Gate · consequential work, governed by evidence
Know what can proceed—and what still needs review.
The Gate judges whether an action has earned enough evidence to proceed, needs review, or should stop. Strong outcomes ground that judgment. Authority and execution constraints remain independently enforced.
Competence is not authority. A system may be capable of an action without being permitted to take it.
Connect a decision in ~5 lines (Python / TS), or let the Observatory govern a supported system move.
- 95% / 57%
- quality kept · cost paid
- +39%
- energy spike foresight
- 4 domains
- one engine, no tuning
Right now your agent sends the email, executes the code, refunds the customer, and merges the PR. You're trusting it blind. Warm Winter returns one bounded judgment: act, escalate, or abstain. The host system still enforces permission.
A router cuts your bill. A trust gate tells you which calls you can act on.
Routing is the commodity; calibrated, outcome-verified judgment is what lasts. We're not a cheaper gateway, we answer “should we act on this?” Cost savings is just the visible proof.
When an agent acts on its own and gets it wrong, the problem isn't the mistake. It's that it sounded exactly as sure as when it's right.
Agents can merge code, call tools, and move money while sounding equally certain in stronger and weaker cases. What matters is whether a system can distinguish those cases before action, then learn from what actually happened.
Autonomy follows evidence, not a clock.
A system earns bounded autonomy when it can distinguish which individual actions deserve trust. Repeating an action does not establish that distinction.
There is no countdown to autonomy. It expands only where evidence demonstrates reliable discrimination—and can contract when the world changes.
Evidence progression · reversible
- 01ColdNo defensible operating evidence yet.
- 02MeasuredStrong outcomes are resolving against actual actions.
- 03ResolvingThe system is beginning to distinguish stronger cases from weaker ones.
- 04VerifiedCoverage, calibration, resolution, and freshness support bounded autonomy.
- 05RevalidatedRuntime, evidence, state, or regime changes can reduce earned competence again.
The proof
One engine. Four unrelated domains. Free public data.
Proven out-of-sample on four domains with no per-domain tuning, at $0, before a single live user. It validates the method; live use builds the track record. We say it exactly that way on purpose.
- 95% quality · 57% cost
- Compute (RouteLLM). The calibrated gate keeps 95% of the expensive model's quality while paying 57% of the cost, where a random router needs 69%.
- 39% over the floor
- Energy (CAISO). Real-time price-spike foresight, out-of-sample: Brier 0.101 → 0.062 vs the climatology no-skill floor. Replicated +37%.
- 0.0002 calibration error
- Weather (NOAA). Stated probabilities matched observed reality almost exactly, the honest core the whole method rests on.
What this proves: that a stated probability matches observed reality (calibration), one engine across four domains. What it does not prove: causality or external adoption. A live founder-dogfood record now grades real coding-agent operations; that proves the pipe and starts the learning loop, not customer traction. The AI-infra-cost pain above is the market; our backtest proof is inference routing and energy, not storage.
Judgment is earned case by case.
Competence starts with an action family. Judgment happens in context. Warm Winter accumulates evidence across meaningful action classes so experience can compound. Each decision must also account for the runtime, current state, evidence quality, stakes, authority boundary, and operating regime.
Context-aware competence is being introduced in shadow alongside the existing Gate. Production competence remains action-family based until the new system earns promotion.
→ case-specific judgment → observed strong outcome → competence update
Authority is checked independently before execution. Contextual judgment remains under shadow evaluation.
Act on it
Strong outcomes support bounded judgment for this action family. Act still depends on independent authority and valid current constraints.
Escalate
Some signal, but the full evidence bar is unmet. Route to a stronger model, more compute, or a human, and attach strong outcomes to the actual actions.
Don't fake it
No grounded basis to call this yet. The system says so plainly instead of guessing. The part that says “I don't know this” is the most valuable part.
Calibrated and resolving across all four task categories.
Sufficient coverage, calibrated, real resolution, fresh.
Sufficient coverage, calibrated, real resolution, fresh.
Once the inbound has pushed back, a late arrival is foreseen with near-perfect calibration. Replicates across seasons — winter and summer.
Some signal, but miscalibrated past the 0.01 gate — not yet callable.
No resolution above the no-skill floor yet. Escalate, keep measuring.
The flight schedule alone carries no signal. Abstain until the inbound state is known.
The driving feature is leaked by SATCAT; clean element sets are licensed/parked. It won't fake it.
Historical out-of-sample research results from public data. These labels do not describe a live coding-agent cell or authorize production action.
No global trusted-agent label. Each claim has a scope, evidence boundary, and current state.
More experience is not enough.
A system can repeat the same action thousands of times and learn its average success rate without learning which individual cases deserve trust. If it cannot distinguish an easy case from a dangerous one, it has not earned autonomy—even if both usually succeed.
Calibration
When Warm Winter says 80%, do comparable actions succeed about 80% of the time?
Resolution
Can it tell the 95% cases from the 55% cases?
Coverage
Has enough strong outcome evidence accumulated to support the distinction?
Freshness
Is that evidence still relevant to this runtime and operating regime?
Verified requires all four. Accuracy alone is not enough.
Read the method →Competence is not permission.
Competence
Can this system judge this action under conditions like these?
Authority
Is this runtime, worker, or operator allowed to perform it?
Current validity
Are the required state, evidence, constraints, and safety conditions valid now?
