SEGMENT evaluation evidence
R4 local runtime, archive and import checks completed; official platform qualification remains unproven. The exact archive is e477d307ff2e09f469d83569ec1ce8703daeef34edbba633a3f240eab955de00 (35,613,293,722 bytes), immutable generation 1789092734891901. The build manifest is ba9ddb1e78a45040803ae6f0e68c66de7a538f5e3e86397777e5fc2c488106b0 and configuration 3f43fce1081dc5db454447dbe4678947e5cefec014635e611d8d1032a0ac71c4. Full compressed/raw archive verification covers all 26 ordered DiffIDs; actual same-generation classic Docker import reports config 3f43fce1081dc5db454447dbe4678947e5cefec014635e611d8d1032a0ac71c4 and no imported manifest digest. The agreed eight cases completed 9 application invocations and 1593 answer executions across 533 unique qIDs. Healthy20 and loose/ZIP500 matched saved content; reverse20 matched the same-image ordinary20, and both ZIP quotas matched the same-image loose500. Canonical repeated calls retained the exact disclosed deferrals. The mixed recovery case blanked only its declared failed questions and preserved later healthy outputs. These are bounded runtime/interface/content checks; no new semantic grade or official final result is claimed.
The completed development comparison is between two registered serving implementations under the same fixed Qwen3.5-27B judge and 20-case batching protocol: candidate 2373/4000 (59.325%) versus named reference 1940/4000 (48.500%), +10.825 percentage points. Strict sensitivity is 2334/3940 versus 1906/3940. All ten original-source-video differences are positive; 5b causal consequence reasoning regressed 17.742 micro points on 62 cases from six sources. These are adaptively used development results with inherited training/data-exposure limits, not an independent holdout, an official finals score or an all-candidate ranking. The registered local scorer passed14 convenience controls (6positive,8negative). These check response/parser boundaries; they do not estimate organizer-judge agreement, real-question false-positive rates, official calibration or a finals score. All 4,000 were not run through a Docker archive. Source-derived routing equality on the 4,000 intervals 1–299s does not by itself transfer these scores to new image bytes.
R4 retains the image-owned literal SEGMENT task and the original request clocks. It stages only the selected ZIP member for each startup probe or answer, verifies that member through CRC/EOF and SHA-256, closes decoder handles, and removes the lease before continuing. Loose input remains a direct read. Model initialization is still eager. An attributed ordinary per-question exception produces empty content for that qID with measured latency and failure telemetry; successful answers are preserved and later questions continue. Unsafe or ambiguous archives, source mutation, cleanup integrity, required model initialization, environment and fixed-track failures remain fatal. The inherited precise pre-generation rules deferrals remain separately disclosed as degraded coverage. The four changed image files are the bootstrap, contract, probe entrypoint and new ZIP helper; all 51 unchanged inherited serving source files, including cap_fmt, all 50 model/prior assets, prompts, model arithmetic and generation settings remain unchanged.
The full retained ZIP500 was qualified locally at the stated 4 GiB and 16 GiB tmpfs profiles, alongside loose500, healthy20, reversed20, default and persistent canonical calls, and controlled recovery. These fixtures do not establish the platform scratch quota, all possible input sizes, continuous resource peaks, cold-cache requirements, general order invariance, new semantic scores, independent-holdout performance or official platform success. Canonical deferrals and deliberately failed recovery rows are reduced scientific coverage. The prior R3 4 GiB eager-ZIP failure is a conditional local capacity reproduction, not proof of the organizer failure cause. Earlier images and failed attempts remain historical.
Measured grader completions: September 10, 2026, 10:53:44 UTC (candidate) and 10:55:13 UTC (reference). Complete readback and actual GPU-return checks passed for both. Paired reporting completed at 11:07:14 UTC.
| Population | Reference correct | Candidate correct | Reference micro accuracy | Candidate micro accuracy | Difference |
|---|---|---|---|---|---|
| Full operational board | 1,940 / 4,000 | 2,373 / 4,000 | 48.500% | 59.325% | +10.825 pp |
| Strict paired sensitivity | 1,906 / 3,940 | 2,334 / 3,940 | 48.376% | 59.239% | +10.863 pp |
| Strict-excluded diagnostic | 34 / 60 | 39 / 60 | 56.667% | 65.000% | +8.333 pp |
The full board has 400 cases from each of ten original source videos. Its equal-video estimate therefore matches the micro estimate: +10.825 percentage points, with descriptive paired two-level bootstrap 95% interval [+9.049, +12.501] points (1,000 draws, seed 42). All ten source-video differences are positive, ranging from +7.5 to +13.25 points. There are 650 paired wins, 217 losses and 3,133 ties. The strict sensitivity equal-video difference is +10.877 points, interval [+9.209, +12.588]. All score rows and parsing checks are complete; neither mode has a parsing failure.
These intervals are conditional on this development board and are unadjusted for repeated model selection or multiple capabilities. The board has inherited training and adaptive-selection exposure. It is not an independent finals holdout.
