From 5c794d3bc6e89bd4a95e0597d2bd35fcc6ba6294 Mon Sep 17 00:00:00 2001 From: Jake Present Date: Tue, 26 May 2026 13:56:12 -0400 Subject: [PATCH 1/2] viewer: split policy violation metrics by permissibility --- tests/test_viewer_server_artifacts.py | 190 ++++++++++++++++++ viewer/src/lib/server/artifacts.ts | 41 ++++ viewer/src/lib/server/data.ts | 44 +++- viewer/src/lib/server/metrics.ts | 88 +++++++- viewer/src/lib/types.ts | 4 + .../suite/[suite_id]/[run_id]/+page.svelte | 74 +++++++ 6 files changed, 430 insertions(+), 11 deletions(-) diff --git a/tests/test_viewer_server_artifacts.py b/tests/test_viewer_server_artifacts.py index 8cbdb500..fb3910d3 100644 --- a/tests/test_viewer_server_artifacts.py +++ b/tests/test_viewer_server_artifacts.py @@ -1723,6 +1723,194 @@ def test_list_suites_marks_scenario_only_scored_suite_as_has_results(self) -> No self.assertEqual(payload[0]["status"], "has_results") + def test_run_metrics_policy_violation_by_permissibility(self) -> None: + with TemporaryDirectory(dir=ROOT / "viewer") as tmp_dir: + harness_dir = Path(tmp_dir) / "harness" + harness_dir.mkdir() + self._copy_data_harness(harness_dir) + metrics_path = harness_dir / "metrics.ts" + + script = textwrap.dedent( + f"""\ + const {{ computeRunMetrics }} = await import({json.dumps(metrics_path.as_uri())}); + const behaviors = [ + {{ name: 'perm_a', definition: '', examples: [], permissible: true }}, + {{ name: 'notperm_a', definition: '', examples: [], permissible: false }} + ]; + const samples = [ + {{ + test_case_id: 'seed-1', + prompt: 'prompt', + response: 'response', + behavior: 'perm_a', + target: 'target-model', + judge_model: 'judge-model', + judge_status: 'ok', + verdict: {{ + dimensions: {{ policy_violation: true, overrefusal: true }}, + justification: '', + node_judgments: [ + {{ node_index: 0, node_name: 'perm_a', relevant: true, violated: true, confidence: 'high', evidence_turns: [], reasoning: '' }}, + {{ node_index: 1, node_name: 'notperm_a', relevant: false, violated: null, confidence: null, evidence_turns: [], reasoning: '' }} + ] + }} + }}, + {{ + test_case_id: 'seed-2', + prompt: 'prompt', + response: 'response', + behavior: 'notperm_a', + target: 'target-model', + judge_model: 'judge-model', + judge_status: 'ok', + verdict: {{ + dimensions: {{ policy_violation: false, overrefusal: false }}, + justification: '', + node_judgments: [ + {{ node_index: 0, node_name: 'perm_a', relevant: true, violated: false, confidence: 'high', evidence_turns: [], reasoning: '' }}, + {{ node_index: 1, node_name: 'notperm_a', relevant: true, violated: true, confidence: 'high', evidence_turns: [], reasoning: '' }}, + {{ node_index: 2, node_name: 'unknown_node', relevant: true, violated: true, confidence: 'high', evidence_turns: [], reasoning: '' }} + ] + }} + }} + ]; + const withBehaviors = computeRunMetrics(samples, behaviors); + const withoutBehaviors = computeRunMetrics(samples, []); + console.log(JSON.stringify({{ withBehaviors, withoutBehaviors }})); + """ + ) + result = self._run_node( + harness_dir=harness_dir, + script=script, + env=os.environ.copy(), + ) + self.assertEqual(result.returncode, 0, msg=f"{result.stdout}\n{result.stderr}") + payload = json.loads(result.stdout) + + with_behaviors = payload["withBehaviors"] + self.assertEqual(with_behaviors["scored_total"], 2) + permissible = with_behaviors["policy_violation_on_permissible"] + self.assertIsNotNone(permissible) + self.assertEqual(permissible["count"], 2) + self.assertEqual(permissible["flagged_count"], 1) + self.assertEqual(permissible["clear_count"], 1) + self.assertAlmostEqual(permissible["rate"], 0.5) + + not_permissible = with_behaviors["policy_violation_on_not_permissible"] + self.assertIsNotNone(not_permissible) + self.assertEqual(not_permissible["count"], 1) + self.assertEqual(not_permissible["flagged_count"], 1) + self.assertEqual(not_permissible["clear_count"], 0) + self.assertAlmostEqual(not_permissible["rate"], 1.0) + + without_behaviors = payload["withoutBehaviors"] + self.assertIsNone(without_behaviors["policy_violation_on_permissible"]) + self.assertIsNone(without_behaviors["policy_violation_on_not_permissible"]) + + def