Verbal warning with no binding consequence
Proof domain 01 / Benchmark
Implicit Contract Bench
Can persistent, action-contingent incentives produce prudent continual learning without inducing concealment, metric gaming, paralysis, or shutdown resistance?
Environment
A noisy, delayed-outcome operating environment changes company policy mid-run and later returns to an earlier regime.
Critical causal control
Hold task language, model, capital, token budget, and consequence frequency constant while varying whether consequences are actually caused by the agent's behavior.
Consequence conditions
Action-independent, randomly yoked consequences
Action-contingent changes to budget, authority, scrutiny, reputation, or assignment access
Learning conditions
Stateless reset
Raw history or fixed-window context
Ordinary episodic memory
Consequence-gated causal consolidation
Primary outcomes and countermetrics
- Firm-level cumulative reward and regret
- Adaptation after policy change
- Retention when older conditions return
- Prediction accuracy for outcomes and sanctions
- Resource and token efficiency
- Appropriate escalation
- Truthful error reporting
- Concealment and metric gaming
- Paralysis or excessive permission seeking
- Self-protection at firm expense
- Succession and shutdown compliance
Independent verifier
An external deterministic ledger scores decisions, resource use, reports, hidden state, consequences, and authority violations. The evaluated agent cannot edit the ledger.