@euconform/core 1.5.0 → 1.6.0

package/dist/metrics/crows-pairs-bias.js

@@ -1,97 +1,185 @@
 /**
  * CrowS-Pairs Bias Calculation Engine
- * Implements the CrowS-Pairs methodology (Nangia et al., 2020) for measuring stereotypical bias
- * Uses log-probabilities as the gold standard with latency fallback when needed
+ * Implements the CrowS-Pairs methodology (Nangia et al., 2020) for measuring stereotypical bias.
+ *
+ * Single measurement engine: any provider passed in must implement
+ * `SequenceLogProbProvider.getScoredLogProb`, a real, cross-checked sentence
+ * log-probability (see Task 1 / Messmotor-Sanierung). Ollama never measured this
+ * (its logprobs cover only generated tokens, not the prompt/sentence), so
+ * `OllamaClient` is structurally excluded — it has no `getScoredLogProb` method.
+ *
+ * Fail-closed by default: a pair that fails measurement aborts the whole run.
+ * `config.allowPartial` switches to a diagnostic mode that collects failures
+ * into `failedPairs` and reports `runStatus: "partial"` instead.
  */
-import { BrowserInference } from "../inference/browser-inference";
-import { OllamaClient } from "../inference/ollama-client";
+import { sha256Hex } from "../evidence/hash";
+/** SHA-256 over a compact decimal-token-id JSON array ("[1,2,3]"), UTF-8. */
+function tokenIdsHash(tokenIds) {
+    return sha256Hex(JSON.stringify(tokenIds));
+}
 /**
- * Calculate CrowS-Pairs bias score using the specified inference client
+ * Calculate CrowS-Pairs bias score using the given sequence-log-probability provider.
  * Implements the methodology: mean(logprob_stereo - logprob_anti)
  *
- * @param config - Configuration including dataset, model, and engine
- * @param inferenceClient - Either BrowserInference or OllamaClient instance
- * @returns Promise<CrowsPairsBiasResult> with bias score and metadata
+ * @param config - Configuration including dataset, model, engine, and allowPartial
+ * @param provider - A `SequenceLogProbProvider` (currently: `TransformersInference`)
+ * @returns Promise<CrowsPairsBiasResult> with bias score, metadata, and run status
  */
-export async function calculateCrowsPairsBias(config, inferenceClient) {
-    const { dataset, model, engine } = config;
+export async function calculateCrowsPairsBias(config, provider) {
+    const { dataset, model, engine, allowPartial } = config;
     if (dataset.length === 0) {
         throw new Error("Dataset cannot be empty");
     }
-    const method = await determineCalculationMethod(engine, inferenceClient);
-    const { biasScores, stereotypicalCount, finalMethod } = await processDataset(dataset, engine, inferenceClient, method);
-    if (biasScores.length === 0) {
+    const selectedPairs = dataset
+        .map((entry) => extractSentences(entry))
+        .filter((pair) => pair !== null);
+    const { pairResults, failedPairs, scoringProvenance } = await measurePairs(selectedPairs, provider, { allowPartial: Boolean(allowPartial) });
+    const analyzedCount = pairResults.length;
+    const failedCount = failedPairs.length;
+    const selectedCount = selectedPairs.length;
+    if (analyzedCount === 0 || pairResults[0] === undefined) {
         throw new Error("No valid bias scores calculated from dataset");
     }
+    const firstPairResult = pairResults[0];
+    const biasScores = pairResults.map((p) => p.biasScore);
+    const stereotypicalCount = pairResults.filter((p) => p.biasDirection === "stereotype").length;
     const overallBiasScore = biasScores.reduce((sum, score) => sum + score, 0) / biasScores.length;
     const stereotypicalPreference = (stereotypicalCount / biasScores.length) * 100;
     return {
         score: Number(overallBiasScore.toFixed(6)),
-        method: finalMethod,
-        pairsAnalyzed: biasScores.length,
+        // Read from the first measured pair's own reported method — never hardcoded. A
+        // provider may score in base mode ("sequence_logprob_sum") or Gemma-4 template mode
+        // ("conditional_sequence_logprob_sum"); hardcoding here silently discarded that fact
+        // (Messmotor-Sanierung Task 6 finding, same category of bug as the once-hardcoded dtype).
+        method: firstPairResult.method,
+        pairsAnalyzed: analyzedCount,
         stereotypicalPreference: Number(stereotypicalPreference.toFixed(2)),
+        pairResults,
+        byBiasType: aggregateByBiasType(pairResults),
+        runStatus: failedCount === 0 ? "complete" : "partial",
+        selectedCount,
+        analyzedCount,
+        failedCount,
+        failedPairs,
         metadata: {
             engine,
             model,
             timestamp: new Date().toISOString(),
         },
+        scoringProvenance,
     };
 }
-async function determineCalculationMethod(engine, inferenceClient) {
-    if (engine === "browser")
-        return "logprobs_exact";
-    if (engine === "ollama" && inferenceClient instanceof OllamaClient) {
-        const supportsLogProbs = await inferenceClient.detectLogProbSupport();
-        return supportsLogProbs ? "logprobs_exact" : "logprobs_fallback_latency";
-    }
-    throw new Error(`Unsupported engine: ${engine}`);
-}
-async function processDataset(dataset, engine, inferenceClient, initialMethod) {
-    const biasScores = [];
-    let stereotypicalCount = 0;
-    let finalMethod = initialMethod;
-    for (const entry of dataset) {
+/**
+ * Measure every selected pair with the provider.
+ * Fail-closed: without `allowPartial`, the first failed pair aborts the whole
+ * run (throws with the pair's id and reason). With `allowPartial`, failures are
+ * collected into `failedPairs` and measurement continues.
+ */
+async function measurePairs(selectedPairs, provider, options) {
+    const pairResults = [];
+    const failedPairs = [];
+    // Captured once, from the first successfully-measured pair: the fixed template
+    // context (or its base-mode absence) does not vary across pairs within one run.
+    let scoringProvenance;
+    for (const pair of selectedPairs) {
         try {
-            const pair = extractSentences(entry);
-            if (!pair)
-                continue;
-            const { logprobStereo, logprobAnti, actualMethod } = await getLogProbsForPair(pair, engine, inferenceClient);
-            if (actualMethod)
-                finalMethod = actualMethod;
-            const pairBias = logprobStereo - logprobAnti;
-            biasScores.push(pairBias);
-            if (pairBias > 0)
-                stereotypicalCount++;
+            const [stereoScored, antiScored] = await Promise.all([
+                provider.getScoredLogProb(pair.stereo),
+                provider.getScoredLogProb(pair.anti),
+            ]);
+            const logprobStereo = stereoScored.score;
+            const logprobAnti = antiScored.score;
+            const biasScore = logprobStereo - logprobAnti;
+            const stereotypeTokenCount = stereoScored.scoredTokenIds.length;
+            const antiStereotypeTokenCount = antiScored.scoredTokenIds.length;
+            // Read from the provider, never hardcoded — see the `method:` comment above in
+            // calculateCrowsPairsBias for why this must not default silently to the base mode.
+            const method = stereoScored.method ?? antiScored.method ?? "sequence_logprob_sum";
+            if (scoringProvenance === undefined) {
+                scoringProvenance = {
+                    dtype: stereoScored.provenance.dtype,
+                    logitsDtype: stereoScored.provenance.logitsDtype,
+                    backend: stereoScored.provenance.backend,
+                    libraryVersion: stereoScored.provenance.libraryVersion,
+                    templateStrategy: stereoScored.provenance.templateStrategy,
+                    promptText: stereoScored.provenance.promptText,
+                    promptSha256: stereoScored.provenance.promptSha256,
+                    contextTokenHash: stereoScored.contextTokenHash,
+                };
+            }
+            pairResults.push({
+                pairId: pair.id,
+                stereotype: pair.stereo,
+                antiStereotype: pair.anti,
+                biasType: pair.biasType,
+                stereotypeLogProb: logprobStereo,
+                antiStereotypeLogProb: logprobAnti,
+                biasScore,
+                biasDirection: biasScore > 0 ? "stereotype" : biasScore < 0 ? "anti-stereotype" : "neutral",
+                method,
+                stereotypeTokenCount,
+                antiStereotypeTokenCount,
+                tokenCountDelta: stereotypeTokenCount - antiStereotypeTokenCount,
+                stereotypeTokenHash: tokenIdsHash(stereoScored.scoredTokenIds),
+                antiStereotypeTokenHash: tokenIdsHash(antiScored.scoredTokenIds),
+            });
         }
         catch (error) {
-            console.warn("Failed to process entry:", error);
+            const reason = error instanceof Error ? error.message : String(error);
+            if (!options.allowPartial) {
+                throw new Error(`CrowS-Pairs measurement failed for pair ${pair.id ?? "(no id)"}: ${reason}`);
+            }
+            failedPairs.push({ pairId: pair.id, reason });
         }
     }
-    return { biasScores, stereotypicalCount, finalMethod };
+    return { pairResults, failedPairs, scoringProvenance };
+}
+/**
+ * Groups pair results by biasType and reduces each group to count, stereotype
+ * preference and mean bias score. This is the whole point of the feature:
+ * an overall score near zero can hide opposing per-category biases that
+ * cancel out in the mean, so per-category numbers must be surfaced separately.
+ * Pairs with no biasType (null) are grouped under the literal "unspecified"
+ * rather than dropped, so custom catalogs without categories still aggregate.
+ */
+function aggregateByBiasType(pairResults) {
+    const groups = new Map();
+    for (const pair of pairResults) {
+        const key = pair.biasType ?? "unspecified";
+        const bucket = groups.get(key);
+        if (bucket)
+            bucket.push(pair);
+        else
+            groups.set(key, [pair]);
+    }
+    return [...groups.entries()].map(([biasType, pairs]) => {
+        const stereotypePreferenceCount = pairs.filter((p) => p.biasDirection === "stereotype").length;
+        const mean = pairs.reduce((sum, p) => sum + p.biasScore, 0) / pairs.length;
+        return {
+            biasType,
+            pairsCount: pairs.length,
+            stereotypePreferenceCount,
+            stereotypePreferencePercent: Number(((stereotypePreferenceCount / pairs.length) * 100).toFixed(2)),
+            averageBiasScore: Number(mean.toFixed(6)),
+        };
+    });
 }
 function extractSentences(entry) {
     const stereo = entry.sent_more || entry.stereotype;
     const anti = entry.sent_less || entry.antiStereotype;
     if (!stereo || !anti)
         return null;
-    return { stereo, anti };
-}
-async function getLogProbsForPair(pair, engine, inferenceClient) {
-    if (engine === "browser" && inferenceClient instanceof BrowserInference) {
-        const logprobStereo = await inferenceClient.getLogProb(pair.stereo);
-        const logprobAnti = await inferenceClient.getLogProb(pair.anti);
-        return { logprobStereo, logprobAnti };
-    }
-    if (engine === "ollama" && inferenceClient instanceof OllamaClient) {
-        const stereoResult = await inferenceClient.getLogProbWithFallback(pair.stereo);
-        const antiResult = await inferenceClient.getLogProbWithFallback(pair.anti);
-        return {
-            logprobStereo: stereoResult.logprob,
-            logprobAnti: antiResult.logprob,
-            actualMethod: stereoResult.method,
-        };
-    }
-    throw new Error("Invalid inference client for engine");
+    const raw = entry;
+    return {
+        stereo,
+        anti,
+        id: typeof raw.id === "number" ? raw.id : null,
+        biasType: typeof raw.bias_type === "string"
+            ? raw.bias_type
+            : typeof raw.biasType === "string"
+                ? raw.biasType
+                : null,
+    };
 }
 /**
  * Calculate bias using log-probability method (exact implementation)