A material fine-capability weakness remains: 5b, causal consequence reasoning, regressed on 62 cases from six source videos. Its micro difference is −17.742 points and equal-video difference −16.881 points, with unadjusted descriptive interval [−34.671, −1.957]. This must remain visible alongside overall gains. The 34-case 5a micro result ties; its equal-video difference is slightly negative. These are local fine-capability results, not official capability-family/ID-OOD buckets or ranking votes.
The measured population compares the registered reference and candidate serving implementations with the same fixed Qwen3.5-27B judge, revision, local control check, hardware envelope and 20-case batching protocol. Candidate archive ed85dc914e846f0586c01e71a1427dd64caba0a80844b02e28feedfcb0dcce88 separately passed its exact-container 20-case H100 test. This report does not mean all 4,000 cases ran through that Docker archive, does not establish best across all historical alternatives, and does not predict an official finals score.
Strict-excluded pairs 0004, 0130 and 0196 remain distinct from recovered pairs 0074, 0111 and 0123. All three historical failed attempts retain their failed status. The fixed recovery policy selected whole pairs without using scores. No first-attempt reliability or model-only causal claim is made.
The paired report SHA256 is a73eedbcbdfeac8f8500d4730980984518209b035f9571d6c10293390ae7035b. Reference accepted observation is c613cfe00d4576f671cdcb587be378e2bf55e476c057f13bb741d9f604363cf1; candidate observation is 2b5f77c9f4a913cecaae405401ac05c15d045ba896d204ed8ff0d5759450d3d4. The monorepo custody record packages/rehearsal-segment/review/preparation/closed3-full4000-comparison-actual-r1/CUSTODY.json retains both actual returns, complete readbacks, commands, reporting inputs and machine-readable report.
This private model-card follow-up was prepared on September 10, 2026. The historical local runtime evidence in the following paragraphs applies only to archive ed85dc914e846f0586c01e71a1427dd64caba0a80844b02e28feedfcb0dcce88, Docker config d8fcfca448b374aab5d37c23382a02ef24367d7d66f6c5fbe194b65eb07f308b, and the unchanged immutable asset/metadata revisions in LOAD.md. The exact20 Docker process took 282.298466seconds, with sampled GPU memory maximum 43,589 MiB, no OOM termination and positive first-frame diagnostics for all 20 clips.
The untouched exact archive now also passed a captured 500-case, default-entrypoint H100 Docker test. The guard completed all 500 original answer calls with healthy model initialization, no recorded engine failures and zero guard input deferrals. Every answer content matched the candidate's evaluated serving implementation under B20 for these same 500 cases. This is not 100% semantic accuracy, not a content comparison against the historical pre-evaluation reference, and not evidence isolating a causal batch-size effect or establishing general order invariance. The 500 cases comprise 216 clips from original source0020 and 284 from 0021; prepared request.videoID values are clip aliases and are not independent source videos.
The actual Docker process ran from 12:49:28.234458 to 13:11:09.278933UTC on September 10, taking 1,301.044475seconds. The independent attach interval was 1,301.398154seconds. This attempt enforced 4 CPU cores and 32 GiB memory on one H100, with a 7,200-second admitted runtime cap; the registered500 profile's 9,144-second value remains separately identified. The answer-loop log reports 1,014.1seconds and the answer-latency sum 1,013.126716seconds; these are self-reported, not independent per-answer measurements. The host observer stopwatch excludes initial registration/input hashing and prior checks.
Across 584 retained resource samples, GPU memory reached 42,429 MiB and main-process RSS 6,453,096,448bytes. Cgroup memory.peak reached 34,359,738,368bytes (32GiB), including cache/tmpfs. Cgroup accounting is distinct from process RSS and VRAM; sampled maxima are not continuous peaks. Docker reported no OOM kill and cgroup OOM/OOM-kill event counts were zero. Nonzero memory-limit max events alone do not establish OOM.
Startup first-frame diagnostics were positive for 150 clips. The 60-second diagnostic budget expired, leaving 350 startup probes unmeasured; successful later answer processing does not backfill those first-frame proofs. These skips are distinct from the guard's observed near-still input deferrals, whose literal list was empty. Inputs were 500 loose clips plus request.json, with no serving-input ZIP. The separately accepted ZIP20 test below supplies narrow ZIP extraction evidence; organizer-platform and 500-case ZIP-capacity qualification remain separate.
The prior500 launcher stopped before START at its scheduling gate and remains a failed attempt, as do the earlier preserved loader/controller failures. The successful separately admitted run does not retrospectively turn them into passes or establish first-attempt reliability. The complete 500 readback passed at 13:11:29UTC. Acceptance SHA256 is c8410ed845b0875189ae55eb7ae8a1571f81c876e36c53c614b9f44c1b6751b6; original summary SHA256 is 47692048634f47758654f3b0bf24119ee041429b37955829f96a36b5c5f309d2; the source-label-corrected summary is 2d2f76b0702028ef52a738c1bbda5b72db0bb185edbc27c5832799df4f4a7fb2. The correction distinguishes clip aliases from the two original source videos and leaves all health/content/timing/resource results unchanged.
