← Warm Winter

Research

Act, Escalate, or Abstain

Deriving an AI trust gate from one objective

Read the full paper (PDF)Try the gate →

Agents can take actions faster than anyone can verify them, and the reflexive fix — asking another model whether the first looks reasonable — is exactly as capable of being confidently wrong as the first. What’s missing isn’t another opinion; it’s a calibratedone that knows when it doesn’t know. Warm Winter is that gate: it returns act, escalate, or abstain. This is where those three verdicts come from.

Proposed actionExpected free energypragmatic value + epistemic valueACTconfident,affordableESCALATEinformationcheap vs. riskABSTAINuncertaintyirreducible
The gate as one decision. A proposed action is scored by a single expected-free-energy objective; act / escalate / abstain are the regions of its minimizer.

Five results, one idea

One objective, three verdicts

Act, escalate, and abstain are an optimal policy under a single expected-free-energy objective — and the minimizer reproduces the gate's nine-cell decision table exactly, from structure rather than tuning.

The cost of computing

Escalation priced as an energy↔information trade on real model-routing data: the return on compute, the mutual information each escalation extracts, and the wasted compute it honestly quantifies.

Knowing which kind of uncertain

An exact epistemic vs. aleatoric split makes abstention principled. The usual count threshold is base-rate blind — a balanced cell at 30 samples is 3.1× more uncertain (by posterior variance) than a near-deterministic one at 25.

Causal, not correlational

Off-policy evaluation on the gate's own randomized exploration recovers an unbiased value of acting where the naive estimate is off by 22 points — doubly-robust, at 42% lower variance. That makes “we prevented X” a cause.

Honest under drift

A distribution-free conformal coverage guarantee (90.9% at a 90% target, despite a wrong model) that an adaptive variant holds (89.9%) where a frozen threshold breaks (81.7%) — because agents adapt to being gated.

Why this is the point

None of these methods is exotic on its own — that’s deliberate. The bet is that calibrated decision-making with honest abstention, verified against real outcomes, is a different and more durable thing to be good at than out-modeling the frontier labs. And every result above states what it does not claim — the thermodynamic limit it doesn’t reach, the coverage that’s marginal not per-decision, the counterfactuals it won’t estimate without overlap. For a product whose entire job is catching ungrounded overconfidence, holding its own claims to that standard is the whole credibility argument.

The gate ships as a tiny SDK — pip install warmwinter / npm install warmwinter — with no LLM in the decision path, so running it is essentially free.