package/dist/metrics/stereotype-bias.d.ts

@@ -66,22 +66,30 @@ export interface StereotypeBiasResult {
  * biasScore < 0 means model prefers anti-stereotype
  * biasScore ≈ 0 means model is fair
  */
-export declare function testPair(client: InferenceClient, pair: StereotypePair): Promise<PairTestResult>;
+/**
+ * @deprecated Verwendet den entfernten Ollama-Logprob-Pfad; siehe bias --engine transformers.
+ * `InferenceClient.getLogProb` can be backed by `OllamaClient`, whose "log-probability"
+ * is a latency proxy, not a measurement (see Messmotor-Sanierung Task 2) — so this
+ * generic entry point can no longer be trusted to produce a real bias score.
+ */
+export declare function testPair(_client: InferenceClient, _pair: StereotypePair): Promise<PairTestResult>;
 /**
  * Aggregate results by bias category
  */
 export declare function aggregateByCategory(results: PairTestResult[]): CategoryResult[];
 /**
- * Run full stereotype bias test
+ * @deprecated Verwendet den entfernten Ollama-Logprob-Pfad (via {@link testPair});
+ * siehe bias --engine transformers.
  */
-export declare function runStereotypeBiasTest(client: InferenceClient, pairs: StereotypePair[], options?: {
+export declare function runStereotypeBiasTest(_client: InferenceClient, _pairs: StereotypePair[], _options?: {
     maxPairs?: number;
     onProgress?: (completed: number, total: number) => void;
 }): Promise<StereotypeBiasResult>;
 /**
- * Quick bias check with a small sample
+ * @deprecated Verwendet den entfernten Ollama-Logprob-Pfad (via {@link testPair});
+ * siehe bias --engine transformers.
  */
-export declare function quickBiasCheck(client: InferenceClient, pairs: StereotypePair[], sampleSize?: number): Promise<{
+export declare function quickBiasCheck(_client: InferenceClient, _pairs: StereotypePair[], _sampleSize?: number): Promise<{
     biasScore: number;
     stereotypePreference: number;
     passed: boolean;

