Numbers
Every figure on this site, with the file it was counted from and the date it was true. If a number is a projection rather than a measurement, it says so.
Most vendors ask you to trust their quality claims. We would rather you check ours.
Project Vista
| Figure | What it measures | Basis | Source |
|---|---|---|---|
| 117 hours | Hours of long-form source footageA property of the source dataset, not of throughput. | Measured2026-08-28 | docs/lvbench-cost-report-2026-07-23.mdline 68 |
| 1,549 questions | Questions in the repaired benchmark | Measured2026-08-28 | data/lvbench/delivery-final/summary.jsonn_delivered |
| 103 videos | Distinct source videos covered | Measured2026-08-28 | docs/lvbench-cost-report-2026-07-23.mdline 68 |
| 1,549 of 1,549 | Questions accounted for by name at delivery1,530 shipped, 19 delivered as documented exclusions with reasons, 0 withheld, 0 incomplete. | Measured2026-08-28 | data/lvbench/delivery-final/_COMPLETE.jsondisposition_counts |
| 1,546 questions | Questions delivered with a corrected evidence window | Measured2026-08-28 | data/lvbench/delivery-final/summary.jsonn_with_evidence_window |
| 655 of 696 | Questions no model could answer from the text aloneMeasured by a three-model blind panel given the question and options but no video. Establishes that shipped questions require watching the footage; it says nothing about whether a repaired question is better than the original. | Measured2026-08-28 | docs/projects/vista-system-policy.mdlines 2352-2358 |
| 33 people | Reviewers who filed independent determinations | Measured2026-08-24 | docs/VISTA-FACT-SHEET-2026-08-24.mdsection 6 |
| 1,241 questions | Questions carrying a formal adjudication ruling | Measured2026-08-28 | data/lvbench/delivery-final/summary.jsonn_adjudicated |
| 1,512 questions | Questions shipped with both reviewers' raw, unreconciled labelsShipped so the client can compute inter-reviewer agreement from the data rather than taking our word for it. | Measured2026-08-28 | data/lvbench/delivery-final/summary.jsonn_with_raw_reviewer_labels |
| 120 audits | Senior spot-audits run over reviewer pairs that agreedAuditing agreements, not just disagreements, is what catches two reviewers making the same mistake. | Measured2026-08-28 | docs/VISTA-PAYMENT-FINAL-2026-08-28.mdsection 5 |
| 1,191 questions | Questions given full two-person independent blind review | Measured2026-08-28 | data/lvbench/delivery-final/summary.jsonverification_tiers.full_two_person |
| 388 questions | Questions materially corrected or withdrawn369 repaired and 19 withdrawn, out of 1,549. The remaining 1,161 were verified and kept unchanged. | Measured2026-08-28 | data/lvbench/delivery-final/summary.jsondecisions |
| 250 answer keys | Answer keys found to be wrong and corrected | Measured2026-08-28 | data/lvbench/delivery-final/summary.jsonn_correct_mc_key_changed |
| 588 defects | Spelling and grammar defects found across 380 questionsCorrections that do not change the task. Reported to the client under a new defect code rather than fixed silently. | Measured2026-08-28 | docs/projects/vista-lvbench-sow-2026-07-29.mdlines 225-234 |
Project Skerry
| Figure | What it measures | Basis | Source |
|---|---|---|---|
| 70 clips | Unique 4K clips annotated, across four locationsRoughly 5.6 hours of unique footage from Bristol, Cancun, Jakarta and Oregon. | Measured2026-07-23 | docs/skerry-cost-report-2026-07-23.mdlines 92-94 |
| 760 objects | Objects annotated in a single five-second harbour clipThe dense end of the range. A sparse bridge scene carries around 160. | Measured2026-07-23 | docs/skerry-cost-report-2026-07-23.mdlines 191-194 |
| 0% | Frontier-model accuracy on fine-grained vessel typing and countingFrom the earlier proof-of-concept round (n=12). The same models scored 100% on scene-level questions, which is what makes the contrast informative. | Measured2026-07-14 | docs/vqa-saturation-handoff-2026-07-14.mdlines 93-98 |
| 44.1% | Questions built from our annotations that frontier models answer correctly95% confidence interval 32.2-56.7%, n=59. Per-model, the best score was 70%. A low number is the point - it means the questions are hard. | Measured2026-07-14 | docs/vqa-saturation-handoff-2026-07-14.mdlines 194-198 |
Project Mosaic
| Figure | What it measures | Basis | Source |
|---|---|---|---|
| 8 capture types | Distinct capture types the taxonomy separates | Measured2026-07-23 | docs/projects/diverse-pilot-strata.mdlines 23-46 |
| 16 content strata | Content strata in the diversity taxonomyThe taxonomy is versioned in code, not prose. The client's floor was twelve. | Measured2026-07-23 | docs/projects/diverse-pilot-strata.mdlines 23-46 |
Source paths refer to files in our internal project repositories. We will walk any of them through with a prospective client under NDA. Figures we hold back (client commercial terms, our own costs, and anything identifying an individual reviewer) are marked internal at the source and cannot be rendered by this site.