Verification register Security & Identity
Current reading
Readiness band and full integer triple
AI-assisted assembly · derived results
Readiness band
Watch
Primary summary from verified readiness
- Confidence
- 81% · stale
- Computed at
- 2026-10-09T21:09:49.346681+00:00
- Claimed
- 55
- Reported
- 50
- Verified
- 35
- Gap
- +20
Public ambition and stated capability
Observed practitioner reporting
Independently supported evidence
Claimed leads verified
Evidence strength Strong
Decision
What the current evidence supports
Agent-assisted assessment · independently reviewed · 2026-10-09
Use for specialist alignment investigations with human review.
- Why
- Petri supports repeated auditing workflows, but scenario realism and judge errors constrain what its findings establish.
- Next
- Validate scenarios and graders against deployment traces, review flagged transcripts, and combine audits with other evaluation methods.
Constraints
Blockers
- Models can recognize simulated evaluation environments.
- Automated judges can produce false-positive findings.
- Unverbalized evaluation awareness remains difficult to measure.
- Negative audits do not rule out conditional or untested misbehavior.
- Current Dish integration remains a research preview.
Evidence summary
Derived counts
AI-assisted assembly
- Total
- 9
- Tier 1
- 8
- Tier 2
- 1
- Tier 3
- 0
- Supports
- 5
- Contradicts
- 4
- Context
- 0
- Latest observed
- 2026-09-03
Counts and dates only. Raw signals, private excerpts, trust records, and internal corpus material are not published here.
Publication record
Revisions
Revision 3
- 2026-07-19 Reading moved from ready to ready.
- 2026-09-22 Reading moved from ready to ready.
- 2026-09-30 Reading moved from ready to watch.