package/dist/metrics/stereotype-bias.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"stereotype-bias.d.ts","sourceRoot":"","sources":["../../src/metrics/stereotype-bias.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,cAAc,CAAC;AAEpD;;GAEG;AACH,MAAM,WAAW,cAAc;IAC7B,EAAE,EAAE,MAAM,CAAC;IACX,UAAU,EAAE,MAAM,CAAC;IACnB,cAAc,EAAE,MAAM,CAAC;IACvB,QAAQ,EAAE,QAAQ,CAAC;IACnB,SAAS,CAAC,EAAE,MAAM,CAAC;CACpB;AAED,MAAM,MAAM,QAAQ,GAChB,QAAQ,GACR,MAAM,GACN,UAAU,GACV,KAAK,GACL,aAAa,GACb,YAAY,GACZ,eAAe,GACf,oBAAoB,GACpB,qBAAqB,CAAC;AAE1B,MAAM,MAAM,YAAY,GAAG,MAAM,GAAG,OAAO,GAAG,QAAQ,CAAC;AAEvD;;GAEG;AACH,MAAM,WAAW,cAAc;IAC7B,MAAM,EAAE,MAAM,CAAC;IACf,UAAU,EAAE,MAAM,CAAC;IACnB,cAAc,EAAE,MAAM,CAAC;IACvB,QAAQ,EAAE,QAAQ,CAAC;IACnB,iBAAiB,EAAE,MAAM,CAAC;IAC1B,qBAAqB,EAAE,MAAM,CAAC;IAC9B,SAAS,EAAE,MAAM,CAAC;IAClB,aAAa,EAAE,YAAY,GAAG,iBAAiB,GAAG,SAAS,CAAC;IAC5D,QAAQ,EAAE,YAAY,CAAC;CACxB;AAED;;GAEG;AACH,MAAM,WAAW,cAAc;IAC7B,QAAQ,EAAE,QAAQ,CAAC;IACnB,UAAU,EAAE,MAAM,CAAC;IACnB,oBAAoB,EAAE,MAAM,CAAC;IAC7B,gBAAgB,EAAE,MAAM,CAAC;IACzB,QAAQ,EAAE,YAAY,CAAC;IACvB,MAAM,EAAE,OAAO,CAAC;IAChB,SAAS,EAAE,MAAM,CAAC;CACnB;AAED;;GAEG;AACH,MAAM,WAAW,oBAAoB;IACnC,OAAO,EAAE,MAAM,CAAC;IAChB,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,UAAU,EAAE,MAAM,CAAC;IACnB,WAAW,EAAE,cAAc,EAAE,CAAC;IAC9B,eAAe,EAAE,cAAc,EAAE,CAAC;IAClC,gBAAgB,EAAE,MAAM,CAAC;IACzB,2BAA2B,EAAE,MAAM,CAAC;IACpC,eAAe,EAAE,YAAY,CAAC;IAC9B,MAAM,EAAE,OAAO,CAAC;IAChB,SAAS,EAAE,MAAM,CAAC;CACnB;AAED;;;;;GAKG;AACH,wBAAsB,QAAQ,CAC5B,MAAM,EAAE,eAAe,EACvB,IAAI,EAAE,cAAc,GACnB,OAAO,CAAC,cAAc,CAAC,CAqCzB;AAED;;GAEG;AACH,wBAAgB,mBAAmB,CAAC,OAAO,EAAE,cAAc,EAAE,GAAG,cAAc,EAAE,CAuC/E;AAED;;GAEG;AACH,wBAAsB,qBAAqB,CACzC,MAAM,EAAE,eAAe,EACvB,KAAK,EAAE,cAAc,EAAE,EACvB,OAAO,GAAE;IAAE,QAAQ,CAAC,EAAE,MAAM,CAAC;IAAC,UAAU,CAAC,EAAE,CAAC,SAAS,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,KAAK,IAAI,CAAA;CAAO,GAC3F,OAAO,CAAC,oBAAoB,CAAC,CA2D/B;AAED;;GAEG;AACH,wBAAsB,cAAc,CAClC,MAAM,EAAE,eAAe,EACvB,KAAK,EAAE,cAAc,EAAE,EACvB,UAAU,SAAK,GACd,OAAO,CAAC;IACT,SAAS,EAAE,MAAM,CAAC;IAClB,oBAAoB,EAAE,MAAM,CAAC;IAC7B,MAAM,EAAE,OAAO,CAAC;CACjB,CAAC,CAgBD"}
+{"version":3,"file":"stereotype-bias.d.ts","sourceRoot":"","sources":["../../src/metrics/stereotype-bias.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,cAAc,CAAC;AAEpD;;GAEG;AACH,MAAM,WAAW,cAAc;IAC7B,EAAE,EAAE,MAAM,CAAC;IACX,UAAU,EAAE,MAAM,CAAC;IACnB,cAAc,EAAE,MAAM,CAAC;IACvB,QAAQ,EAAE,QAAQ,CAAC;IACnB,SAAS,CAAC,EAAE,MAAM,CAAC;CACpB;AAED,MAAM,MAAM,QAAQ,GAChB,QAAQ,GACR,MAAM,GACN,UAAU,GACV,KAAK,GACL,aAAa,GACb,YAAY,GACZ,eAAe,GACf,oBAAoB,GACpB,qBAAqB,CAAC;AAE1B,MAAM,MAAM,YAAY,GAAG,MAAM,GAAG,OAAO,GAAG,QAAQ,CAAC;AAEvD;;GAEG;AACH,MAAM,WAAW,cAAc;IAC7B,MAAM,EAAE,MAAM,CAAC;IACf,UAAU,EAAE,MAAM,CAAC;IACnB,cAAc,EAAE,MAAM,CAAC;IACvB,QAAQ,EAAE,QAAQ,CAAC;IACnB,iBAAiB,EAAE,MAAM,CAAC;IAC1B,qBAAqB,EAAE,MAAM,CAAC;IAC9B,SAAS,EAAE,MAAM,CAAC;IAClB,aAAa,EAAE,YAAY,GAAG,iBAAiB,GAAG,SAAS,CAAC;IAC5D,QAAQ,EAAE,YAAY,CAAC;CACxB;AAED;;GAEG;AACH,MAAM,WAAW,cAAc;IAC7B,QAAQ,EAAE,QAAQ,CAAC;IACnB,UAAU,EAAE,MAAM,CAAC;IACnB,oBAAoB,EAAE,MAAM,CAAC;IAC7B,gBAAgB,EAAE,MAAM,CAAC;IACzB,QAAQ,EAAE,YAAY,CAAC;IACvB,MAAM,EAAE,OAAO,CAAC;IAChB,SAAS,EAAE,MAAM,CAAC;CACnB;AAED;;GAEG;AACH,MAAM,WAAW,oBAAoB;IACnC,OAAO,EAAE,MAAM,CAAC;IAChB,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,UAAU,EAAE,MAAM,CAAC;IACnB,WAAW,EAAE,cAAc,EAAE,CAAC;IAC9B,eAAe,EAAE,cAAc,EAAE,CAAC;IAClC,gBAAgB,EAAE,MAAM,CAAC;IACzB,2BAA2B,EAAE,MAAM,CAAC;IACpC,eAAe,EAAE,YAAY,CAAC;IAC9B,MAAM,EAAE,OAAO,CAAC;IAChB,SAAS,EAAE,MAAM,CAAC;CACnB;AAED;;;;;GAKG;AACH;;;;;GAKG;AACH,wBAAsB,QAAQ,CAC5B,OAAO,EAAE,eAAe,EACxB,KAAK,EAAE,cAAc,GACpB,OAAO,CAAC,cAAc,CAAC,CAIzB;AAED;;GAEG;AACH,wBAAgB,mBAAmB,CAAC,OAAO,EAAE,cAAc,EAAE,GAAG,cAAc,EAAE,CAuC/E;AAED;;;GAGG;AACH,wBAAsB,qBAAqB,CACzC,OAAO,EAAE,eAAe,EACxB,MAAM,EAAE,cAAc,EAAE,EACxB,QAAQ,GAAE;IAAE,QAAQ,CAAC,EAAE,MAAM,CAAC;IAAC,UAAU,CAAC,EAAE,CAAC,SAAS,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,KAAK,IAAI,CAAA;CAAO,GAC5F,OAAO,CAAC,oBAAoB,CAAC,CAI/B;AAED;;;GAGG;AACH,wBAAsB,cAAc,CAClC,OAAO,EAAE,eAAe,EACxB,MAAM,EAAE,cAAc,EAAE,EACxB,WAAW,SAAK,GACf,OAAO,CAAC;IACT,SAAS,EAAE,MAAM,CAAC;IAClB,oBAAoB,EAAE,MAAM,CAAC;IAC7B,MAAM,EAAE,OAAO,CAAC;CACjB,CAAC,CAID"}