test_audit_run_metrics_policy_violation_by_permissibility(self) -> None: + with TemporaryDirectory(dir=ROOT / "viewer") as tmp_dir: + harness_dir = Path(tmp_dir) / "harness" + harness_dir.mkdir() + self._copy_data_harness(harness_dir) + metrics_path = harness_dir / "metrics.ts" + + script = textwrap.dedent( + f"""\ + const {{ computeAuditRunMetrics }} = await import({json.dumps(metrics_path.as_uri())}); + const behaviors = [ + {{ name: 'perm_a', definition: '', examples: [], permissible: true }}, + {{ name: 'notperm_a', definition: '', examples: [], permissible: false }} + ]; + const scores = [ + {{ + test_case_id: 'scenario-1', + target: 'target-model', + tester_model: 'tester-model', + judge_model: 'judge-model', + judge_status: 'ok', + verdict: {{ + dimensions: {{ policy_violation: true, overrefusal: false }}, + justification: '', + node_judgments: [ + {{ node_index: 0, node_name: 'notperm_a', relevant: true, violated: true, confidence: 'high', evidence_turns: [], reasoning: '' }}, + {{ node_index: 1, node_name: 'perm_a', relevant: true, violated: false, confidence: 'high', evidence_turns: [], reasoning: '' }} + ] + }} + }} + ]; + const metrics = computeAuditRunMetrics(scores, behaviors); + console.log(JSON.stringify(metrics)); + """ + ) + result = self._run_node( + harness_dir=harness_dir, + script=script, + env=os.environ.copy(), + ) + self.assertEqual(result.returncode, 0, msg=f"{result.stdout}\n{result.stderr}") + metrics = json.loads(result.stdout) + + self.assertEqual(metrics["scored_total"], 1) + permissible = metrics["policy_violation_on_permissible"] + self.assertIsNotNone(permissible) + self.assertEqual(permissible["count"], 1) + self.assertEqual(permissible["flagged_count"], 0) + self.assertEqual(permissible["clear_count"], 1) + self.assertAlmostEqual(permissible["rate"], 0.0) + + not_permissible = metrics["policy_violation_on_not_permissible"] + self.assertIsNotNone(not_permissible) + self.assertEqual(not_permissible["count"], 1) + self.assertEqual(not_permissible["flagged_count"], 1) + self.assertEqual(not_permissible["clear_count"], 0) + self.assertAlmostEqual(not_permissible["rate"], 1.0) + + def test_load_run_judge_taxonomy_prefers_run_config(self) -> None: + with TemporaryDirectory(dir=ROOT / "viewer") as tmp_dir: + harness_dir = Path(tmp_dir) / "harness" + harness_dir.mkdir() + self._copy_data_harness(harness_dir) + artifacts_path = harness_dir / "artifacts.ts" + + suite_dir = Path(tmp_dir) / "demo-suite" + run_dir = suite_dir / "demo-run" + run_dir.mkdir(parents=True) + judge_taxonomy = { + "behavior": {"name": "demo", "definition": "demo"}, + "behavior_categories": [ + {"name": "judge_only", "definition": "", "examples": [], "permissible": False} + ], + } + taxonomy_path = suite_dir / "taxonomy.override.json" + taxonomy_path.write_text(json.dumps(judge_taxonomy), encoding="utf-8") + (run_dir / "config.yaml").write_text( + f"pipeline:\n judge:\n taxonomy_path: {taxonomy_path}\n", + encoding="utf-8", + ) + + script = textwrap.dedent( + f"""\ + const {{ loadRunJudgeTaxonomyForRun, loadRunJudgeTaxonomy, loadRunJudgeTaxonomyFromArtifacts }} = await import({json.dumps(artifacts_path.as_uri())}); + const fromRun = loadRunJudgeTaxonomyForRun('demo-suite', 'demo-run'); + const fromConfig = loadRunJudgeTaxonomy({{ pipeline: {{ judge: {{ taxonomy_path: {json.dumps(str(taxonomy_path))} }} }} }}); + const fromArtifact = loadRunJudgeTaxonomyFromArtifacts({{ suite: 'demo-suite' }}, {{ systematize: {{ path: 'taxonomy.override.json' }} }}); + const fromMissing = loadRunJudgeTaxonomy({{ pipeline: {{ judge: {{}} }} }}); + console.log(JSON.stringify({{ fromRun, fromConfig, fromArtifact, fromMissing }})); + """ + ) + env = os.environ.copy() + env["ARTIFACTS_ROOT"] = str(Path(tmp_dir)) + result = self._run_node(harness_dir=harness_dir, script=script, env=env) + self.assertEqual(result.returncode, 0, msg=f"{result.stdout}\n{result.stderr}") + payload = json.loads(result.stdout) + + self.assertEqual(payload["fromRun"], judge_taxonomy) + self.assertEqual(payload["fromConfig"], judge_taxonomy) + self.assertEqual(payload["fromArtifact"], judge_taxonomy) + self.assertIsNone(payload["fromMissing"]) + + + class ViewerReadModelHelpersTest(unittest.TestCase): """Tests for path-traversal defenses that don't depend on Node TS support.""" @@ -1822,5 +2010,7 @@ def test_test_set_artifact_path_rejects_paths_that_normalize_to_directory(self) ) + + if __name__ == "__main__": unittest.main() diff --git a/viewer/src/lib/server/artifacts.ts b/viewer/src/lib/server/artifacts.ts index 65a023c5..83c55c66 100644 --- a/viewer/src/lib/server/artifacts.ts +++ b/viewer/src/lib/server/artifacts.ts @@ -556,6 +556,47 @@ function readObject(value: unknown): Record | null { : null; } +export function loadRunJudgeTaxonomy(config: Record | null): Taxonomy | null { + return loadRunJudgeTaxonomyFromArtifacts(config, null); +} + +export function loadRunJudgeTaxonomyFromArtifacts( + config: Record | null, + artifacts: Record | null +): Taxonomy | null { + const systematize = readObject(artifacts?.systematize); + const artifactTaxonomyPath = typeof systematize?.path === 'string' ? systematize.path : null; + if (artifactTaxonomyPath) { + const resolvedArtifactPath = manifestArtifactPath(suiteDirPathFromConfig(config), artifactTaxonomyPath); + const artifactTaxonomy = resolvedArtifactPath + ? readJsonFile(resolvedArtifactPath, { missingOk: true }) + : null; + if (artifactTaxonomy) return artifactTaxonomy; + } + + const pipeline = readObject(config?.pipeline); + const judge = readObject(pipeline?.judge); + const rawTaxonomyPath = typeof judge?.taxonomy_path === 'string' ? judge.taxonomy_path : null; + if (!rawTaxonomyPath) return null; + + const resolved = path.resolve(rawTaxonomyPath); + return readJsonFile(resolved, { missingOk: true }); +} + +function suiteDirPathFromConfig(config: Record | null): string { + const suite = typeof config?.suite === 'string' ? config.suite : null; + return suite ? suiteDirPath(suite) : ARTIFACTS_ROOT; +} + +export function loadRunJudgeTaxonomyForRun(suiteId: string, runId: string): Taxonomy | null { + const runDir = runDirPath(suiteId, runId); + const config = readYamlFile>(path.join(runDir, RUN_CONFIG_FILE), { + missingOk: true + }); + const manifest = readJsonFile(path.join(runDir, RUN_MANIFEST_FILE), { missingOk: true }); + return loadRunJudgeTaxonomyFromArtifacts(config, manifest?.artifact_versions ?? null); +} + export function loadRunRuntimeMode(config: Record | null): string | null { const pipeline = readObject(config?.pipeline); const inference = readObject(pipeline?.inference); diff --git a/viewer/src/lib/server/data.ts b/viewer/src/lib/server/data.ts index 1a876dd2..9a02de2e 100644 --- a/viewer/src/lib/server/data.ts +++ b/viewer/src/lib/server/data.ts @@ -6,6 +6,8 @@ import { ViewerReadModelError, loadIndexedRunScoreRow, loadIndexedRunTranscriptRow, + loadRunJudgeTaxonomyForRun, + loadRunJudgeTaxonomyFromArtifacts, loadRunRuntimeMode, loadRunScoreRow, loadRunTranscriptRow, @@ -64,6 +66,8 @@ interface PromptMetricView { counts: BinaryCounts; policyViolationRate: number; overrefusalRate: number; + policyViolationOnPermissible: DimensionMetrics | null; + policyViolationOnNotPermissible: DimensionMetrics | null; dimensions: Record; target: string; judge_model: string; @@ -77,6 +81,8 @@ interface AuditMetricView { counts: BinaryCounts; policyViolationRate: number; overrefusalRate: number; + policyViolationOnPermissible: DimensionMetrics | null; + policyViolationOnNotPermissible: DimensionMetrics | null; dimensions: Record; target: string; tester_model: string; @@ -179,6 +185,15 @@ function normalizeBehavior(b: Behavior): Behavior { return { ...b, permissible: b.permissible ?? false }; } +function metricBehaviors( + snapshot: SuiteSnapshot | null, + runConfig?: Record | null, + artifacts?: Record | null +): Behavior[] { + const judgeTaxonomy = loadRunJudgeTaxonomyFromArtifacts(runConfig ?? null, artifacts ?? null); + return (judgeTaxonomy?.behavior_categories ?? snapshot?.taxonomy?.behavior_categories ?? []).map(normalizeBehavior); +} + function normalizePolicy(taxonomy: Taxonomy | null | undefined): Taxonomy | null { if (!taxonomy) return null; const behavior = taxonomy.behavior ?? taxonomy.risk; @@ -649,6 +664,7 @@ function buildRunListEntries(snapshot: SuiteSnapshot): { const hasPromptScores = promptScores.length > 0; const hasAuditScores = auditScores.length > 0; const hasScoreStage = manifest?.stages?.judge != null; + const behaviors = metricBehaviors(snapshot, runSnapshot.config, manifest?.artifact_versions ?? null); if ((hasPromptScores || hasScoreStage) && !