mirror of
https://github.com/EvoMap/evolver.git
synced 2026-09-18 21:47:53 +08:00
evolver-publish
26 lines
1.5 KiB
TypeScript
26 lines
1.5 KiB
TypeScript
import type { ValueFactors } from '../schema/problem.js';
|
||
export interface ValueResult {
|
||
score: number;
|
||
factors: ValueFactors;
|
||
}
|
||
/** value = severity × reach × strategicFit × (1+novelty)/(1+costEst). 命名因子, 禁黑盒 (军杰§5.3). */
|
||
export declare function computeValue(f: ValueFactors): ValueResult;
|
||
/** Observation stats for one problem pattern, used to derive its value factors. */
|
||
export interface ProblemValueStats {
|
||
occurrences: number;
|
||
/** false for vague catch-all buckets (e.g. 'general'/'unknown') — they should NOT out-rank specific problems by sheer volume. */
|
||
classified?: boolean;
|
||
/** 0..1 hint from the failure-mode severity (caller maps its taxonomy → a number); default 0.5. */
|
||
severity?: number;
|
||
novelty?: number;
|
||
costEst?: number;
|
||
}
|
||
/** occurrences that earn full reach — a FIXED scale, deliberately NOT normalized to the biggest bucket. */
|
||
export declare const REACH_FULL_AT = 30;
|
||
/**
|
||
* Derive value factors from a problem's observation stats — the tuning learned from observing real agent logs
|
||
* (evolver-v2-observation): a vague high-volume catch-all should not out-trigger specific, actionable problems.
|
||
* So `reach` is on a FIXED occurrence scale (not normalized to the largest bucket, which crushes everything
|
||
* specific), and `strategicFit` is reduced for unclassified problems. Feed the result to {@link computeValue}.
|
||
*/
|
||
export declare function deriveValueFactors(stats: ProblemValueStats): ValueFactors; |