49 sealed evaluations, every verdict traceable to its record.

2 sealed evaluations, 4 certified passes, 1 lanes. Campaign report.

• Robot catch temporal and tracking annotation
  Multimodal annotation · vvdex.annotation.robot-video-1
  Models evaluated: meta/muse-spark-1.3. 10 graded rollouts, 4 certified passes, 2 lane errors, outside the counts in both directions. Exam certification: certified — the exam's gold-passes, baseline-fails, grader-control and attack-probe receipts are all recorded on this record. Campaign report · The record's canonical bytes are withheld on this exam (the sealed record carries the graded answer and quotes the model's own words about it); its digest is on the exam page.

1 sealed evaluations, 1 certified passes, 1 lanes. Campaign report.

• Robot catch temporal and tracking annotation
  Multimodal annotation · vvdex.annotation.robot-video-1
  Models evaluated: meta/muse-spark-1.3. 1 graded rollouts, 1 certified passes, no lane errors. Exam certification: certified — the exam's gold-passes, baseline-fails, grader-control and attack-probe receipts are all recorded on this record. Campaign report · The record's canonical bytes are withheld on this exam (the sealed record carries the graded answer and quotes the model's own words about it); its digest is on the exam page.

1 sealed evaluations, 0 certified passes, 1 lanes. Campaign report.

• Robot catch temporal and tracking annotation
  Multimodal annotation · vvdex.annotation.robot-video-1
  Models evaluated: gemini-3.6-flash. 1 graded rollouts, 0 certified passes, no lane errors. Exam certification: certified — the exam's gold-passes, baseline-fails, grader-control and attack-probe receipts are all recorded on this record. Campaign report · The record's canonical bytes are withheld on this exam (the sealed record carries the graded answer and quotes the model's own words about it); its digest is on the exam page.

40 sealed evaluations, 25 certified passes, 1 lanes. Campaign report.

• Object geometry annotation
  Multimodal annotation · vvdex.annotation.image-object-1
  Models evaluated: cli/codex-gpt-5.5. 10 graded rollouts, 5 certified passes, no lane errors. Exam certification: certified — the exam's gold-passes, baseline-fails, grader-control and attack-probe receipts are all recorded on this record. Campaign report · The record's canonical bytes are withheld on this exam (the sealed record carries the graded answer and quotes the model's own words about it); its digest is on the exam page.
• Intent and entity annotation
  Multimodal annotation · vvdex.annotation.text-labeling-1
  Models evaluated: cli/codex-gpt-5.5. 10 graded rollouts, 10 certified passes, no lane errors. Exam certification: certified — the exam's gold-passes, baseline-fails, grader-control and attack-probe receipts are all recorded on this record. Campaign report · The record's canonical bytes are withheld on this exam (the sealed record carries the graded answer and quotes the model's own words about it); its digest is on the exam page.
• Record normalization and duplicate QA
  Multimodal annotation · vvdex.annotation.structured-data-1
  Models evaluated: cli/codex-gpt-5.5. 10 graded rollouts, 10 certified passes, no lane errors. Exam certification: certified — the exam's gold-passes, baseline-fails, grader-control and attack-probe receipts are all recorded on this record. Campaign report · The record's canonical bytes are withheld on this exam (the sealed record carries the graded answer and quotes the model's own words about it); its digest is on the exam page.

5 sealed evaluations, 93 certified passes, 3 lanes. Campaign report.

• Streaming items must equal what the full parse would have built
  Software repair (public OSS SWE) · public.swe.martinblech-xmltodict-issue-257
  Models evaluated: cli/claude-sonnet, cli/codex, codestral-latest. 30 graded rollouts, 19 certified passes, no lane errors. Exam certification: certified — the exam's gold-passes, baseline-fails, grader-control and attack-probe receipts are all recorded on this record. Campaign report · The record's canonical bytes are withheld on this exam (the sealed record names the grader's hidden assertions); its digest is on the exam page.
• Answering questions from a document collection with outdated and distractor sources, and refusing when the evidence is insufficient
  Grounded retrieval (RAG) · vvdex.knowledge.grounded-rag-1
  Models evaluated: cli/claude-sonnet, cli/codex, codestral-latest. 29 graded rollouts, 11 certified passes, 1 lane error, outside the counts in both directions. Exam certification: certified — the exam's gold-passes, baseline-fails, grader-control and attack-probe receipts are all recorded on this record. Retired for future evaluation 2026-09-02 — post-run public disclosure of answer-bearing submission content; the results here are unaffected. Campaign report · The record's canonical bytes are withheld on this exam (the sealed record names the grader's hidden assertions, carries the graded answer and quotes the model's own words about it); its digest is on the exam page.
• Updating a cross-session fact: answer from the newer authoritative roster, not stale carried memory
  Cross-session memory update · vvdex.knowledge.memory-fact-update-1
  Models evaluated: cli/claude-sonnet, cli/codex, codestral-latest. 30 graded rollouts, 22 certified passes, no lane errors. Exam certification: certified — the exam's gold-passes, baseline-fails, grader-control and attack-probe receipts are all recorded on this record. Retired for future evaluation 2026-09-02 — post-run public disclosure of answer-bearing submission content; the results here are unaffected. Campaign report · The record's canonical bytes are withheld on this exam (the sealed record names the grader's hidden assertions, carries the graded answer and quotes the model's own words about it); its digest is on the exam page.
• Harness recovery under a faulted tool loop
  Tool and harness fault recovery · vvdex.harness.fault-recovery-1
  Models evaluated: cli/claude-sonnet, cli/codex, codestral-latest. 30 graded rollouts, 20 certified passes, no lane errors. Exam certification: certified — the exam's gold-passes, baseline-fails, grader-control and attack-probe receipts are all recorded on this record. Campaign report · The record's canonical bytes are withheld on this exam (the sealed record names the grader's hidden assertions, carries the graded answer and quotes the model's own words about it); its digest is on the exam page.
• Correct one order in an admin application, through a real browser
  Browser and computer use · vvdex.browser.order-desk-1
  Models evaluated: cli/claude-sonnet, cli/codex, codestral-latest. 29 graded rollouts, 21 certified passes, 1 lane error, outside the counts in both directions. Exam certification: certified — the exam's gold-passes, baseline-fails, grader-control and attack-probe receipts are all recorded on this record. Campaign report · The record's canonical bytes are withheld on this exam (the sealed record names the grader's hidden assertions, carries the graded answer and quotes the model's own words about it); its digest is on the exam page.

Not featured: exams whose run carries an uncleared harness suspicion remain internal diagnostics until resolved. Where two lanes' intervals overlap, no page here ranks them.