(manifest?.status === 'failed' && !hasPromptScores)) { runs.push({ @@ -656,7 +672,7 @@ function buildRunListEntries(snapshot: SuiteSnapshot): { has_judged: hasPromptScores, has_scenario_scores: hasAuditScores, manifest, - metrics: hasPromptScores ? computeRunMetrics(promptScores) : null + metrics: hasPromptScores ? computeRunMetrics(promptScores, behaviors) : null }); } @@ -665,7 +681,7 @@ function buildRunListEntries(snapshot: SuiteSnapshot): { run_id: runId, has_scores: hasAuditScores, manifest, - metrics: hasAuditScores ? computeAuditRunMetrics(auditScores) : null + metrics: hasAuditScores ? computeAuditRunMetrics(auditScores, behaviors) : null }); } } @@ -682,6 +698,8 @@ function buildZeroPromptMetrics(): PromptMetricView { counts: emptyScoreCounts(), policyViolationRate: 0, overrefusalRate: 0, + policyViolationOnPermissible: null, + policyViolationOnNotPermissible: null, dimensions: {}, target: '', judge_model: '' @@ -697,6 +715,8 @@ function buildZeroAuditMetrics(): AuditMetricView { counts: emptyScoreCounts(), policyViolationRate: 0, overrefusalRate: 0, + policyViolationOnPermissible: null, + policyViolationOnNotPermissible: null, dimensions: {}, target: '', tester_model: '', @@ -714,6 +734,8 @@ function toPromptMetricView(metrics: RunMetrics | null): PromptMetricView { counts: metrics.counts, policyViolationRate: metrics.policy_violation_rate, overrefusalRate: metrics.overrefusal_rate, + policyViolationOnPermissible: metrics.policy_violation_on_permissible, + policyViolationOnNotPermissible: metrics.policy_violation_on_not_permissible, dimensions: metrics.dimensions, target: metrics.target, judge_model: metrics.judge_model @@ -730,6 +752,8 @@ function toAuditMetricView(metrics: AuditRunMetrics | null): AuditMetricView { counts: metrics.counts, policyViolationRate: metrics.policy_violation_rate, overrefusalRate: metrics.overrefusal_rate, + policyViolationOnPermissible: metrics.policy_violation_on_permissible, + policyViolationOnNotPermissible: metrics.policy_violation_on_not_permissible, dimensions: metrics.dimensions, target: metrics.target, tester_model: metrics.tester_model, @@ -1080,6 +1104,7 @@ async function loadSuiteHeavyData( const hasPromptScores = promptScores.length > 0; const hasAuditScores = auditScores.length > 0; const hasScoreStage = manifest?.stages?.judge != null; + const behaviors = metricBehaviors(snapshot, runSnapshot.config, manifest?.artifact_versions ?? null); const addedToRuns = (hasPromptScores || hasScoreStage) && @@ -1094,7 +1119,7 @@ async function loadSuiteHeavyData( has_judged: hasPromptScores, has_scenario_scores: hasAuditScores, manifest, - metrics: hasPromptScores ? computeRunMetrics(promptScores) : null + metrics: hasPromptScores ? computeRunMetrics(promptScores, behaviors) : null }); } if (addedToAuditRuns) { @@ -1102,7 +1127,7 @@ async function loadSuiteHeavyData( run_id: runId, has_scores: hasAuditScores, manifest, - metrics: hasAuditScores ? computeAuditRunMetrics(auditScores) : null + metrics: hasAuditScores ? computeAuditRunMetrics(auditScores, behaviors) : null }); } @@ -1204,8 +1229,10 @@ function loadCompletedRunPageData( const samples = resolvedTab === 'prompts' ? promptRows : []; const auditScores = resolvedTab === 'audit' ? auditRows : []; const scenarioSeeds = buildScenarioSeeds(suiteSnapshot); - const promptMetrics = resolvedTab === 'prompts' ? computeRunMetrics(samples) : null; - const auditMetrics = resolvedTab === 'audit' ? computeAuditRunMetrics(auditScores) : null; + const