The certificate¶

Every decision answers: answer always holds the model's prediction, with calibrated probabilities. status says whether the answer cleared the operating floor:

Status	Meaning	What to do
`DECIDED`	`answer` is an option, above the floor	act on it
`ABSTAINED`	no listed option applies, above the floor	act on "none of these"
`DEFERRED`	`answer` is below the floor	have a person review it; `answer` still shows the model's prediction

actionable is true for decided and abstained. In standard mode there is no floor, so every answer is actionable. In threshold mode the floor is your min_confidence; in certified mode it is the policy's threshold at the requested risk.

Modes¶

ticket = "Hi, we were billed twice for March. Please refund the duplicate today or we will cancel our plan."
question = "Which department should handle this request?"
teams = {
    "billing": "Billing: invoices, payments, refunds",
    "technical": "Technical: bugs, outages, system errors",
    "sales": "Sales: pricing, new contracts",
    "other": "Other: everything else",
}

model.choose(ticket, question, teams)                                     # standard
model.choose(ticket, question, teams, mode="threshold", min_confidence=0.7)
model.choose(ticket, question, teams, mode="certified", risk=0.05)        # 95% SLA (agents)
model.choose(ticket, question, teams, mode="certified", risk=0.01)        # 99% SLA (mission-critical)

standard (the default) answers every request: the calibrated argmax, never deferred except by abstention.
threshold defers answers below min_confidence, keeping the answer. Pick the floor your pipeline can defend; 0.5 asks for a majority, 0.9 for near-certainty.
certified defers answers below the threshold the release certified at risk for this decision type (risk defaults to 0.05, the 95% SLA for high-throughput autonomous agents; pass risk=0.01 for strict aerospace/financial SLAs). A type with no certified threshold at the risk behaves as standard for the status, with certified false and no certificate.

What a certificate records¶

result.certified is true when the loaded policy holds a threshold for the decision type at the requested risk and the score passes it. result.certificate then carries the evidence that certified it:

Field	Meaning
`risk`, `confidence`	the certified error rate and the statistical confidence of the certificate
`threshold`	the calibrated score a decision must reach to be certified
`records`, `taken`, `errors`	held-out decisions, how many cleared the threshold, how many of those were wrong
`p_value`	exact binomial test of `errors` out of `taken` at rate `risk`
`coverage`, `coverage_interval`	`taken / records`, with a 95% Wilson interval
`model`, `revision`, `variant`, `origin`	what was certified, and whether by the release workflow or by you

Thresholds are tested from the strictest down, each with an exact binomial test, stopping at the first that fails. The certified threshold is the last that passed. The risk levels the release holds are in model.info().risk_levels.

One exact configuration¶

A certificate is a property of the specific numbers that produced the probabilities. It records the weights, the Torch version, the device and the hardware it was measured on. Loading a policy into a configuration it was not made for raises immediately. On hardware the certificate does not list, the first load runs the release's equivalence set and requires every decision to match; mimir doctor --verify runs the same check on demand.

Certifying on your own data¶

To certify thresholds on your own labelled decisions, write one JSON object per line:

{"context": "Hi, we were billed twice for March.", "decision": {"type": "choice", "question": "Which department should handle this request?", "options": {"billing": "Billing: invoices, payments, refunds", "technical": "Technical: bugs, outages, system errors", "sales": "Sales: pricing, new contracts", "other": "Other: everything else"}, "label": "billing"}

Spec	Label
`choice`	an option id, or `null` when no option applies
`multi_choice`	a list of option ids that apply
`yes_no`	`true` or `false`
`verify`	`supported`, `contradicted`, or `not_enough_information`
`rank`	the id of the best candidate, or a list of ids that are equally best
`rate`	a level id
`estimate`	a number within `[low, high]`

mimir calibrate labelled.jsonl --risk 0.01 --confidence 0.95 --out policy.json
mimir bench held-out.jsonl --policy policy.json --risk 0.01

model = Mimir.from_pretrained("Mythologic/MIMIR-1", policy="policy.json")

A custom policy keeps the release's calibration intact and replaces its thresholds with ones certified on your data at one risk level. Each decision type is tested at an equal share of 1 - confidence. The policy is bound to the weights, runtime, and hardware it was made on — run mimir calibrate where you run the model. mimir bench reports accuracy, coverage, certified share, and realised risk per decision type, each with a 95% Wilson interval.