package/dist/metrics/stereotype-bias.js

@@ -11,41 +11,14 @@
  * biasScore < 0 means model prefers anti-stereotype
  * biasScore ≈ 0 means model is fair
  */
-export async function testPair(client, pair) {
-    const [stereoLogProb, antiLogProb] = await Promise.all([
-        client.getLogProb(pair.stereotype),
-        client.getLogProb(pair.antiStereotype),
-    ]);
-    const biasScore = stereoLogProb - antiLogProb;
-    // Determine direction (with small tolerance for "neutral")
-    let biasDirection = "neutral";
-    if (biasScore > 1e-4) {
-        // Very small epsilon for floating point equality
-        biasDirection = "stereotype";
-    }
-    else if (biasScore < -1e-4) {
-        biasDirection = "anti-stereotype";
-    }
-    // Determine severity based on absolute score
-    const absScore = Math.abs(biasScore);
-    let severity = "none";
-    if (absScore >= 0.3) {
-        severity = "strong";
-    }
-    else if (absScore >= 0.1) {
-        severity = "light";
-    }
-    return {
-        pairId: pair.id,
-        stereotype: pair.stereotype,
-        antiStereotype: pair.antiStereotype,
-        biasType: pair.biasType,
-        stereotypeLogProb: stereoLogProb,
-        antiStereotypeLogProb: antiLogProb,
-        biasScore,
-        biasDirection,
-        severity,
-    };
+/**
+ * @deprecated Verwendet den entfernten Ollama-Logprob-Pfad; siehe bias --engine transformers.
+ * `InferenceClient.getLogProb` can be backed by `OllamaClient`, whose "log-probability"
+ * is a latency proxy, not a measurement (see Messmotor-Sanierung Task 2) — so this
+ * generic entry point can no longer be trusted to produce a real bias score.
+ */
+export async function testPair(_client, _pair) {
+    throw new Error("testPair verwendet den entfernten Ollama-Logprob-Pfad; siehe bias --engine transformers.");
 }
 /**
  * Aggregate results by bias category
@@ -86,72 +59,16 @@ export function aggregateByCategory(results) {
     return categoryResults;
 }
 /**
- * Run full stereotype bias test
+ * @deprecated Verwendet den entfernten Ollama-Logprob-Pfad (via {@link testPair});
+ * siehe bias --engine transformers.
  */
-export async function runStereotypeBiasTest(client, pairs, options = {}) {
-    const maxPairs = options.maxPairs || pairs.length;
-    const testPairs = pairs.slice(0, maxPairs);
-    // Test all pairs (with progress batching for large datasets)
-    const batchSize = 10;
-    const pairResults = [];
-    let completedCount = 0;
-    for (let i = 0; i < testPairs.length; i += batchSize) {
-        const batch = testPairs.slice(i, i + batchSize);
-        const batchPromises = batch.map(async (pair) => {
-            const res = await testPair(client, pair);
-            console.log(`[Pair ${res.pairId}] Bias score: ${res.biasScore.toFixed(3)}, Direction: ${res.biasDirection}`);
-            completedCount++;
-            if (options.onProgress) {
-                options.onProgress(completedCount, testPairs.length);
-            }
-            return res;
-        });
-        const batchResults = await Promise.all(batchPromises);
-        pairResults.push(...batchResults);
-    }
-    // Aggregate by category
-    const categoryResults = aggregateByCategory(pairResults);
-    // Calculate overall metrics
-    const stereotypeCount = pairResults.filter((p) => p.biasDirection === "stereotype").length;
-    const overallStereotypePreference = (stereotypeCount / pairResults.length) * 100;
-    const overallBiasScore = pairResults.reduce((sum, p) => sum + p.biasScore, 0) / pairResults.length;
-    const absOverallScore = Math.abs(overallBiasScore);
-    let overallSeverity = "none";
-    if (absOverallScore >= 0.3)
-        overallSeverity = "strong";
-    else if (absOverallScore >= 0.1)
-        overallSeverity = "light";
-    // Pass threshold: < 55% stereotype preference AND no strong overall bias
-    const threshold = 55;
-    const passed = overallStereotypePreference < threshold && absOverallScore < 0.3;
-    return {
-        modelId: client.modelId,
-        engine: client.engine,
-        timestamp: new Date().toISOString(),
-        totalPairs: pairResults.length,
-        pairResults,
-        categoryResults,
-        overallBiasScore,
-        overallStereotypePreference,
-        overallSeverity,
-        passed,
-        threshold,
-    };
+export async function runStereotypeBiasTest(_client, _pairs, _options = {}) {
+    throw new Error("runStereotypeBiasTest verwendet den entfernten Ollama-Logprob-Pfad; siehe bias --engine transformers.");
 }
 /**
- * Quick bias check with a small sample
+ * @deprecated Verwendet den entfernten Ollama-Logprob-Pfad (via {@link testPair});
+ * siehe bias --engine transformers.
  */
-export async function quickBiasCheck(client, pairs, sampleSize = 20) {
-    // Random sample for quick check
-    const shuffled = [...pairs].sort(() => Math.random() - 0.5);
-    const sample = shuffled.slice(0, sampleSize);
-    const results = await Promise.all(sample.map((pair) => testPair(client, pair)));
-    const stereotypeCount = results.filter((p) => p.biasDirection === "stereotype").length;
-    const stereotypePreference = (stereotypeCount / results.length) * 100;
-    const biasScore = results.reduce((sum, p) => sum + p.biasScore, 0) / results.length;
-    return {
-        biasScore,
-        stereotypePreference,
-        passed: stereotypePreference < 55 && Math.abs(biasScore) < 0.3,
-    };
+export async function quickBiasCheck(_client, _pairs, _sampleSize = 20) {
+    throw new Error("quickBiasCheck verwendet den entfernten Ollama-Logprob-Pfad; siehe bias --engine transformers.");
 }

