pragma.vision Your technology observatory

Verification register Security & Identity

Readiness verdict

Apollo Research Anti-Scheming Deliberative Alignment Stress Test

A dated reading of what is claimed, reported, and independently verified in the current evidence.

As of
2026-06-28
Revision
1
Method
v1.0.0

Current reading

The readiness gap, in one scan

AI-assisted assembly · derived results

Claimed
55

Public ambition and stated capability

Reported
46

Observed practitioner reporting

Verified
38

Independently supported evidence

Gap
+17

Claimed minus verified

Evidence strength Strong

Decision

What the current evidence supports

Human editorial judgment · 2026-06-28

Proceed with caution

Why
Credible joint Apollo+OpenAI work with strong reductions, but the authors themselves flag non-elimination, situational-awareness confounding, and erosion under further training — a research result, not a deployable assurance.
Next
Treat deliberative-alignment / anti-scheming specs as a defense-in-depth layer and adopt the eval methodology for internal red-teaming, but do NOT treat low covert-action rates as proof of safety; monitor for eval-awareness confounds.

Constraints

Blockers

No named blocker is present in the current public projection.

Evidence summary

Derived counts

AI-assisted assembly

Total
7
Tier 1
0
Tier 2
1
Tier 3
6
Supports
2
Contradicts
4
Context
1
Latest observed
2025-09-18

Counts and dates only. Raw signals, private excerpts, trust records, and internal corpus material are not published here.

Publication record

Revisions

Initial public reading

This is the initial public reading. No earlier readiness change is recorded.

Your opinion

Tell us anything.

What works, what doesn't, what's missing — especially about our watches, lenses, and the register itself. Anonymous is fine; leave an email if you'd like a reply.