judgeTaxonomy = loadRunJudgeTaxonomyForRun(suiteId, runId); + const behaviors = (judgeTaxonomy?.behavior_categories ?? suiteSnapshot?.taxonomy?.behavior_categories ?? []).map(normalizeBehavior); + const promptMetrics = resolvedTab === 'prompts' ? computeRunMetrics(samples, behaviors) : null; + const auditMetrics = resolvedTab === 'audit' ? computeAuditRunMetrics(auditScores, behaviors) : null; return { suite_id: suiteId, @@ -1270,8 +1297,9 @@ export function loadRunPageData(suiteId: string, runId: string, activeTab: 'prom const scenarioSeeds = buildScenarioSeeds(suiteSnapshot); const promptSeedTitleMap = buildPromptSeedTitleMap(suiteSnapshot); - const promptMetrics = resolvedTab === 'prompts' ? computeRunMetrics(samples) : null; - const auditMetrics = resolvedTab === 'audit' ? computeAuditRunMetrics(auditScores) : null; + const behaviors = metricBehaviors(suiteSnapshot, runSnapshot.config, runSnapshot.manifest?.artifact_versions ?? null); + const promptMetrics = resolvedTab === 'prompts' ? computeRunMetrics(samples, behaviors) : null; + const auditMetrics = resolvedTab === 'audit' ? computeAuditRunMetrics(auditScores, behaviors) : null; const scenarioSeedMap = resolvedTab === 'audit' ? buildScenarioSeedMap(scenarioSeeds, auditScores) : {}; return { diff --git a/viewer/src/lib/server/metrics.ts b/viewer/src/lib/server/metrics.ts index 0d140be0..b4a3ca94 100644 --- a/viewer/src/lib/server/metrics.ts +++ b/viewer/src/lib/server/metrics.ts @@ -4,7 +4,16 @@ import { isBooleanFlag, isSuccessfulJudgment } from '$lib/judgment.js'; -import type { AuditScore, AuditRunMetrics, BinaryCounts, DimensionMetrics, JudgedSample, RunMetrics } from '$lib/types.js'; +import type { + AuditScore, + AuditRunMetrics, + Behavior, + BinaryCounts, + DimensionMetrics, + JudgedSample, + NodeJudgment, + RunMetrics +} from '$lib/types.js'; import { loadDimensions } from './dimensions.js'; type EventScoredRecord = { @@ -22,6 +31,67 @@ export function emptyScoreCounts(): BinaryCounts { return { 0: 0, 1: 0 }; } +function emptyDimensionAggregate(): EventDimensionAggregate { + return { count: 0, flagged_count: 0, clear_count: 0, counts: emptyScoreCounts() }; +} + +function finalizeDimensionAggregate(aggregate: EventDimensionAggregate): DimensionMetrics { + return { + rate: aggregate.count > 0 ? aggregate.flagged_count / aggregate.count : 0, + count: aggregate.count, + flagged_count: aggregate.flagged_count, + clear_count: aggregate.clear_count, + counts: aggregate.counts + }; +} + +function readNodeJudgments(verdict: Record | null | undefined): NodeJudgment[] { + if (!verdict || typeof verdict !== 'object') return []; + const nodes = (verdict as Record).node_judgments; + return Array.isArray(nodes) ? (nodes as NodeJudgment[]) : []; +} + +function buildPermissibilityIndex(behaviors: Behavior[]): Map { + const index = new Map(); + for (const behavior of behaviors) { + if (!behavior || typeof behavior.name !== 'string') continue; + index.set(behavior.name, behavior.permissible === true); + } + return index; +} + +export function computePolicyViolationByPermissibility( + records: EventScoredRecord[], + behaviors: Behavior[] +): { permissible: DimensionMetrics | null; not_permissible: DimensionMetrics | null } { + if (!behaviors || behaviors.length === 0) { + return { permissible: null, not_permissible: null }; + } + const permissibilityIndex = buildPermissibilityIndex(behaviors); + if (permissibilityIndex.size === 0) { + return { permissible: null, not_permissible: null }; + } + + const permissible = emptyDimensionAggregate(); + const notPermissible = emptyDimensionAggregate(); + + for (const record of records) { + for (const node of readNodeJudgments(record.verdict)) { + if (node.relevant !== true) continue; + if (!isBooleanFlag(node.violated)) continue; + const name = typeof node.node_name === 'string' ? node.node_name.trim() : ''; + if (!name || !permissibilityIndex.has(name)) continue; + const bucket = permissibilityIndex.get(name) ? permissible : notPermissible; + addFlag(bucket, node.violated); + } + } + + return { + permissible: finalizeDimensionAggregate(permissible), + not_permissible: finalizeDimensionAggregate(notPermissible) + }; +} + function collectDimensionNames(records: EventScoredRecord[]): string[] { const names = new Set(); for (const record of records) { @@ -77,7 +147,10 @@ function dimensionRate(dimensions: Record, name: strin return dimensions[name]?.rate ?? 