Harness-Induced Belief Divergence

The harness does not only change what an agent does. With task, environment and model all held fixed, changing only the harness changes what the agent believes about its own run — its estimate of progress, risk, recoverability, the constraints it is under, the failure mode it is in, its uncertainty, its expected future success, the repair cost, and its choice of next action.

This is the mechanism underneath the Binding Constraint Thesis, and it is stranger than a performance delta. Two instances of the same model on the same task can disagree about whether the run is already failing. One believes it is recoverable and keeps spending; the other believes it is lost and stops. The scaffolding is shaping the agent's self-model, not merely its action space.

The practical consequence is that an agent's own progress report is a property of its harness. Any supervisory design that trusts the agent's self-assessment — "should I retry?", "am I stuck?", "is this done?" — inherits whatever bias the harness introduces. It also means a belief-rollout diagnostic is a legitimate harness instrument: you can measure the scaffolding by measuring what the agent comes to believe under it.

Key Claims

• The harness alters an agent's multi-step beliefs across nine dimensions — progress, risk, recoverability, constraints, failure mode, uncertainty, future success, repair cost, next action — with task, environment and model fixed. Evidence: moderate (preprint) (Harness-Induced Belief Divergence)
• A belief-rollout diagnostic can measure harness effects through the agent's self-model rather than through task outcomes alone. Evidence: moderate (preprint) (Harness-Induced Belief Divergence)