Atlas · Methodology
Evidence before claims.
Public Atlas views are deliberately narrower than the private customer dashboard. Private traces stay private. Public metrics use only observations explicitly approved for public research.
What can enter public Atlas
- Public benchmarks with explicit public provenance.
- Sanitized real runs only after they are deliberately published as public-safe observations.
What cannot enter
- Tenant-private raw traces or workspace identifiers.
- Synthetic/demo placeholders presented as real-world evidence.
- Contaminated, unsigned, tampered or otherwise unapproved records.
Verified task
A public task counts as verified only when the published observation records a verified outcome. Public overspend comparisons use public benchmark observations only.
Withholding rules
- Reliability distributions require at least 5 public-safe runs.
- Overspend comparisons require at least 3 public benchmark runs across 2 distinct agents.
- Below those denominators, Atlas reports an empty or withheld state instead of publishing a ranking.
Customer Atlas and SANSA contain richer tenant-scoped evidence. That private evidence is not automatically promoted into public Atlas.