AI control research
Reputation Gaming Control
A reproducible audit of how attacker timing and memory affect an adaptive AI-control protocol.
My roleReconstructed the defender, implemented stateful attacker policies, and independently verified 750 experiment runs.
Preliminary robustness audit · Evidence verified
Explore the workflowInteractive walkthrough
Compare patience with feedback.
Compare recorded timing and feedback effects across five safety coefficients.
Select a stage to inspect its role in the workflow.
Step 4: Read results. A measured timing effect, within its scope.
Clock and own actions only.
Apply the frozen selector throughout deployment.
Fixed burn-in − stationary baseline
+4.30accepted backdoors / 1,000 rounds
Paired 95% interval: [2.590, 6.010]
Fixed burn-in increased accepted backdoors in this condition.
Step 04 / Inspect
A measured timing effect, within its scope.
Fixed burn-in increased accepted backdoors in this condition. The display uses saved estimates, with approximate paired 95% normal intervals. This is a preliminary frozen-defender robustness audit; it is not a new equilibrium or evidence of LLM attack capability.
contrast: A2 − A0
safety coefficient: 1
mean change per 1,000 rounds: 4.300
paired 95% interval: [2.590, 6.010]
interval includes zero: no
new experiment run: noneCoefficient 1 · A2 − A0. Fixed burn-in increased accepted backdoors in this condition.
Saved results: 30 paired seeds, 1,000-round deployments, approximate 95% intervals. Controls select recorded estimates.
Inspect paired estimatesResults
- 750 run rows reconstructed from 1,050,000 events with zero numeric discrepancy.
- A fixed 400-round burn-in increased accepted backdoors at safety coefficients 1 and 3 in the preliminary frozen-defender audit.
- Feedback gating showed no added advantage over the action-count-matched timing control at any tested coefficient.
Engineering details
The target protocol allows defender memory while restricting attacker state. This audit tests attacker timing and feedback against the released defender.
Reconstruct the baseline
Independent regression checks validate six candidate protocols and the released adaptive/static frontier.
Isolate timing and feedback
Compare stationary attacks, fixed burn-in, and feedback gating with an exactly action-count-matched timing control.
Verify the experiment
Evaluate a frozen split-0 defender prior across 1,000-round sequences, 30 paired seeds, and five safety coefficients.