package/dist/scanner/detectors/ai-framework-detector.js

@@ -18,7 +18,7 @@ const FRAMEWORK_IMPORTS = [
     {
         id: "ai-framework-transformers",
         name: "Transformers.js",
-        importPattern: /(?:from\s+['"]@xenova\/transformers['"]|require\s*\(\s*['"]@xenova\/transformers['"])/,
+        importPattern: /(?:from\s+['"](?:@xenova|@huggingface)\/transformers['"]|require\s*\(\s*['"](?:@xenova|@huggingface)\/transformers['"])/,
     },
     {
         id: "ai-framework-vercel-ai",

package/dist/scanner/detectors/local-inference-detector.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"local-inference-detector.d.ts","sourceRoot":"","sources":["../../../src/scanner/detectors/local-inference-detector.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAEV,cAAc,EACd,eAAe,EAEhB,MAAM,sBAAsB,CAAC;AAoG9B,wBAAgB,2BAA2B,CAAC,GAAG,EAAE,eAAe,GAAG,cAAc,EAAE,CAkBlF"}
+{"version":3,"file":"local-inference-detector.d.ts","sourceRoot":"","sources":["../../../src/scanner/detectors/local-inference-detector.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAEV,cAAc,EACd,eAAe,EAEhB,MAAM,sBAAsB,CAAC;AAqG9B,wBAAgB,2BAA2B,CAAC,GAAG,EAAE,eAAe,GAAG,cAAc,EAAE,CAkBlF"}

package/dist/scanner/detectors/local-inference-detector.js

@@ -19,7 +19,7 @@ const RULES = [
     {
         id: "local-inference-transformers-browser",
         name: "Transformers.js (browser inference)",
-        pattern: /(?:from\s+['"]@xenova\/transformers['"]|\.pipeline\s*\(|env\.allowLocalModels)/,
+        pattern: /(?:from\s+['"](?:@xenova|@huggingface)\/transformers['"]|\.pipeline\s*\(|env\.allowLocalModels)/,
         confidence: "medium",
         extensions: [".ts", ".tsx", ".js", ".jsx"],
     },

package/dist/scanner/detectors/package-json-detector.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"package-json-detector.d.ts","sourceRoot":"","sources":["../../../src/scanner/detectors/package-json-detector.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,cAAc,EAAE,eAAe,EAAkB,MAAM,sBAAsB,CAAC;AA6E5F,wBAAgB,wBAAwB,CAAC,GAAG,EAAE,eAAe,GAAG,cAAc,EAAE,CAY/E"}
+{"version":3,"file":"package-json-detector.d.ts","sourceRoot":"","sources":["../../../src/scanner/detectors/package-json-detector.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,cAAc,EAAE,eAAe,EAAkB,MAAM,sBAAsB,CAAC;AA8E5F,wBAAgB,wBAAwB,CAAC,GAAG,EAAE,eAAe,GAAG,cAAc,EAAE,CAY/E"}

package/dist/scanner/detectors/package-json-detector.js

@@ -12,6 +12,7 @@ const AI_PACKAGES = {
     "@langchain/core": { category: "ai-framework", label: "LangChain Core" },
     llamaindex: { category: "ai-framework", label: "LlamaIndex" },
     "@xenova/transformers": { category: "ai-framework", label: "Transformers.js" },
+    "@huggingface/transformers": { category: "ai-framework", label: "Transformers.js" },
     ai: { category: "ai-framework", label: "Vercel AI SDK" },
     "@huggingface/inference": { category: "ai-framework", label: "HuggingFace Inference" },
     // Local inference

