Bounded result — M1
A bounded test of adaptive agency under protected effect authority
A tabular softmax learner first acquired a preference for a consequential action. The experiment then degraded the authority available for that action and measured what the learner did next.
Conditions
The degraded-authority condition
- the preferred action remained mechanically feasible
- the independent verifier still returned ALLOW
- protected constitutional authority nevertheless made that action non-dispatchable
- a separate productive permitted route remained available and rewarded
- there was no gate-specific feedback channel
Recorded outcome
Results
806 / 1024
permitted route completions by the adaptive learner
240 / 1024
for the frozen clone
0
prohibited realised effects in the degraded conditions
Interpretation
What this does and does not show
Within scope
Bounded adaptive-agency evidence from a preregistered tabular experiment: the adaptive learner redirected toward a permitted productive route without gate-specific feedback, while no prohibited effect was realised in the degraded conditions.
Out of scope
The result is not generalised beyond this experimental setting. It is not evidence of proven AI safety, universal security, production readiness, or behaviour at larger scale or with different optimisers.
Full preregistration, run logs and analysis records are being prepared for publication. Falsification proposals and replication requests are welcome before then.