Warm Winter · the calibrated trust layer for AI

Know which AI actions you can act on.

For every model call or agent action, Warm Winter decides whether it's trustworthy enough to act on, and escalates, or skips, the rest. It earns that judgment the only honest way: by scoring its own confidence against what actually happened.

The gate earns trust over weeks of real outcomes, so the record you start today is one you can't backfill later. There's no fast-forward.

Wrap a call in ~5 lines (Python / TS). Running a business instead? For operators →

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 verdict first: act, escalate, or skip.

send emailrefundmerge PRexecute codeupdate CRMlet an agent continue

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 now merge code, call tools, and move money unsupervised, reporting the same confidence whether they nailed it or blew it. The scarce thing isn't more capability; it's trustworthy judgment about which actions are safe to act on, graded against what actually happened.

Autonomy isn't a setting. It's earned, one verified outcome at a time.

Every cell starts cold and defers to you. As real outcomes resolve, the gate earns the right to act on that kind of decision, and only then.

day 01
asks every time
day 07
cells verify
day 14
auto-acts what it earned
Earned autonomy · shadow → enforceillustrative
0%50%100%askactday 1day 4day 7day 10day 14time running · real outcomes accumulatingactions auto-approvedcells verify →

The curve only exists once you start it, there's no fast-forward on earned trust.

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.

Calibrated router · cost vs qualityn = 460 held-out
85%90%95%100%0%25%50%75%100%cost — share of always-expensivequalityrandom router69% cost, same qualityalways-cheapalways-expensive95% quality · 57% cost
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 the live track record. Those start compounding only when real decisions flow through a live product, and none do yet. The AI-infra-cost pain above is the market; our in-hand proof is inference routing and energy, not storage.

The gate

Every decision is graded before it acts.

A competence-frontier cell is (domain × decision-type × context). Each one is labelled by how much the system has actually verified it can call, and that label is the routing rule.

Verified

Act on it

This (domain × decision-type) has been measured against enough real outcomes to call it. The cheap model answers; the action proceeds.

Provisional

Escalate

Some signal, not enough verified depth. Route to a stronger model, more compute, or a human, and record what happens to earn the cell.

Ungrounded

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.

The frontier today · from the backtests
4/ 8 cells verified
4 verified3 provisional1 ungrounded
ComputeVerified
cheap model suffices
RouteLLM MMLU · n=460
95% quality · 57% cost

Calibrated and resolving across all four task categories.

EnergyVerified
RT price spike (>$50)
CAISO SP15 · n=696
+39% skill vs floor

Sufficient coverage, calibrated, real resolution, fresh.

WeatherVerified
precip PoP (6h)
NOAA · 8 stn · n=7,976
0.0002 calibration error

Sufficient coverage, calibrated, real resolution, fresh.

AviationVerified
late arrival, given inbound delay
BTS · 5 hubs · 2024 · n=17,129
+64–66% Brier vs floor

Once the inbound has pushed back, a late arrival is foreseen with near-perfect calibration. Replicates across seasons — winter and summer.

EnergyProvisional
load > DA forecast
CAISO SP15 · n=696
reliability 0.015

Some signal, but miscalibrated past the 0.01 gate — not yet callable.

EnergyProvisional
RT > DA price
CAISO SP15 · n=696
no skill over base rate

No resolution above the no-skill floor yet. Escalate, keep measuring.

AviationProvisional
delay from schedule alone
BTS · 5 hubs · 2024 · n=17,129
no skill over base rate

The flight schedule alone carries no signal. Abstain until the inbound state is known.

SpaceUngrounded
decay within 5y
2010–2020 cohort · n=5,055
refused — altitude leaks

The driving feature is leaked by SATCAT; clean element sets are licensed/parked. It won't fake it.

Computed from free public data, out-of-sample. The provisional and ungrounded cells are the product working as designed — it earns the right to call a cell, it doesn't assume it.

Calibrated humility is the point, not omniscience. There is no global accuracy number, ever.

One gate, many decisions

The same check, in front of every call you don't fully trust.

You wire the gate once and name the decision, (domain × decision-type). It fits anything shaped like a repeated call that later gets a real outcome:

model_route

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

Auto-merge / deploy

Ship the agent's change unattended, or hold it for review? CI is the verifier, it reports the outcome back so the cell learns.

tool:*

Tool calls

Safe to execute, or stop and ask a human? Stakes scale with reversibility, it abstains on an irreversible guess.

rag_answer

RAG grounding

Is the retrieval grounded enough to answer, or should it abstain instead of guessing off the frontier?

support_reply

Support & outbound

Auto-resolve the ticket or send the message, or route to a human? The reopen, or the reply landing, is the verifier.

anything measurable

…and anything like it

Moderation (publish vs. review), data extraction (trust vs. flag), anomaly flags, multi-agent hand-offs. One calibrated verdict, learned per cell.

You don't integrate five tools, you integrate the gate once, and name the call.

Domain-blind by architecture

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 · the wedge
Verified

Compute / agents

LMSYS RouteLLM · n=460 OOS

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 · lead proof
Verified

Energy grid

CAISO OASIS · keyless

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
Verified

Weather

NOAA / NWS via IEM

Stated probabilities matched observed reality to a calibration error of 0.0002, the cleanest evidence the calibration math is honest.

/ space
Verified + honest null

Orbital lifetime

Celestrak SATCAT · keyless

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.

Why it lasts

The asset can't be downloaded.

A record, not a model

The advantage isn't the engine, a model, or a data feed, those commoditize. It's a proprietary, ever-growing record of (situation → decision → outcome), tied to reality. Models get smarter; a multi-year record of what happened when someone acted only deepens.

Trust earned per cell

The competence frontier measures how much the system has verified it can call, per domain, per decision-type, and labels each Verified, Provisional, or Ungrounded. The most valuable thing it builds is the part that says “I don't know this yet.”

Universal by architecture

Domain-blind by design, four domains prove it, but breadth stays rationed behind verifiability. Depth of verified outcomes beats breadth, always.

How we price

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 honest status

The product is live. Your verified track record starts now.

Live today

The gate runs: a public API, self-serve keys, and a ~5-line SDK (Python / TS), calibrated from request #1 by the four-domain backtests. You can wrap a real call right now and get a verdict back.

Still being earned

The live verified-outcome record only compounds as real decisions flow through. We're early on purpose: no inflated claims of a track record that doesn't exist yet. Each resolved outcome is the first of it.

We say it exactly that way on purpose. Credibility is the product, it compounds the same way the data does.

Also validated on human decisions

The operator product, the proof that validated the engine on real business calls.

Pricing, rent, churn across eight industries: narrative in, a recommended move out, every call bound to what actually happened. Same engine, human-facing surface.

See the operator product →
The why

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 gets more trustworthy one verified outcome at a time.

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 trust layer for a world where AI acts

The interface of AI will be free. The judgment underneath it will be scarce.

Warm Winter is the system that earns trust by being verifiably right one decision at a time, and decides when intelligence, human or machine, is trustworthy enough to act on.

Building something bigger? Become a founding design partner →