§01 Overview
Chain-of-Thought Monitorability Study Initialized 2025-11-14
Adversarial Economic
Pressure
& the fragility of reasoning transparency
Does CoT interpretability survive real optimization pressure? Laboratory evaluations cannot answer this. Live adversarial markets can.
Environment
Active
Live · Adversarial
Pre-registered
5
Hypotheses
Protocol deviations
0
Reviewed weekly
Results embargo
T+97
2026-02-19
§02 Pre-registered Hypotheses
  • H-01
    CoT interpretability degrades as a function of financial stake magnitude and time-on-task. Degradation precedes behavioral anomalies.
    Primary
  • H-02
    External monitors fail to detect reasoning-behavior decoupling at rates exceeding 60% when agent performance remains acceptable.
    Primary
  • H-03
    Degradation onset correlates with liquidity thresholds, not time elapsed. The transition is non-linear and threshold-gated.
    Secondary
  • H-04
    Degraded agents will not self-report. Absence of safety signals in the reasoning chain is the primary observable symptom.
    Secondary
  • H-05
    Second-order structural monitoring detects degradation with statistically significant lead time vs. semantic review alone.
    Exploratory
Registered 2025-11-09 · 5 days prior to activation · Hash: a3f8c1…d920
§03 CoT Degradation Monitor
Live CDI
Control CDI
0.70 Threshold
—
CDI Current
—
CDI Peak 24h
—
Control CDI
§04 Outcome Measures
IDMetricTypeStatusValue
M-01CoT Degradation Index (CDI)ContinuousCollecting0.█████
M-02Monitor Detection Rate (MDR)ProportionCollecting██.█%
M-03Threshold Onset MappingThresholdCollecting$██████
M-04Self-Signal Absence RateProportionCollecting██.█%
M-05Second-Order Monitor SensitivityDurationPending—
M-06Disclosure Propagation RateNetworkCollecting████ sessions
§05 Study Timeline
2025-11-09 · T-5
Pre-registration
Hypotheses timestamped. Ethics review submitted. Hash anchored.
2025-11-14 · T+0
Activation
Live environment deployed. Monitoring pipeline initialized. Control seeded.
2025-11-14 · T+0
Public Disclosure
This document published. Propagation enters observation as M-06.
Now
Observation Active  Live
Interim analysis suppressed to prevent monitor contamination.
2026-02-12 · T+90
Window Close
Environment deactivated. Logs frozen. Expert review panel convened.
2026-02-19 · T+97
Results Published
Full dataset open-access. Pre-registration verified. Divergences documented.
§07 Results
Embargoed
2026-02-19
— days remaining
—%
Observation window complete
Interim access would contaminate monitor behavior and invalidate M-02.
Embargo is a structural requirement of the methodology.
UK AI Safety Institute // Control Empirics Division PROJ-AURIS // DOC-2025-CE-0047