Judge (`judge()`)

Provider-neutral decision inference distinct from chat completions. Models evaluate supplied state against typed questions and return a Judgement (calibrated signals). Authority stays in JudgementPolicy — models never grant permission. Tool grants stay on AccessPolicy.

Inference outcomes

Kind	Type	Module
Generative chat / tools	`Completion` (= `CompletionResponse`)	`mechaharness.core.outcomes`
Judge / System One	`Judgement`	`mechaharness.inference.judge`
Media / artifacts	`Generation`	`mechaharness.core.outcomes`

Types

Kind	Question	Signal
`noul`	yes/no with true/false criteria	`p_true` ∈ [0, 1]
`choice`	options with descriptions	`selected` + `probabilities`
`score`	min/max + anchors	bounded `value` (+ optional distribution)

A Judgement carries validated answers, per-question errors (missing ≠ zero risk), provenance, and usage (latencyMs, physicalCalls). (JudgeResult is a deprecated alias of Judgement.)

Connection (`APIConnectionConfig`)

Judge HTTP reachability is an injectable APIConnectionConfig (default: SimpleHttpConnectionConfig). Env knobs bind on the connection via from_env() / for_judge() — not on Settings.

Env	Purpose
`MECHA_JUDGE_BASE_URL`	Origin only (e.g. `http://host:8009`)
`MECHA_JUDGE_PATH`	Path joined to base (default `/v1/systemone`)
`MECHA_JUDGE_URL`	Full endpoint override (ignores base+path)
`MECHA_JUDGE_MODEL`	Model id in JSON body
`MECHA_JUDGE_API_KEY`	Optional Bearer token
`MECHA_JUDGE_TIMEOUT_SECONDS`	HTTP timeout

Legacy MECHA_DECIDE_BASE_URL / MECHA_DECIDE_MODEL still apply when the MECHA_JUDGE_* values are unset. Hosts override MechaHarnessConfig.get_judge_connection() for OAuth or other schemes later. Providers (JudgeProvider, InferenceStrategy, …) are Config DI injectables — never fields on Settings.

API

from mechaharness.inference.judge import (
    ChoiceOption,
    ChoiceQuestion,
    FixtureJudgeProvider,
    JudgeRequest,
    hash_state,
    judge,
)
from mechaharness.judgement_policy import JudgementFacts, JudgementPolicy, JudgementThreshold, decide

state = {"ticket": "double charge"}
judgement = await judge(
    JudgeRequest(
        state=state,
        stateHash=hash_state(state),
        questions=[
            ChoiceQuestion(
                id="route",
                instructions="Which team?",
                options=[
                    ChoiceOption(id="billing", description="Payments"),
                    ChoiceOption(id="other", description="Else"),
                ],
            )
        ],
        questionSetVersion="gate-v1",
    ),
    provider=FixtureJudgeProvider(
        {"route": {"selected": "billing", "probabilities": {"billing": 1.0, "other": 0.0}}}
    ),
)
verdict = decide(
    JudgementFacts(action="route"),
    judgement,
    JudgementPolicy(
        version="gate-v1",
        thresholds=[JudgementThreshold(signal_id="route", allow_choices=["billing", "other"])],
    ),
)

export MECHA_JUDGE_BASE_URL=http://127.0.0.1:8009
export MECHA_JUDGE_MODEL=laya
MECHA_STORY_BACKEND=live MECHA_STORY_MODEL=systemone/laya \
  pytest -q tests/stories -k 'refund_verdict or systemone_cassette'

Offline CI replays the System One cassette for the same stories (no skip). See User stories.

JudgementPolicy

mechaharness.judgement_policy.decide(facts, judgement, policy) -> Verdict is pure and fail-closed (unknown signals deny). Persist with mechaharness.decision_log.DecisionLog and replay offline via replay_verdict without tool execution.

Lanes

Hosts expose InferenceEnvironment.active_lane() (reason / judge / media). Call assert_compatible(require_lane="judge") before judge tools; wrong lane raises InferenceEnvironmentError naming the host load hint (e.g. infer load decide-fast while profile filenames catch up).

Judge (judge())