economy2s wall / cgroup CPU s | Resilient wall / cgroup CPU s |
|---|---|---|---|
| 10 s, 720p24 | 13.92 / 22.77 | 17.29 / 28.78 | 94.69 / 166.22 |
| 3 s, 720p24 | 5.81 / 8.44 | 6.48 / 10.40 | 29.85 / 51.95 |
These are original-version/corpus measurements, not a 0.7.2a1 performance claim or SLA. Current container memory limits are deployment limits, not measured RAM consumption. A positive early-stop scan must not be compared as if equivalent to an unmarked full scan. Use P6 for a new matched comparison; do not combine Mac/server or old adapter timings.
Observed short and removal limits
In the same two-scene record, all modes recovered the two fresh 10 s and two fresh 3 s exports, and all four identity CRF23 transcodes per mode. A 1–4 s excerpt recovered 0/2 Economy5s versus 2/2 Economy2s and Resilient; an excerpt between Economy windows recovered 0/2 both Economy modes versus 2/2 Resilient. A leading 0.2 s trim of the 3 s exports recovered 0/2 Economy5s versus 2/2 the other modes. These are scenario counts, not estimated probabilities.
The paired public-remover image experiment removed IDs in 4/4 trials across two
scenes/two IDs; matching identity controls recovered. Resilient video after removal
recovered by aggregation for crowd, while face returned not_found. Crop and
letterbox remain known limitations. A correct ID alone never authenticates an
edited copy. See the complete versioned record for parameters and logs.
Public claim rules
Publish claims with SDK version, source/CI identity, machine, corpus and exact stage/transform scope. Do not describe test count as attack count, sampled reads as whole-video confidence, a memory cap as RSS, or standard C2PA parsing as public trust/conformance enrollment. Include failures and incomplete outcomes.