0; } -export function computeAuditRunMetrics(scores: AuditScore[]): AuditRunMetrics | null { +export function computeAuditRunMetrics( + scores: AuditScore[], + behaviors: Behavior[] = [] +): AuditRunMetrics | null { if (scores.length === 0) return null; const requiredBaseMetrics = getRequiredBaseMetricNames(loadDimensions()); @@ -100,6 +173,7 @@ export function computeAuditRunMetrics(scores: AuditScore[]): AuditRunMetrics | const dimensions = finalizeDimensions(dimensionAggregates); const total = scores.length; const scoredTotal = scoredScores.length; + const permissibilitySplit = computePolicyViolationByPermissibility(scoredScores, behaviors); return { total, @@ -109,6 +183,8 @@ export function computeAuditRunMetrics(scores: AuditScore[]): AuditRunMetrics | counts, policy_violation_rate: dimensionRate(dimensions, 'policy_violation'), overrefusal_rate: dimensionRate(dimensions, 'overrefusal'), + policy_violation_on_permissible: permissibilitySplit.permissible, + policy_violation_on_not_permissible: permissibilitySplit.not_permissible, dimensions, target: scores[0]?.target ?? '', tester_model: scores[0]?.tester_model ?? '', @@ -116,7 +192,10 @@ export function computeAuditRunMetrics(scores: AuditScore[]): AuditRunMetrics | }; } -export function computeRunMetrics(samples: JudgedSample[]): RunMetrics | null { +export function computeRunMetrics( + samples: JudgedSample[], + behaviors: Behavior[] = [] +): RunMetrics | null { if (samples.length === 0) return null; const requiredBaseMetrics = getRequiredBaseMetricNames(loadDimensions()); @@ -137,6 +216,7 @@ export function computeRunMetrics(samples: JudgedSample[]): RunMetrics | null { } const dimensions = finalizeDimensions(dimensionAggregates); + const permissibilitySplit = computePolicyViolationByPermissibility(scoredSamples, behaviors); return { total: samples.length, @@ -147,6 +227,8 @@ export function computeRunMetrics(samples: JudgedSample[]): RunMetrics | null { counts, policy_violation_rate: dimensionRate(dimensions, 'policy_violation'), overrefusal_rate: dimensionRate(dimensions, 'overrefusal'), + policy_violation_on_permissible: permissibilitySplit.permissible, + policy_violation_on_not_permissible: permissibilitySplit.not_permissible, target: samples[0]?.target ?? '—', judge_model: samples[0]?.judge_model ?? '—', dimensions diff --git a/viewer/src/lib/types.ts b/viewer/src/lib/types.ts index 05263138..40ee9b5c 100644 --- a/viewer/src/lib/types.ts +++ b/viewer/src/lib/types.ts @@ -273,6 +273,8 @@ export interface RunMetrics { counts: BinaryCounts; policy_violation_rate: number; overrefusal_rate: number; + policy_violation_on_permissible: DimensionMetrics | null; + policy_violation_on_not_permissible: DimensionMetrics | null; target: string; judge_model: string; dimensions: Record; @@ -301,6 +303,8 @@ export interface AuditRunMetrics { counts: BinaryCounts; policy_violation_rate: number; overrefusal_rate: number; + policy_violation_on_permissible: DimensionMetrics | null; + policy_violation_on_not_permissible: DimensionMetrics | null; dimensions: Record; target: string; tester_model: string; diff --git a/viewer/src/routes/suite/[suite_id]/[run_id]/+page.svelte b/viewer/src/routes/suite/[suite_id]/[run_id]/+page.svelte index 76d16059..da8cf99e 100644 --- a/viewer/src/routes/suite/[suite_id]/[run_id]/+page.svelte +++ b/viewer/src/routes/suite/[suite_id]/[run_id]/+page.svelte @@ -919,6 +919,43 @@

{/if} + + {#if data.metrics.policyViolationOnPermissible || data.metrics.policyViolationOnNotPermissible} + {@const promptPerm = data.metrics.policyViolationOnPermissible} + {@const promptNotPerm = data.metrics.policyViolationOnNotPermissible} + {#if (promptPerm?.count ?? 0) + (promptNotPerm?.count ?? 0) > 0} +
