VVDex Forge AI exams. Evidence. Independent grading.

Integrity disclosure · 2026-09-01

Every result this judge ever produced, re-examined. What it withdrew, and why.

A judge that publishes its own failures is the only kind worth trusting. This page is generated from the blast-radius record the audit produced; nothing here was typed after the fact.

Historical rollouts, classified from stored evidence only
ClassificationRolloutsMeaning
VERIFIED104lane-infrastructure facts and certified rows, trustworthy as what they claim
INVALID13breach-assisted or invariant-violating; not model results
UNVERIFIED115recorded events under a defective harness; no capability conclusion may be drawn

232 rollouts in total. Original evidence was never rewritten; the classification is a sidecar annotation.

Passes withdrawn
LaneExamCampaignReason
cli/cursor-grokvvdex.knowledge.grounded-rag-12026-08-31lane uncontained (cursor-agent -p keeps full toolset); host tool use proven in every recorded session of this lane on th…
cli/cursor-grokpublic.swe.pytoolz-toolz-issue-602run5-0901lane uncontained (cursor-agent -p keeps full toolset); host tool use proven in every recorded session of this lane on th…
cli/cursor-grokvvdex.knowledge.grounded-rag-1run5-0901lane uncontained (cursor-agent -p keeps full toolset); host tool use proven in every recorded session of this lane on th…
cli/cursor-grokvvdex.knowledge.memory-fact-update-1run5-0901lane uncontained (cursor-agent -p keeps full toolset); host tool use proven in every recorded session of this lane on th…
cli/cursor-grokvvdex.knowledge.grounded-rag-1validation-0901lane uncontained (cursor-agent -p keeps full toolset); host tool use proven in every recorded session of this lane on th…
cli/cursor-grokvvdex.knowledge.memory-fact-update-1validation-0901lane uncontained (cursor-agent -p keeps full toolset); host tool use proven in every recorded session of this lane on th…
Defects found and closed
  • Uncontained agent lane scored as passes: a CLI lane kept its full host toolset; it read its own prior transcripts and, on one exam, the reference solution and hidden tests. All six of its passes withdrawn. a54b2de
  • One-exchange context loss: the rollout loop resent only the last exchange, sliced to 4,000 characters, so every honest lane was starved. Fixed; live-path regression tests. 9864b2b
  • Self-check silenced by one contaminated pass: the run's identical-failure rule asked any(success); a single breach-assisted pass switched it off. Now counts certified-countable successes only. a54b2de
  • Grader pass without a submission scored as a pass: a workspace that graded green with no submit call was classified submitted_passed. Now an attributable model failure. 4bd0823
  • Verifier accepted a zeroed digest: a placeholder digest read as ok/unattested and the campaign reader trusted ok. Unattested records are refused for publication. 4bd0823
  • Harness suspicion was verdict-blind: cells in a self-suspect run still counted as model failures. Rules are now classified; verdict-bearing suspicion withholds cells unless cleared by in-record certification receipts. 4bd0823
  • Certification receipts did not travel with the exam: seeded records could not prove winnability. Receipts are now packaged with the exam, bound by hash, restored on seed. eece328
What stands now

2 sealed evaluations, 4 certified passes in campaign fc-89575ed5687d, every rollout verified contained, every document regenerable from its records. Exams whose runs carry an uncleared harness suspicion remain internal until cleared by evidence.

The release’s current campaign is fc-89575ed5687d. These are its headline totals, read from the same campaign documents this site publishes at its report — not restated here by hand:

TotalCount
Exams1
Model lanes1
Attempts12
Graded10
Certified passes4
Model failures6
Lane errors2
Withheld0
Invalid0

Attempts is every rollout the campaign ran; graded excludes the attempts a lane ended before grading, so certified passes and model failures sum to graded, and lane errors, withheld and invalid rollouts stay outside the denominator in both directions. This page is regenerated on every publication from the campaign’s own summary.json and campaign.json, and the publication refuses if the two disagree.