vernier

Error bars on a dataset vendor's quality claim. Build AI publishes hand-visibility and active-manipulation rates for Egocentric-10K / Egocentric-100K, judged once by gemini-2.5-flash with no human gold, no interval, and no test that the judge scores a factory floor and a home kitchen on the same scale. This release is the data behind an independent, pre-registered measurement of that claim: human labels against a written rubric, a live open-weights judge on the same frames, Build AI's own stored labels, and every computed result. No image is redistributed in this dataset -- frames are identified by frame_id and fetched from the vendor's own release.

Full writeup, with the argument and the charts: https://caio.theodoro.dev/blog/vernier-judge-errors-run-one-way.

Code, protocol and decision log: https://github.com/caiotheodoro/vernier. Frame-by-frame view: Space. Distilled probe (negative result): model. Machine-checked verdict: MEASUREMENT_CARD.json in this repo -- 18 claims, 1 items that could not be checked, verdict NOT_VERIFIED.

What was found

Live judge: Qwen/Qwen3-VL-8B-Instruct-FP8, self-hosted, temperature 0, prompt P0a (Build AI's own wording), N = 10,000 frames per release.

figure	measured	published	diff (pp)	within ±2 pp
Egocentric-10K, ≥1 hand	95.4	96.4	0.97	yes
Egocentric-10K, 2 hands	82.7	76.3	6.32	no
Egocentric-10K, active manipulation	91.3	91.7	0.38	yes
Egocentric-100K, ≥1 hand	96.1	97.0	0.86	yes
Egocentric-100K, 2 hands	85.2	79.0	6.14	no
Egocentric-100K, active manipulation	92.1	92.8	0.62	yes

Same judge, same gap on the same figure (2 hands) on both the superseded and the current release. The 100K rows are a disclosed extension, not pre-registered.

Against human gold (one rater, n = 153 primary labels, rubric v1.2.0):

• Rubric is decidable: intra-rater AC1 0.876 (hand count), 0.899 (manipulation), n = 34 blind re-labels, gate 0.70.
• Judge–human AC1 0.860 (hand count, 95% CI [0.791, 0.921]), 0.887 (manipulation, 95% CI [0.815, 0.945]).
• PPI++-corrected active-manipulation prevalence, Egocentric-10K: 85.3% (95% CI [78.7, 91.9]) vs published 91.7%; judge alone 90.0%. Not clustered by worker (no worker id on these frames) -- a lower bound on the true width.
• Every judge error in the gold set inflates the statistic. 28 errors across 153 frames: hand count 16 over-counts out of 59 frames that could be over-counted and 0 under-counts out of 129; manipulation 10 out of 33 and 2 out of 120. Build AI's own stored gemini-2.5-flash labels carry the same sign on the same frames, so this is not one model's quirk.
• The published lead over EPIC-KITCHENS-100 does not survive correction, and is not refuted either. Exploratory, not pre-registered: the active-manipulation margin goes from a published +6.62pp to +0.51pp (95% CI [-9.64, 10.67]). The sign flips and the interval still covers the published value, so this fails to resolve the margin rather than refuting it. Using the vendor's own 9,800 stored labels per corpus as the model arm instead of our 200-frame judged pool (DECISIONS.md D089), settling it needs about 184 gold frames per arm at 80% power, against the 63 and 60 that exist; on the 200-frame judged pool the comparison is not answerable by labelling at any power. This card has now given four wrong answers to that one question: 59, on a calculation that scaled the whole standard error with the gold sample (D088); 226, on the small judged pool; 89, which is the boundary case and therefore 50% power, quoted as though it were a resolution (D092); and every one of those was computed with a tuning parameter missing PPI++'s (1 + n/N) divisor, which moved the answer again (D096).
• Domain bias (H5): judge error on manipulation 4.8% (Egocentric, n = 63) vs 8.3% (EPIC-KITCHENS-100, n = 60). Underpowered by design at this n; reported, not concluded.
• Calibration (H7): ECE 0.104 / 0.078; greedy decoding puts ~all confidence in one bin, so this is a weak curve by construction.
• Distillation (H6): DINOv2-small linear probe, teacher fidelity 0.693 vs target 0.90; the 0.80 agreement floor is unreachable at 95% confidence on n = 46 calibration gold. Negative result, shipped as such.

Every statement above is one row of the results config, with the file and key it came from.

Configs

config	rows	one row is
human_labels	213	a HumanLabel: frame_id, rater, pass, rubric_rev, hands_visible, manipulation, edge_case, difficulty, note, labelled_at, seconds_spent
gold_judged	600	a JudgeResponse from the live judge on the three 200-frame gold pools (sample added)
stored_labels	29400	a JudgeResponse reconstructed from Build AI's own published gemini-2.5-flash labels (P0b)
membership	52700	a FrameRef: which pre-registered sample each frame belongs to, with the provenance the source release does (not) ship
results	18	one MEASUREMENT_CARD.json claim: record_type, record_ref, statement

Schemas are the pydantic models in CONTRACTS.md (copied here). Raw result files are under results/. RUBRIC.md is the annotation rubric the human labels follow; PRE-REGISTRATION.md is the protocol, frozen before any data was drawn.

Join on sample, not on corpus. In membership, corpus names the published collection a frame was released in, so every row from the Egocentric-10K-Evaluation release carries egocentric-10k -- including the Ego4D and EPIC-KITCHENS-100 evaluation arms, which ship inside that same release. The sub-corpus lives in sample: E10k-ego / G200-ego are the audited corpus, E10k-ego4d / G200-ego4d are Ego4D, E10k-epic / G200-epic are EPIC-KITCHENS-100. Every analysis in this project keys on sample for that reason. Grouping by corpus will pool three different corpora into one and give you nonsense.

Limitations

One rater, so no inter-rater agreement -- the blind re-label is the substitute, and it ran a median of 2.4 hours after the primary pass rather than the seven days the protocol specified (0 of 34 pairs met it), so it measures consistency within a session and cannot rule out recall (DECISIONS.md D076). The rater had read the audited prompt before writing the rubric. Human gold is n = 153, reduced from the pre-registered 600; every interval is correspondingly wide and every negative result is ambiguous between "no effect" and "cannot see one". No cluster-robust interval exists yet: the evaluation release ships no worker id.

Two checks on the intervals themselves (data/ppi_robustness.json, DECISIONS.md D093). A percentile bootstrap that resamples both pools and re-estimates the PPI++ tuning parameter per replicate agrees with the analytic widths across all twelve arms, ratios 0.85 to 1.07, so the reported intervals can be read at face value. Gold-label noise is the other story: perturbing the gold at the rate the rater's own blind re-label disagreed with itself leaves the hand-visibility results exactly where they are, because the rater reproduced that indicator on all 35 re-labelled frames, but would widen the manipulation margin's interval by 24 to 27%. That widening is an understatement this release reports and does not correct. All of this is in the repo's docs/RED-TEAM.md.

Contributing a second rater

Re-label G200-ego (200 frames, ids in membership) against RUBRIC.md with the repo's make human-labels RATER=<you>, and open a discussion here with the file. Inter-rater agreement is the one number this project cannot produce alone.

License

Apache-2.0 for labels, records and code. Source frames remain under Build AI's release terms and are not included.