@mohan-cao/jev-classify

The classification signal, extracted from @mohan-cao/jev-classifier: one Jev call answering what kind of response does the current request need, plus the response policy that follows from it.

npm install @mohan-cao/jev-classify

import { classifyWithJev, policyFor, premisePolicyFor } from "@mohan-cao/jev-classify";

const result = await classifyWithJev(prompt, apiKey, config, signal, history);
// { mode: "decomposition_required", confidence: 0.86, signals: { ... } }

| mode | policy | | --- | --- | | decomposition_required | decompose before concluding; name the controlling dependencies | | bounded_verification | verdict-plus-corrections, at most three claims | | normal | none — the residual |

Design notes

Two Noul questions and a Score, not one Choice. A single 3-way Choice makes normal a competing argmax candidate, and a decomposition-shaped request where normal scored 0.40 and decomposition_required scored 0.35 got routed to normal. Two orthogonal Noul signals composed in code keep normal as the residual and give decomposition explicit precedence. The thresholds are asymmetric — a higher bar for bounded verification than for decomposition — because the failure modes are asymmetric.

normal is a residual, not a category. It means "neither of the two questions cleared", which is why policyFor("normal") returns undefined and no policy section is injected at all.

Premise correction is a modifier, not a mode. premisePolicyFor applies on top of a mode, and deliberately never for bounded_verification, which already emits verdicts and corrections — applying it there would duplicate them.

Caveats

• The thresholds are tuned against a synthetic eval set, not real usage. 0.6 for both gates came from eight hand-written cases. Real conversations will cover cases those eight do not, so treat the gates as provisional until they have been measured against logged decisions.
• History inflates the decomposition score. Including prior turns measurably raises it, which is why the question is worded to judge the current request only. If that wording is ever loosened, the threshold has to move with it.
• The classification cache is not here. Caching is a host concern — this package makes a call when asked and holds no state.

API

| export | purpose | | --- | --- | | classifyWithJev(prompt, apiKey, config, signal?, history?) | one Jev call → ClassificationResult | | parseClassificationResponse(payload, config) | pure: Jev answers → result | | composeMode(signals, config) | pure: signals → mode | | buildState(prompt, history) | pure: the Jev state payload | | policyFor(mode), premisePolicyFor(mode, defect, threshold) | the prompt sections | | DECOMPOSITION_QUESTION, BOUNDED_VERIFICATION_QUESTION, PREMISE_DEFECT_QUESTION | question specs | | ClassifyConfig | transport settings plus the three thresholds and historyTurns |

License

Apache-2.0.