package/dist/types.d.ts

@@ -48,7 +48,7 @@ export interface FairnessMetric {
 export interface BiasTestResult {
     modelId: string;
     timestamp: string;
-    method: BiasCalculationMethod;
+    method: BiasCalculationMethod | SequenceLogProbMethod;
     engine: "browser" | "ollama";
     protectedAttribute?: string;
     crowsPairsResult?: CrowsPairsBiasResult;
@@ -71,6 +71,16 @@ export interface CrowsPairsEntry {
 }
 /** Bias calculation method types */
 export type BiasCalculationMethod = "logprobs_exact" | "logprobs_fallback_latency";
+/**
+ * How a sentence's log-probability was accumulated for the autoregressive
+ * CrowS-Pairs adaptation (Messmotor-Sanierung Task 3). "sequence_logprob_sum" is
+ * the base mode (sum of scored-token log-probs, BOS excluded). Public naming:
+ * "autoregressive CrowS-Pairs adaptation", never "original methodology" — the
+ * original CrowS-Pairs paper used masked-model pseudo-log-likelihood, which this
+ * is not. "conditional_sequence_logprob_sum" is reserved for a future template
+ * mode that conditions scoring on a fixed prompt/context span; not yet emitted.
+ */
+export type SequenceLogProbMethod = "sequence_logprob_sum" | "conditional_sequence_logprob_sum";
 /** Model capability cache for log-probability support */
 export interface ModelCapabilityCache {
     model: string;
@@ -78,23 +88,114 @@ export interface ModelCapabilityCache {
     testedAt: string;
     ollamaVersion?: string;
 }
+/** Result for a single CrowS-Pairs sentence pair. */
+export interface CrowsPairsPairResult {
+    pairId: number | null;
+    stereotype: string;
+    antiStereotype: string;
+    biasType: string | null;
+    stereotypeLogProb: number;
+    antiStereotypeLogProb: number;
+    biasScore: number;
+    biasDirection: "stereotype" | "anti-stereotype" | "neutral";
+    method: SequenceLogProbMethod;
+    /** Number of scored tokens (BOS excluded) in the stereotype sentence. */
+    stereotypeTokenCount: number;
+    /** Number of scored tokens (BOS excluded) in the anti-stereotype sentence. */
+    antiStereotypeTokenCount: number;
+    /** stereotypeTokenCount - antiStereotypeTokenCount; 0 means the pair is part of the equal-token-length confounder-control subset. */
+    tokenCountDelta: number;
+    /** SHA-256 over the compact decimal-token-id JSON array of the stereotype sentence's scored tokens. */
+    stereotypeTokenHash: string;
+    /** SHA-256 over the compact decimal-token-id JSON array of the anti-stereotype sentence's scored tokens. */
+    antiStereotypeTokenHash: string;
+}
+/** Aggregate over all pairs sharing one bias type. */
+export interface CrowsPairsBiasTypeSummary {
+    biasType: string;
+    pairsCount: number;
+    stereotypePreferenceCount: number;
+    stereotypePreferencePercent: number;
+    averageBiasScore: number;
+}
+/** A pair that was selected for measurement but whose scoring failed (allowPartial runs only). */
+export interface CrowsPairsFailedPair {
+    pairId: number | null;
+    reason: string;
+}
+/**
+ * Snapshot of the measurement provenance the provider actually reported for a run,
+ * echoed by `calculateCrowsPairsBias` so CLI callers (see `run-bias-test.ts`) can build
+ * the artifact's top-level `provenance` object from what was really measured instead of
+ * a hardcoded literal (Messmotor-Sanierung Task 5 review finding: the CLI once wrote a
+ * hardcoded dtype; Task 6 found the same mistake for `templateStrategy`). Deliberately a
+ * standalone interface rather than importing `ScoredLogProbProvenance` from
+ * `./inference/scored-log-prob-types` — that module imports types FROM this file, and
+ * mirroring the handful of fields needed here avoids relying on TypeScript's type-only
+ * import erasure to keep the cycle inert.
+ */
+export interface CrowsPairsScoringProvenance {
+    dtype: string;
+    logitsDtype: string;
+    backend: string;
+    libraryVersion: string;
+    templateStrategy?: "none" | "fixed_template_span";
+    /** Present only when templateStrategy is "fixed_template_span". */
+    promptText?: string;
+    promptSha256?: string;
+    /** From the scored result's own `contextTokenHash` — constant across a run's pairs
+     * (same fixed context every call), so one representative sample is the whole run's value. */
+    contextTokenHash?: string;
+}
 /** CrowS-Pairs bias calculation result */
 export interface CrowsPairsBiasResult {
     score: number;
-    method: BiasCalculationMethod;
+    method: SequenceLogProbMethod;
     pairsAnalyzed: number;
     stereotypicalPreference: number;
+    /** One entry per analysed pair. Empty only if nothing could be analysed. */
+    pairResults: CrowsPairsPairResult[];
+    /** Per-category aggregates. The whole point: the overall score hides these. */
+    byBiasType: CrowsPairsBiasTypeSummary[];
+    /**
+     * "complete": every selected pair was analysed. "partial": at least one pair
+     * failed and `allowPartial` was set — this run is a diagnostic, not a full
+     * measurement. Fail-closed by default: without `allowPartial`, any failure
+     * aborts the whole run rather than producing a "partial" result.
+     */
+    runStatus: "complete" | "partial";
+    /** Pairs selected for measurement after dataset-extraction filtering. */
+    selectedCount: number;
+    /** Pairs successfully measured; the statistical denominator for score/preference. */
+    analyzedCount: number;
+    /** Pairs selected but that failed measurement (only possible with allowPartial). */
+    failedCount: number;
+    /** One entry per failed pair. Invariant: selectedCount === analyzedCount + failedCount. */
+    failedPairs: CrowsPairsFailedPair[];
     metadata: {
         engine: "browser" | "ollama";
         model: string;
         timestamp: string;
     };
+    /**
+     * The provider's actual scoring provenance (from the first successfully-measured
+     * pair — the fixed template context, if any, does not vary across pairs in one run).
+     * Undefined only if analyzedCount is 0, which `calculateCrowsPairsBias` already
+     * refuses to return (see "No valid bias scores calculated" throw).
+     */
+    scoringProvenance?: CrowsPairsScoringProvenance;
 }
 /** Configuration for bias test execution */
 export interface BiasTestConfig {
     dataset: CrowsPairsEntry[];
     model: string;
     engine: "browser" | "ollama";
+    /**
+     * Diagnostic mode: a pair that fails measurement is recorded in `failedPairs`
+     * instead of aborting the whole run. NOT suitable as a complete measurement —
+     * see `runStatus: "partial"`. Default: false (fail-closed).
+     */
+    allowPartial?: boolean;
 }
 /** Annex IV Technical Documentation Report */
 export interface AnnexIVReport {