An ACT is a judgment inside a declared boundary. It is not a universal permission token. In physical systems, Warm Winter can contribute calibrated judgment. Machine authority and safety remain independently enforced by the physical execution system.
Trust can move backward.
A verified action class does not stay verified forever. Model changes, runtime changes, new tools, stale observations, operating drift, failures, or unfamiliar conditions can reduce the evidence supporting a decision. Warm Winter is designed to expand autonomy where outcomes justify it—and contract it when they no longer do.
From the field · September 2026
Fixed-confidence categories learned average success without proving case-specific discrimination. Shadow cells did not promote on volume alone.
Inspect the engineering finding →The same check, in front of every call you don't fully trust.
Name an action family and bind each call to its actual context and later outcome. The current production Gate groups competence by action family; contextual judgment is being tested in shadow. These are examples of decisions it can be wired to assess:
Cut model cost
Is the cheap model trustworthy enough here, or escalate to the expensive one? Measured: 95% of the quality at 57% of the cost.
Auto-merge / deploy
Hold the agent's change for review or consider an authorized merge. CI can supply a strong result when the exact action and run are bound to its outcome; an open decision does not close automatically.
Tool calls
Does evidence support this tool call, or should it stop for review? Permission, reversibility, and current constraints still govern execution.
RAG grounding
Is the retrieval grounded enough to answer, or should it abstain instead of guessing off the frontier?
Support & outbound
Auto-resolve the ticket or send the message, or route to a human? The reopen, or the reply landing, is the verifier.
…and anything like it
Moderation, data extraction, anomaly flags, and agent handoffs each require their own action and outcome boundary.
You don't integrate five tools, you integrate the gate once, and name the call.
Four domains. One engine. No per-domain tuning.
The engine fits everything, that's exactly why fit can't choose the wedge. Breadth is rationed behind verifiability: undeniably right about one measurable thing, then the next.
Compute / agents
Decides when a cheap model is trustworthy enough to answer, and when to escalate. 95% of the expensive model's quality at 57% of the cost, all four task categories verified.
Energy grid
Foresaw real-time price spikes out-of-sample: Brier 0.101 → 0.062, a 39% gain over the no-skill floor. Replicated +37% on a held-out summer.
Weather
Stated probabilities matched observed reality to a calibration error of 0.0002, the cleanest evidence the calibration math is honest.
Orbital lifetime
A clean orbital-survival result, and an honest ungrounded on the one cell it couldn't ground. It refused to fake a prediction rather than leak the answer.
The asset can't be downloaded.
An outcome record
Decisions are useful for competence only when their outcomes attach to the actual action and carry enough strength to grade it. A larger weak-label record alone does not earn autonomy.
Trust earned in context
Evidence compounds across action families, while a specific judgment depends on current runtime, state, evidence, and regime. Contextual competence remains shadow-first until prospective results support promotion.
Universal by architecture
Domain-blind by design, four domains prove it, but breadth stays rationed behind verifiability. Depth of verified outcomes beats breadth, always.
You keep what you save.
No seats, no per-token markup, no priced-as-if-live tiers. The model follows the value, and it only starts when the loop is real.
Gain-share on verified savings
We charge a share of the spend we verifiably avoid, measured against your own baseline, on outcomes that actually happened. If it doesn't save you money, it doesn't cost you any.
A flat trust-layer subscription
A predictable platform fee for the gate itself, the calibration, the competence frontier, the verified-outcome record, once it runs over your live systems.
Free to start
Self-serve a key and run as much as you want while you prove it out, you only ever pay a share of spend we verifiably save. Early partners lock their rate and get a direct line to the founder.
Wrap your first call in about five minutes. Free to start, no card. Prefer to talk first? Become a design partner.
Start free →Engagements run under our Business Terms and DPA , advisory only; you remain the decision-maker and your systems execute, not Warm Winter.
The product is live. Your outcome record starts now.
The Gate runs through a public API and self-serve keys. Historical four-domain backtests test the method; your action family must earn its own evidence. You can wrap a real call and start a bounded shadow record.
The founder-dogfood record is now compounding across real coding-agent operations. The external customer record is still unearned. We're early on purpose: no inflated claim that internal proof-of-pipe is customer traction.
We say it exactly that way on purpose. Credibility is the product, it compounds the same way the data does.
I'm building the thing I want standing between AI and the decisions that matter.
AI is about to act on our behalf everywhere, in our code, our money, our infrastructure, our lives. The part that unsettles me isn't that it'll be wrong sometimes. It's that it sounds exactly as sure of itself when it's wrong as when it's right.
Warm Winter is my answer: a layer that earns trust the only honest way, by remembering what actually happened every time it made a call, and being willing to say “I don't know this yet.” No omniscience, no hype, no track record I haven't earned. Judgment that improves only when strong outcomes show calibration and case-specific resolution under current conditions.
The interface of AI will be everywhere and free. The judgment underneath it, grounded in reality, honest about its own edges, won't be. If a world where AI acts is coming either way, I want the thing it checks itself against to be honest. That's what I'm building, solo and in the open. I'd rather earn your trust slowly and keep it than win it with a pitch.
, Enrique Vigil, founder · LinkedIn · a direct line
The interface of AI will be free. The judgment underneath it will be scarce.
Warm Winter records outcomes, tests whether its judgments discriminate, and reports where a bounded action has earned evidence for Act, Escalate, or Abstain.
Building something bigger? Become a founding design partner →