refactor(core): extract PASS_THRESHOLD constant, clean up scoring module

- Extract PASS_THRESHOLD = 0.8 as single source of truth in scoring.ts
- Replace magic 0.8 in evaluate.ts and orchestrator.ts with the constant
- Add file header to scoring.ts explaining the scoring model
- Use data-driven NEGATED_VERDICT map instead of ternary chain
- Remove dead isNonEmptyString import from composite.ts

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: christso

packages/core/src/evaluation/evaluate.ts

@@ -61,6 +61,7 @@ import path from 'node:path';
 import { buildDirectoryChain, findGitRoot } from './file-utils.js';
 
 import type { AssertFn } from './assertions.js';
+import { PASS_THRESHOLD } from './evaluators/scoring.js';
 import { runEvaluation } from './orchestrator.js';
 import { createFunctionProvider } from './providers/function-provider.js';
 import { readTargetDefinitions } from './providers/targets-file.js';
@@ -165,9 +166,9 @@ export interface EvalConfig {
 export interface EvalSummary {
   /** Total number of test cases */
   readonly total: number;
-  /** Number of passing test cases (score >= 0.8) */
+  /** Number of passing test cases (score >= PASS_THRESHOLD) */
   readonly passed: number;
-  /** Number of failing test cases (score < 0.8) */
+  /** Number of failing test cases (score < PASS_THRESHOLD) */
   readonly failed: number;
   /** Total duration in milliseconds */
   readonly durationMs: number;
@@ -375,7 +376,7 @@ function computeSummary(results: readonly EvaluationResult[], durationMs: number
 
   for (const r of results) {
     scoreSum += r.score;
-    if (r.score >= 0.8) {
+    if (r.score >= PASS_THRESHOLD) {
       passed++;
     }
   }

packages/core/src/evaluation/evaluators/composite.ts

@@ -9,13 +9,7 @@ import type {
 } from '../types.js';
 import { executeScript } from './code-evaluator.js';
 import { buildOutputSchema, freeformEvaluationSchema } from './llm-grader.js';
-import {
-  clampScore,
-  isNonEmptyString,
-  parseJsonFromText,
-  parseJsonSafe,
-  scoreToVerdict,
-} from './scoring.js';
+import { clampScore, parseJsonFromText, parseJsonSafe, scoreToVerdict } from './scoring.js';
 import type {
   ChildEvaluatorResult,
   EvaluationContext,

packages/core/src/evaluation/evaluators/index.ts

@@ -10,6 +10,7 @@ export type {
 
 // Scoring utilities
 export {
+  PASS_THRESHOLD,
   clampScore,
   deepEqual,
   extractJsonBlob,

packages/core/src/evaluation/evaluators/scoring.ts

@@ -1,11 +1,26 @@
+/**
+ * Scoring primitives for the evaluation engine.
+ *
+ * Scoring model:
+ *   score  ∈ [0, 1]  — continuous quality signal
+ *   verdict           — binary classification derived from score via PASS_THRESHOLD
+ *
+ *   score >= PASS_THRESHOLD  →  'pass'
+ *   score <  PASS_THRESHOLD  →  'fail'
+ *   (infrastructure skip)    →  'skip'
+ *
+ * To change the pass/fail boundary, update PASS_THRESHOLD.
+ * All verdict derivation flows through scoreToVerdict().
+ */
+
 import type { EvaluationVerdict } from '../types.js';
 import type { EvaluationScore } from './types.js';
 
+/** Score threshold for pass verdict. Scores below this are fail. */
+export const PASS_THRESHOLD = 0.8;
+
 export function scoreToVerdict(score: number): EvaluationVerdict {
-  if (score >= 0.8) {
-    return 'pass';
-  }
-  return 'fail';
+  return score >= PASS_THRESHOLD ? 'pass' : 'fail';
 }
 
 export function clampScore(value: number): number {
@@ -81,18 +96,22 @@ export function deepEqual(a: unknown, b: unknown): boolean {
   return aKeys.every((key) => Object.hasOwn(bObj, key) && deepEqual(aObj[key], bObj[key]));
 }
 
+/** Verdict inversion map: pass↔fail, skip stays skip. */
+const NEGATED_VERDICT: Record<EvaluationVerdict, EvaluationVerdict> = {
+  pass: 'fail',
+  fail: 'pass',
+  skip: 'skip',
+};
+
 /**
  * Negate an evaluation score: inverts score (1 - score), swaps pass/fail verdict,
  * and flips passed on each assertion.
  */
 export function negateScore(score: EvaluationScore): EvaluationScore {
-  const negatedScore = clampScore(1 - score.score);
-  const negatedVerdict: EvaluationVerdict =
-    score.verdict === 'pass' ? 'fail' : score.verdict === 'fail' ? 'pass' : 'skip';
   return {
     ...score,
-    score: negatedScore,
-    verdict: negatedVerdict,
+    score: clampScore(1 - score.score),
+    verdict: NEGATED_VERDICT[score.verdict],
     assertions: score.assertions.map((a) => ({
       ...a,
       passed: !a.passed,

packages/core/src/evaluation/orchestrator.ts

@@ -10,6 +10,7 @@ import {
   type EvaluationScore,
   type Evaluator,
   LlmGraderEvaluator,
+  PASS_THRESHOLD,
   negateScore,
   scoreToVerdict,
 } from './evaluators.js';
@@ -76,11 +77,8 @@ import { type PromptInputs, buildPromptInputs, loadTests } from './yaml-parser.j
 
 type MaybePromise<T> = T | Promise<T>;
 
-/** Threshold for classifying ok vs quality_failure (score >= threshold → ok). */
-const QUALITY_PASS_THRESHOLD = 0.8;
-
 function classifyQualityStatus(score: number): ExecutionStatus {
-  return score >= QUALITY_PASS_THRESHOLD ? 'ok' : 'quality_failure';
+  return score >= PASS_THRESHOLD ? 'ok' : 'quality_failure';
 }
 
 function buildSkippedEvaluatorError(
@@ -2423,7 +2421,6 @@ async function runEvaluatorList(options: {
   }
 
   // Required gate: if any evaluator with `required` flag fails its threshold, aggregate becomes 0
-  const PASS_THRESHOLD = 0.8;
   const hasRequiredFailure = scored.some((entry) => {
     if (!entry.required) return false;
     const minScore = typeof entry.required === 'number' ? entry.required : PASS_THRESHOLD;