The answer is easy. The confidence is the hard part.

A decision model earns its keep when it says “route to billing, 0.82” and the 0.82 is real — so you can escalate on doubt or trade quality against cost. Train on hard labels (the usual way) and the model still picks well but turns overconfident; you’re left patching calibration after the fact.

A known-answer world, dressed in language.

You define a decision as a synthetic world whose answer is computable. certo renders it to natural language, and a small encoder learns to read the state and each option and emit calibrated, typed probabilities. The answer key both trains and grades.

Calibration recovered from plain English.

Measured against the exact posterior (KL, lower is better) on held-out wording. Training on the full distribution lands essentially on the answer; single-label targets sit near “ignore the evidence.”

Define a decision, generate data, ship a classifier.

The interface certo ships (research preview — API stabilizing).

# 1 — describe your decision as a known-answer synthetic world
from certo.synth import EvidenceWorld
world = EvidenceWorld(classes=["billing", "tech", "refunds"], features=12)
data  = world.generate(n=20_000)          # (state, options, EXACT target)

# 2 — train, calibrated by construction (soft/distributional targets)
from certo import train
model = train(data, backbone="ModernBERT-base", target="soft")
model.save("my-router")

# 3 — use it locally: state + runtime options -> calibrated probabilities
from certo import DecisionModel
m = DecisionModel.load("my-router")
r = m.decide(state="card charged twice, refund please",
             options=[{"name":"billing", "description":"charges & payments"},
                      {"name":"tech",    "description":"app problems"}],
             kind="choice")
r.probs        # {"billing": 0.86, "tech": 0.08, "OUT": 0.06}  — calibrated
r.abstain      # True when the top probability is below your threshold

When one forward pass isn’t enough, reason — briefly.

Certo-R1: a Qwen3-4B fine-tuned for concise, verified reasoning — behind the same API.

A calibrated encoder is cheap and trustworthy, but hard, reasoning-heavy decisions (multi-step arithmetic, timezone math, nested exceptions) need generation. Certo-R1 reasons in a sentence or two and returns the chosen option — non-inferior to full verbose reasoning at ~15× fewer tokens and ~8–11× lower latency. The natural architecture is a cascade: the calibrated encoder decides the easy majority; low-confidence cases escalate to Certo-R1.

from certo import ReasoningDecisionModel
m = ReasoningDecisionModel.load("altslate/certo-r1-qwen3-4b", device="cuda")
r = m.decide(
    state="A parcel weighs 2.4 kg. Shipping is $3/kg, any part of a kg rounds up. Charged $9.",
    options=[{"id":"yes", "description":"charged correctly"},
             {"id":"no",  "description":"charged incorrectly"}],
    instructions="Was the customer charged correctly?")
r["top"]        # 'yes'
r["rationale"]  # 'Round 2.4->3 kg; correct = 3x3 = $9; paid $9 -> correct. FINAL ANSWER: A'

One model, three decisions, calibrated probabilities.

The unified decider: one Qwen3-4B backbone, one scoring head — Choice, Score and Noul, probabilities-first.

A single head answers all three decision shapes — pick one of many (Choice), rate on a scale (Score), or yes/no (Noul) — and always returns a probability distribution, not just a label. It can answer in one fast forward pass, or write a short rationale first; both paths return a calibrated distribution, so reasoning improves the answer without giving up trustworthy confidence. Software reads the distribution into the typed response you asked for.

Where it shines: train it on the decision types you’ll actually serve — routing, triage, intent, eligibility, ordinal scoring, verification — and it is accurate and well-calibrated across all three shapes from one endpoint. It’s built for the decisions you can define, not for zero-shot transfer to arbitrary unseen specifications; give it examples of a decision type and it does that type well.

We checked this on a different source (tool-calling + science decisions it never trained on): picking the right action from described options transferred almost perfectly zero-shot (0.99), while a brand-new decision type — “call a tool now, or ask first?” — needed a few examples before it became reliable and calibrated (0.57 → 0.88). Choice over described options generalizes; new rubrics and new decision types are learned, not assumed.

What it’s for — and what it isn’t.