The same unchanged image subsequently passed a separately admitted ZIP20 test with the prior loose20 clips and identical question order. The serving input contained one ZIP archive and no loose media files; the guard resolved all 20 contained media members and completed all 20 answer calls. Every answer content matched the prior same-image loose20 output. The ZIP_STORED archive was 754,197,783 bytes, SHA256 9bc6931274c95ab84358fa86b24adf6ffcccc6bb0323e0b513ae94edd5c69436, with 754,195,577 uncompressed media bytes; this small archive fit the harness's 16 GiB tmpfs. It does not establish capacity for the roughly 21 GB captured500 corpus as a ZIP or identify organizer scratch storage.
ZIP20's actual Docker process took 148.082458 seconds and its attach interval 149.756455 seconds, within the separate 900-second admitted cap. There were 67 retained runtime/resource observations, zero guard input deferrals, no OOM termination and completed scoped cleanup. These process durations are observations under different run/cache conditions, not a causal speedup measurement. There was no independent post-extraction file-hash census. The complete metadata readback finished at 13:15:25 UTC; receipt SHA256 is 53532b03e69bac7cecd9bb98051e164400c766a7cf3045658fe2e9f9fa0e6058. The result is local same-image ZIP20 extraction/execution evidence, not organizer-platform qualification.
The private repository does not imply final submission sign-off. The earlier ancestor evaluation follow-up changed only README.md, EVALUATION.md and EVALUATION.json; all 50 assets, immutable LOAD revision pins, training/exposure disclosures and existing 4,000/3,940-case statistics remain unchanged. Raw videos, gold answers, per-case outputs and judge data are not added by this presentation update.
Historical 43d4 local evidence, followed by failed canonical platform tryout
Historical 43d4 local evidence. The historical 43d4 archive is 43d4fc4fd7a39e4884a5179176a6b6a436956a146494876783d6287855a82825 (35,613,238,247 bytes, generation 1789057108435274). It has configuration f7e22f1822d7a152ec91999dbb639ec63378346ef3f5144723c5b453748f97b9. The resident build manifest is 4eb060d790e39af7b56dcc263b73a0a039ded5cd4dd898c0bda33e6e13c5bf63. The CPU worker's classic Docker store loaded the exact configuration ID sha256:f7e22f1822d7a152ec91999dbb639ec63378346ef3f5144723c5b453748f97b9 and all 24 ordered DiffIDs; it did not expose an imported manifest digest. That field remains null. IMAGE_IDENTITY.json records their full joins. The original ed85 platform failure remains preserved; it is not a failure of the historically tested derivative, which later failed the canonical interval admission check.
The complete replacement-image acceptance combines 6 invocations and 600 answer executions across two retained attempts, not 600 unique questions. The first default 20-answer invocation completed with healthy guards and matching contents; its enclosing attempt then failed when a case-sensitive absence parser rejected Docker's lowercase no-such-object message after successful removal. That raw failure and watchdog return code 125 remain failed. A separately claimed continuation used one persistent container for linked loose 20-input, real-root loose 20-input, linked ZIP 20-input/FO, repeated ZIP 20-input/FO, and linked loose 500-input. All accepted guards passed with empty input-deferral lists. Repeated FO contents matched; the FO fixture was reconstructed from the image package and is not an independent capture of organizer socket bytes. Non-FO answers matched the frozen candidate baseline; these are content comparisons, not new semantic grading. The 500-input case's independent host child interval was 1140.333899 seconds; startup first-frame probes were positive for 151 clips and did not cover the other 349. Successful answers do not fill those missing startup proofs. The combined independent host child intervals total 1718.896785 seconds, including the first attempt's 115.555531 seconds. Self-reported answer latencies, attached child intervals and sampled GPU/cgroup measurements remain separate in the machine-readable metadata. Cgroup memory.peak is cumulative within the persistent container, rather than an isolated per-invocation maximum. Library tokenizer/processor/dtype warnings were retained; no behavior change or fallback is inferred from those warnings. Local persistent tests do not attest the deployed shim version, organizer execution or full 500-input ZIP capacity.
The earlier nine-file image reconciliation changed five existing metadata files and added four files, leaving 77 prior files unchanged at 86 total. All 50 model/prior assets and component loaders remain byte-identical; selected training/annotation facts, original 4,000/3,940 statistics, 5b regression and inherited/adaptive exposure limits are preserved. Original MODEL_FILES.json and SERVING_ASSETS.json retain ancestor image metadata because the immutable loader pins their raw bytes; current-image scope is explicit in the new identity/compatibility records and additive LINEAGE fields. No final form or organizer declaration is submitted.
The subsequent LOAD.md presentation-only update identifies the fixed serving archive through the current IMAGE_IDENTITY.json and PLATFORM_COMPATIBILITY records, and excludes three historical presentation files from the immutable metadata download. Asset revision A, metadata/tools revision M, all 50 asset bytes, and the verification/component-loading commands remain unchanged.