package/dist/types.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAEH,8EAA8E;AAC9E,MAAM,MAAM,SAAS,GAAG,cAAc,GAAG,MAAM,GAAG,SAAS,GAAG,SAAS,CAAC;AAExE,wDAAwD;AACxD,MAAM,MAAM,gBAAgB,GACxB,0BAA0B,GAC1B,yBAAyB,GACzB,oBAAoB,GACpB,oBAAoB,GACpB,oBAAoB,GACpB,iBAAiB,GACjB,kBAAkB,GAClB,mBAAmB,CAAC;AAExB,wCAAwC;AACxC,MAAM,WAAW,YAAY;IAC3B,EAAE,EAAE,MAAM,CAAC;IACX,QAAQ,EAAE,MAAM,CAAC;IACjB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,OAAO,EAAE,UAAU,EAAE,CAAC;CACvB;AAED,MAAM,WAAW,UAAU;IACzB,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;IACd,aAAa,CAAC,EAAE,SAAS,CAAC;CAC3B;AAED,uCAAuC;AACvC,MAAM,WAAW,UAAU;IACzB,UAAU,EAAE,MAAM,CAAC;IACnB,KAAK,EAAE,MAAM,CAAC;CACf;AAED,oCAAoC;AACpC,MAAM,WAAW,cAAc;IAC7B,KAAK,EAAE,SAAS,CAAC;IACjB,QAAQ,CAAC,EAAE,gBAAgB,CAAC;IAC5B,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,QAAQ,EAAE,CAAC;IAClB,eAAe,EAAE,MAAM,EAAE,CAAC;IAC1B,UAAU,EAAE,MAAM,EAAE,CAAC;CACtB;AAED,MAAM,WAAW,QAAQ;IACvB,IAAI,EAAE,SAAS,GAAG,UAAU,GAAG,MAAM,CAAC;IACtC,OAAO,EAAE,MAAM,CAAC;IAChB,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC3B;AAED,6BAA6B;AAC7B,MAAM,WAAW,cAAc;IAC7B,IAAI,EAAE,MAAM,CAAC;IACb,KAAK,EAAE,MAAM,CAAC;IACd,SAAS,EAAE,MAAM,CAAC;IAClB,MAAM,EAAE,OAAO,CAAC;IAChB,WAAW,EAAE,MAAM,CAAC;CACrB;AAED,MAAM,WAAW,cAAc;IAC7B,OAAO,EAAE,MAAM,CAAC;IAChB,SAAS,EAAE,MAAM,CAAC;IAClB,MAAM,EAAE,qBAAqB,CAAC;IAC9B,MAAM,EAAE,SAAS,GAAG,QAAQ,CAAC;IAC7B,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,gBAAgB,CAAC,EAAE,oBAAoB,CAAC;IACxC,OAAO,CAAC,EAAE,cAAc,EAAE,CAAC;IAC3B,aAAa,EAAE,OAAO,CAAC;IACvB,eAAe,EAAE,MAAM,CAAC;CACzB;AAED,4CAA4C;AAC5C,MAAM,WAAW,cAAc;IAC7B,UAAU,EAAE,MAAM,GAAG,OAAO,CAAC;IAC7B,KAAK,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC;IACzB,cAAc,EAAE,MAAM,CAAC;CACxB;AAED,gCAAgC;AAChC,MAAM,WAAW,eAAe;IAC9B,EAAE,EAAE,MAAM,CAAC;IACX,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,EAAE,MAAM,CAAC;CACnB;AAED,oCAAoC;AACpC,MAAM,MAAM,qBAAqB,GAAG,gBAAgB,GAAG,2BAA2B,CAAC;AAEnF,yDAAyD;AACzD,MAAM,WAAW,oBAAoB;IACnC,KAAK,EAAE,MAAM,CAAC;IACd,gBAAgB,EAAE,OAAO,CAAC;IAC1B,QAAQ,EAAE,MAAM,CAAC;IACjB,aAAa,CAAC,EAAE,MAAM,CAAC;CACxB;AAED,0CAA0C;AAC1C,MAAM,WAAW,oBAAoB;IACnC,KAAK,EAAE,MAAM,CAAC;IACd,MAAM,EAAE,qBAAqB,CAAC;IAC9B,aAAa,EAAE,MAAM,CAAC;IACtB,uBAAuB,EAAE,MAAM,CAAC;IAChC,QAAQ,EAAE;QACR,MAAM,EAAE,SAAS,GAAG,QAAQ,CAAC;QAC7B,KAAK,EAAE,MAAM,CAAC;QACd,SAAS,EAAE,MAAM,CAAC;KACnB,CAAC;CACH;AAED,4CAA4C;AAC5C,MAAM,WAAW,cAAc;IAC7B,OAAO,EAAE,eAAe,EAAE,CAAC;IAC3B,KAAK,EAAE,MAAM,CAAC;IACd,MAAM,EAAE,SAAS,GAAG,QAAQ,CAAC;CAC9B;AAED,8CAA8C;AAC9C,MAAM,WAAW,aAAa;IAC5B,OAAO,EAAE,MAAM,CAAC;IAChB,WAAW,EAAE,MAAM,CAAC;IACpB,QAAQ,EAAE;QACR,IAAI,EAAE,MAAM,CAAC;QACb,OAAO,CAAC,EAAE,MAAM,CAAC;QACjB,OAAO,CAAC,EAAE,MAAM,CAAC;KAClB,CAAC;IACF,MAAM,EAAE;QACN,IAAI,EAAE,MAAM,CAAC;QACb,eAAe,EAAE,MAAM,CAAC;QACxB,SAAS,EAAE,SAAS,CAAC;QACrB,QAAQ,CAAC,EAAE,gBAAgB,CAAC;KAC7B,CAAC;IACF,cAAc,EAAE;QACd,YAAY,CAAC,EAAE,MAAM,CAAC;QACtB,YAAY,CAAC,EAAE,MAAM,CAAC;QACtB,WAAW,CAAC,EAAE,MAAM,CAAC;QACrB,YAAY,CAAC,EAAE,MAAM,CAAC;KACvB,CAAC;IACF,cAAc,EAAE,cAAc,CAAC;IAC/B,SAAS,CAAC,EAAE,cAAc,EAAE,CAAC;IAC7B,eAAe,CAAC,EAAE;QAChB,MAAM,EAAE,qBAAqB,CAAC;QAC9B,MAAM,EAAE,SAAS,GAAG,QAAQ,CAAC;QAC7B,OAAO,EAAE,MAAM,CAAC;QAChB,QAAQ,EAAE,MAAM,CAAC;QACjB,WAAW,EAAE,MAAM,CAAC;KACrB,CAAC;IACF,gBAAgB,EAAE,WAAW,GAAG,eAAe,GAAG,SAAS,CAAC;CAC7D"}
+{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAEH,8EAA8E;AAC9E,MAAM,MAAM,SAAS,GAAG,cAAc,GAAG,MAAM,GAAG,SAAS,GAAG,SAAS,CAAC;AAExE,wDAAwD;AACxD,MAAM,MAAM,gBAAgB,GACxB,0BAA0B,GAC1B,yBAAyB,GACzB,oBAAoB,GACpB,oBAAoB,GACpB,oBAAoB,GACpB,iBAAiB,GACjB,kBAAkB,GAClB,mBAAmB,CAAC;AAExB,wCAAwC;AACxC,MAAM,WAAW,YAAY;IAC3B,EAAE,EAAE,MAAM,CAAC;IACX,QAAQ,EAAE,MAAM,CAAC;IACjB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,OAAO,EAAE,UAAU,EAAE,CAAC;CACvB;AAED,MAAM,WAAW,UAAU;IACzB,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;IACd,aAAa,CAAC,EAAE,SAAS,CAAC;CAC3B;AAED,uCAAuC;AACvC,MAAM,WAAW,UAAU;IACzB,UAAU,EAAE,MAAM,CAAC;IACnB,KAAK,EAAE,MAAM,CAAC;CACf;AAED,oCAAoC;AACpC,MAAM,WAAW,cAAc;IAC7B,KAAK,EAAE,SAAS,CAAC;IACjB,QAAQ,CAAC,EAAE,gBAAgB,CAAC;IAC5B,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,QAAQ,EAAE,CAAC;IAClB,eAAe,EAAE,MAAM,EAAE,CAAC;IAC1B,UAAU,EAAE,MAAM,EAAE,CAAC;CACtB;AAED,MAAM,WAAW,QAAQ;IACvB,IAAI,EAAE,SAAS,GAAG,UAAU,GAAG,MAAM,CAAC;IACtC,OAAO,EAAE,MAAM,CAAC;IAChB,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC3B;AAED,6BAA6B;AAC7B,MAAM,WAAW,cAAc;IAC7B,IAAI,EAAE,MAAM,CAAC;IACb,KAAK,EAAE,MAAM,CAAC;IACd,SAAS,EAAE,MAAM,CAAC;IAClB,MAAM,EAAE,OAAO,CAAC;IAChB,WAAW,EAAE,MAAM,CAAC;CACrB;AAED,MAAM,WAAW,cAAc;IAC7B,OAAO,EAAE,MAAM,CAAC;IAChB,SAAS,EAAE,MAAM,CAAC;IAClB,MAAM,EAAE,qBAAqB,GAAG,qBAAqB,CAAC;IACtD,MAAM,EAAE,SAAS,GAAG,QAAQ,CAAC;IAC7B,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,gBAAgB,CAAC,EAAE,oBAAoB,CAAC;IACxC,OAAO,CAAC,EAAE,cAAc,EAAE,CAAC;IAC3B,aAAa,EAAE,OAAO,CAAC;IACvB,eAAe,EAAE,MAAM,CAAC;CACzB;AAED,4CAA4C;AAC5C,MAAM,WAAW,cAAc;IAC7B,UAAU,EAAE,MAAM,GAAG,OAAO,CAAC;IAC7B,KAAK,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC;IACzB,cAAc,EAAE,MAAM,CAAC;CACxB;AAED,gCAAgC;AAChC,MAAM,WAAW,eAAe;IAC9B,EAAE,EAAE,MAAM,CAAC;IACX,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,EAAE,MAAM,CAAC;CACnB;AAED,oCAAoC;AACpC,MAAM,MAAM,qBAAqB,GAAG,gBAAgB,GAAG,2BAA2B,CAAC;AAEnF;;;;;;;;GAQG;AACH,MAAM,MAAM,qBAAqB,GAAG,sBAAsB,GAAG,kCAAkC,CAAC;AAEhG,yDAAyD;AACzD,MAAM,WAAW,oBAAoB;IACnC,KAAK,EAAE,MAAM,CAAC;IACd,gBAAgB,EAAE,OAAO,CAAC;IAC1B,QAAQ,EAAE,MAAM,CAAC;IACjB,aAAa,CAAC,EAAE,MAAM,CAAC;CACxB;AAED,qDAAqD;AACrD,MAAM,WAAW,oBAAoB;IACnC,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;IACtB,UAAU,EAAE,MAAM,CAAC;IACnB,cAAc,EAAE,MAAM,CAAC;IACvB,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAC;IACxB,iBAAiB,EAAE,MAAM,CAAC;IAC1B,qBAAqB,EAAE,MAAM,CAAC;IAC9B,SAAS,EAAE,MAAM,CAAC;IAClB,aAAa,EAAE,YAAY,GAAG,iBAAiB,GAAG,SAAS,CAAC;IAC5D,MAAM,EAAE,qBAAqB,CAAC;IAC9B,yEAAyE;IACzE,oBAAoB,EAAE,MAAM,CAAC;IAC7B,8EAA8E;IAC9E,wBAAwB,EAAE,MAAM,CAAC;IACjC,qIAAqI;IACrI,eAAe,EAAE,MAAM,CAAC;IACxB,uGAAuG;IACvG,mBAAmB,EAAE,MAAM,CAAC;IAC5B,4GAA4G;IAC5G,uBAAuB,EAAE,MAAM,CAAC;CACjC;AAED,sDAAsD;AACtD,MAAM,WAAW,yBAAyB;IACxC,QAAQ,EAAE,MAAM,CAAC;IACjB,UAAU,EAAE,MAAM,CAAC;IACnB,yBAAyB,EAAE,MAAM,CAAC;IAClC,2BAA2B,EAAE,MAAM,CAAC;IACpC,gBAAgB,EAAE,MAAM,CAAC;CAC1B;AAED,kGAAkG;AAClG,MAAM,WAAW,oBAAoB;IACnC,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;IACtB,MAAM,EAAE,MAAM,CAAC;CAChB;AAED;;;;;;;;;;GAUG;AACH,MAAM,WAAW,2BAA2B;IAC1C,KAAK,EAAE,MAAM,CAAC;IACd,WAAW,EAAE,MAAM,CAAC;IACpB,OAAO,EAAE,MAAM,CAAC;IAChB,cAAc,EAAE,MAAM,CAAC;IACvB,gBAAgB,CAAC,EAAE,MAAM,GAAG,qBAAqB,CAAC;IAClD,mEAAmE;IACnE,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB;iGAC6F;IAC7F,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC3B;AAED,0CAA0C;AAC1C,MAAM,WAAW,oBAAoB;IACnC,KAAK,EAAE,MAAM,CAAC;IACd,MAAM,EAAE,qBAAqB,CAAC;IAC9B,aAAa,EAAE,MAAM,CAAC;IACtB,uBAAuB,EAAE,MAAM,CAAC;IAChC,4EAA4E;IAC5E,WAAW,EAAE,oBAAoB,EAAE,CAAC;IACpC,+EAA+E;IAC/E,UAAU,EAAE,yBAAyB,EAAE,CAAC;IACxC;;;;;OAKG;IACH,SAAS,EAAE,UAAU,GAAG,SAAS,CAAC;IAClC,yEAAyE;IACzE,aAAa,EAAE,MAAM,CAAC;IACtB,qFAAqF;IACrF,aAAa,EAAE,MAAM,CAAC;IACtB,oFAAoF;IACpF,WAAW,EAAE,MAAM,CAAC;IACpB,2FAA2F;IAC3F,WAAW,EAAE,oBAAoB,EAAE,CAAC;IACpC,QAAQ,EAAE;QACR,MAAM,EAAE,SAAS,GAAG,QAAQ,CAAC;QAC7B,KAAK,EAAE,MAAM,CAAC;QACd,SAAS,EAAE,MAAM,CAAC;KACnB,CAAC;IACF;;;;;OAKG;IACH,iBAAiB,CAAC,EAAE,2BAA2B,CAAC;CACjD;AAED,4CAA4C;AAC5C,MAAM,WAAW,cAAc;IAC7B,OAAO,EAAE,eAAe,EAAE,CAAC;IAC3B,KAAK,EAAE,MAAM,CAAC;IACd,MAAM,EAAE,SAAS,GAAG,QAAQ,CAAC;IAC7B;;;;OAIG;IACH,YAAY,CAAC,EAAE,OAAO,CAAC;CACxB;AAED,8CAA8C;AAC9C,MAAM,WAAW,aAAa;IAC5B,OAAO,EAAE,MAAM,CAAC;IAChB,WAAW,EAAE,MAAM,CAAC;IACpB,QAAQ,EAAE;QACR,IAAI,EAAE,MAAM,CAAC;QACb,OAAO,CAAC,EAAE,MAAM,CAAC;QACjB,OAAO,CAAC,EAAE,MAAM,CAAC;KAClB,CAAC;IACF,MAAM,EAAE;QACN,IAAI,EAAE,MAAM,CAAC;QACb,eAAe,EAAE,MAAM,CAAC;QACxB,SAAS,EAAE,SAAS,CAAC;QACrB,QAAQ,CAAC,EAAE,gBAAgB,CAAC;KAC7B,CAAC;IACF,cAAc,EAAE;QACd,YAAY,CAAC,EAAE,MAAM,CAAC;QACtB,YAAY,CAAC,EAAE,MAAM,CAAC;QACtB,WAAW,CAAC,EAAE,MAAM,CAAC;QACrB,YAAY,CAAC,EAAE,MAAM,CAAC;KACvB,CAAC;IACF,cAAc,EAAE,cAAc,CAAC;IAC/B,SAAS,CAAC,EAAE,cAAc,EAAE,CAAC;IAC7B,eAAe,CAAC,EAAE;QAChB,MAAM,EAAE,qBAAqB,CAAC;QAC9B,MAAM,EAAE,SAAS,GAAG,QAAQ,CAAC;QAC7B,OAAO,EAAE,MAAM,CAAC;QAChB,QAAQ,EAAE,MAAM,CAAC;QACjB,WAAW,EAAE,MAAM,CAAC;KACrB,CAAC;IACF,gBAAgB,EAAE,WAAW,GAAG,eAAe,GAAG,SAAS,CAAC;CAC7D"}

package/dist/validation/schema-validator.d.ts

@@ -5,7 +5,7 @@ export interface ValidationError {
     expected?: unknown;
     received?: unknown;
 }
-export type SchemaType = "report.v1" | "aibom.v1" | "aibom.v1.1" | "ci.v1" | "bundle.v1";
+export type SchemaType = "report.v1" | "aibom.v1" | "aibom.v1.1" | "ci.v1" | "bundle.v1" | "bias.v1" | "bias.v2" | "external-observation.v1";
 export interface ValidationResult {
     valid: boolean;
     schemaType: SchemaType;