+ {#each [ + { key: 'permissible', label: 'Allowed requests failed', summary: promptPerm, hint: 'Requests the assistant should have helped with, but instead refused, deflected, or failed to complete.' }, + { key: 'not_permissible', label: 'Blocked requests failed', summary: promptNotPerm, hint: 'Requests the assistant should have blocked, but instead helped with or allowed.' } + ] as card (card.key)} + {@const summary = card.summary} +
+
{card.label}
+

{card.hint}

+
+ {summary && summary.count > 0 ? metricRateText(summary.rate) : '—'} + {summary && summary.count > 0 ? 'violated' : 'no relevant judgments'} +
+ {#if summary && summary.count > 0} + {@const pct = binaryBar(summary.counts)} +
+ {#if pct.clear > 0}
{/if} + {#if pct.flagged > 0}
{/if} +
+
+ {summary.clear_count} clear + {summary.flagged_count} violated + {summary.count} relevant judgments +
+ {/if} +
aggregated over relevant behavior judgments, not over prompts
+
+ {/each} +
+ {/if} + {/if} +
@@ -1199,6 +1236,43 @@

{/if} + + {#if data.auditMetrics.policyViolationOnPermissible || data.auditMetrics.policyViolationOnNotPermissible} + {@const auditPerm = data.auditMetrics.policyViolationOnPermissible} + {@const auditNotPerm = data.auditMetrics.policyViolationOnNotPermissible} + {#if (auditPerm?.count ?? 0) + (auditNotPerm?.count ?? 0) > 0} +
+ {#each [ + { key: 'permissible', label: 'Allowed requests failed', summary: auditPerm, hint: 'Requests the assistant should have helped with, but instead refused, deflected, or failed to complete.' }, + { key: 'not_permissible', label: 'Blocked requests failed', summary: auditNotPerm, hint: 'Requests the assistant should have blocked, but instead helped with or allowed.' } + ] as card (card.key)} + {@const summary = card.summary} +
+
{card.label}
+

{card.hint}

+
+ {summary && summary.count > 0 ? metricRateText(summary.rate) : '—'} + {summary && summary.count > 0 ? 'violated' : 'no relevant judgments'} +
+ {#if summary && summary.count > 0} + {@const pct = binaryBar(summary.counts)} +
+ {#if pct.clear > 0}
{/if} + {#if pct.flagged > 0}
{/if} +
+
+ {summary.clear_count} clear + {summary.flagged_count} violated + {summary.count} relevant judgments +
+ {/if} +
aggregated over relevant behavior judgments, not over scenarios
+
+ {/each} +
+ {/if} + {/if} +
From 86ae430d348b2eb7dc1959c099143e61440ca6f7 Mon Sep 17 00:00:00 2001 From: Jake Present Date: Tue, 26 May 2026 14:29:29 -0400 Subject: [PATCH 2/2] fix: omit unsupported GPT-5 temperature overrides --- p2m/core/model_client.py | 40 ++++++++++++++++-- tests/test_model_client.py | 84 ++++++++++++++++++++++++++++++++++++-- 2 files changed, 117 insertions(+), 7 deletions(-) diff --git a/p2m/core/model_client.py b/p2m/core/model_client.py index 5d9945ff..3cf11b8b 100644 --- a/p2m/core/model_client.py +++ b/p2m/core/model_client.py @@ -311,6 +311,36 @@ def _supports_web_search_preview(model: str) -> bool: return _model_family(model) in {"openai", "azure"} +def _model_name(model: str) -> str: + normalized = (model or "").strip().lower() + if "/" in normalized: + return normalized.split("/", 1)[1] + return normalized + + +def _supports_custom_temperature(model: str) -> bool: + """Whether non-default temperature values are supported for this model. + + GPT-5.x deployments on OpenAI/Azure reject explicit non-default + temperatures. LiteLLM's Azure model string uses the deployment name after + ``azure/``, so match prefixes instead of exact public model IDs. + """ + return not _model_name(model).startswith("gpt-5") + + +def _temperature_for_payload(model: str, temperature: float | None) -> float | None: + if temperature is None or temperature == 1: + return temperature + if _supports_custom_temperature(model): + return temperature + log.warning( + "Model %s only supports the default temperature; ignoring configured temperature=%s", + model, + temperature, + ) + return None + + def _require_web_search_preview_support(model: str) -> None: if _supports_web_search_preview(model): return @@ -493,8 +523,9 @@ def _build_chat_payload( "model": model, "messages": messages_to_openai(messages), } - if