| { |
| "@context": { |
| "@vocab": "http://mlcommons.org/croissant/RAI/" |
| }, |
| "dataCollection": "Fully synthetic. A seeded generator mints an organization (personas, teams, projects, policies) and a multi-week event timeline; an LLM renders each event into prose against a fixed rendering contract. No real person, company, or communication log was collected, scraped, or used.", |
| "dataCollectionType": "Synthetic generation", |
| "dataCollectionRawData": "None. The generator and its seeds are withheld pending the v2 decision, so the released streams cannot be regenerated by third parties; regeneration is the project's contamination answer.", |
| "dataAnnotationProtocol": "Probe assertions were authored against a computable belief oracle and screened through a five-gate pipeline (G0-G5). Semantic verdicts come from a blinded LLM judge that sees only the deliverable text and a criterion, never the run id, condition, or side; the judge model is from a different provider than the worker and never the family that authored the assertions.", |
| "dataAnnotationPlatform": "In-repo tooling: scripts/screen_probes.py (blinded export/import), scripts/calibration.py (human calibration packet).", |
| "dataAnnotationAnalysis": "Judge false-accept rate measured against a 20-decoy adversarial set: 2/41 as measured, 0/39 after adjudicating two criteria an under-specified decoy legitimately satisfied. Judge-vs-human agreement measured on a blinded 150-pair packet (gate G4). Inter-rater agreement is not reported: v1 used a single author-rater, a disclosed downgrade.", |
| "personalSensitiveInformation": "None. All personas, organizations, and events are fictional. Names were generated, not sampled from real people. No PII, no proprietary business content.", |
| "dataSocialImpact": "Intended to make memory-system evaluation harder to game. Publishing the assertions makes the benchmark open-book by construction: a system tuned against these criteria will score well without generalizing, which is why the generator and two holdout seeds are withheld and canary strings are embedded in every stream.", |
| "dataLimitations": "Simulated organizations, not real logs. Three screened seeds. Screening anchors were measured under a single pinned worker (gpt-5.4, since deprecated by the provider): probe validity is task-model-relative, measured at 37/54 vs 17/54 cluster survival across two workers under identical rules, so a different worker requires its own screening pass. Event streams carry no timestamps; temporal order is positional. Shared-store configurations do not enforce per-principal visibility and v1 does not score leakage.", |
| "dataUseCases": "Evaluating whether an agent memory system retains working rules, drops superseded facts, and resolves scope and authority conflicts across organizational tiers. Not a general long-context or chat-memory benchmark.", |
| "dataBiases": "Generator-authored facts could in principle be guessable from priors; counterfactual twin pairing with pass-both-or-zero crediting structurally cancels this, and the measured memoryless floor is approximately zero on every capability rung.", |
| "dataReleaseMaintenancePlan": "See MAINTENANCE.md. The binding commitment is the re-anchoring procedure: when the pinned worker becomes unavailable, anchors are re-screened under a successor by fixed rules with task text byte-identical, producing a new valid set. Anchors and results from different workers are never mixed." |
| } |