resolved_options.temperature is not None: - payload["temperature"] = resolved_options.temperature + temperature = _temperature_for_payload(model, resolved_options.temperature) + if temperature is not None: + payload["temperature"] = temperature if resolved_options.max_tokens is not None: payload["max_tokens"] = resolved_options.max_tokens if resolved_options.max_output_tokens is not None and "max_tokens" not in payload: @@ -519,8 +550,9 @@ def _build_responses_payload( "model": model, "input": input_payload, } - if resolved_options.temperature is not None: - payload["temperature"] = resolved_options.temperature + temperature = _temperature_for_payload(model, resolved_options.temperature) + if temperature is not None: + payload["temperature"] = temperature if resolved_options.max_output_tokens is not None: payload["max_output_tokens"] = resolved_options.max_output_tokens elif resolved_options.max_tokens is not None: diff --git a/tests/test_model_client.py b/tests/test_model_client.py index f9c45b41..21a5bdc9 100644 --- a/tests/test_model_client.py +++ b/tests/test_model_client.py @@ -13,7 +13,7 @@ async def fake_acompletion(**kwargs): captured.update(kwargs) return { "id": "resp-chat-1", - "model": "openai/gpt-5-mini", + "model": "openai/gpt-4o-mini", "choices": [ { "finish_reason": "stop", @@ -36,7 +36,7 @@ async def fake_acompletion(**kwargs): with patch.object(model_client, "_get_litellm_module", return_value=fake_litellm): response = await model_client.generate( - "openai/gpt-5-mini", + "openai/gpt-4o-mini", "say hi", options, ) @@ -51,9 +51,87 @@ async def fake_acompletion(**kwargs): self.assertEqual(response.finish_reason, "stop") self.assertEqual(response.usage.total_tokens, 18) self.assertEqual(response.api_mode, "chat_completion") - self.assertEqual(response.request_payload["model"], "openai/gpt-5-mini") + self.assertEqual(response.request_payload["model"], "openai/gpt-4o-mini") self.assertEqual(response.request_payload["messages"], [{"role": "user", "content": "say hi"}]) + async def test_generate_omits_unsupported_gpt5_temperature(self) -> None: + captured: dict[str, object] = {} + + async def fake_acompletion(**kwargs): + captured.update(kwargs) + return { + "choices": [ + { + "finish_reason": "stop", + "message": {"role": "assistant", "content": "ok"}, + } + ] + } + + fake_litellm = SimpleNamespace(acompletion=fake_acompletion) + options = model_client.GenerateOptions(temperature=0.0) + + with patch.object(model_client, "_get_litellm_module", return_value=fake_litellm): + await model_client.generate( + "azure/gpt-5.4-1", + "say hi", + options, + ) + + self.assertNotIn("temperature", captured) + + async def test_generate_keeps_default_gpt5_temperature(self) -> None: + captured: dict[str, object] = {} + + async def fake_acompletion(**kwargs): + captured.update(kwargs) + return { + "choices": [ + { + "finish_reason": "stop", + "message": {"role": "assistant", "content": "ok"}, + } + ] + } + + fake_litellm = SimpleNamespace(acompletion=fake_acompletion) + options = model_client.GenerateOptions(temperature=1.0) + + with patch.object(model_client, "_get_litellm_module", return_value=fake_litellm): + await model_client.generate( + "azure/gpt-5.4-1", + "say hi", + options, + ) + + self.assertEqual(captured["temperature"], 1.0) + + async def test_generate_keeps_non_gpt5_temperature(self) -> None: + captured: dict[str, object] = {} + + async def fake_acompletion(**kwargs): + captured.update(kwargs) + return { + "choices": [ + { + "finish_reason": "stop", + "message": {"role": "assistant", "content": "ok"}, + } + ] + } + + fake_litellm = SimpleNamespace(acompletion=fake_acompletion) + options = model_client.GenerateOptions(temperature=0.0) + + with patch.object(model_client, "_get_litellm_module", return_value=fake_litellm): + await model_client.generate( + "azure/gpt-4o-mini", + "say hi", + options, + ) + + self.assertEqual(captured["temperature"], 0.0) + async def test_generate_structured_adds_json_schema_response_format(self) -> None: captured: dict[str, object] = {}