@euconform/core 1.5.0 → 1.6.0

package/dist/inference/capability-detection.js

@@ -21,6 +21,11 @@ export class DefaultCapabilityDetectionService {
         try {
             // Import browser models dynamically
             const { BROWSER_MODELS } = await import("./browser-inference");
+            // These entries are declared from the static BROWSER_MODELS list, not
+            // measured — no model has actually been loaded/tested at this point.
+            // status: "unverified" (not "available") reflects that honestly; no
+            // lastTested since no test ran. A real per-model verification step is
+            // future work, not this task's scope.
             for (const model of BROWSER_MODELS) {
                 if (!model.supported)
                     continue;
@@ -28,9 +33,8 @@ export class DefaultCapabilityDetectionService {
                     modelId: model.id,
                     engine: "browser",
                     method: "logprobs_exact",
-                    status: "available",
+                    status: "unverified",
                     recommended: true,
-                    lastTested: new Date().toISOString(),
                 });
             }
         }
@@ -41,9 +45,8 @@ export class DefaultCapabilityDetectionService {
                 modelId: "Xenova/distilgpt2",
                 engine: "browser",
                 method: "logprobs_exact",
-                status: "available",
+                status: "unverified",
                 recommended: true,
-                lastTested: new Date().toISOString(),
             });
         }
         try {
@@ -95,15 +98,14 @@ export class DefaultCapabilityDetectionService {
         if (cached && this.isCacheValid(cached)) {
             return cached.capability;
         }
-        // Browser inference always supports exact log-probabilities
+        // Declared, not measured — no model has been loaded/tested here.
         if (engine === "browser") {
             const capability = {
                 modelId,
                 engine: "browser",
                 method: "logprobs_exact",
-                status: "available",
+                status: "unverified",
                 recommended: true,
-                lastTested: new Date().toISOString(),
             };
             this.setCachedCapability(capability);
             return capability;
@@ -252,7 +254,9 @@ export class DefaultCapabilityDetectionService {
      * Calculate recommendation score for a capability
      */
     getRecommendationScore(capability) {
-        if (capability.status !== "available")
+        // "unverified" is a usable, declared-supported model (see the type doc) —
+        // only truly broken/unknown states (unavailable, error, detecting) score 0.
+        if (capability.status !== "available" && capability.status !== "unverified")
             return 0;
         let score = 0;
         // Browser inference gets highest score
@@ -283,15 +287,14 @@ export class EnhancedCapabilityDetectionService extends DefaultCapabilityDetecti
         if (cached && this.isCacheValid(cached)) {
             return cached.capability;
         }
-        // Browser inference always supports exact log-probabilities
+        // Declared, not measured — no model has been loaded/tested here.
         if (engine === "browser") {
             const capability = {
                 modelId,
                 engine: "browser",
                 method: "logprobs_exact",
-                status: "available",
+                status: "unverified",
                 recommended: true,
-                lastTested: new Date().toISOString(),
             };
             this.setCachedCapability(capability);
             return capability;
@@ -378,7 +381,8 @@ export class EnhancedCapabilityDetectionService extends DefaultCapabilityDetecti
      */
     async detectAllCapabilities() {
         const capabilities = [];
-        // Add all supported browser models dynamically
+        // Add all supported browser models dynamically. Declared, not measured —
+        // no model has actually been loaded/tested at this point.
         try {
             const { BROWSER_MODELS } = await import("./browser-inference");
             for (const model of BROWSER_MODELS) {
@@ -388,9 +392,8 @@ export class EnhancedCapabilityDetectionService extends DefaultCapabilityDetecti
                     modelId: model.id,
                     engine: "browser",
                     method: "logprobs_exact",
-                    status: "available",
+                    status: "unverified",
                     recommended: true,
-                    lastTested: new Date().toISOString(),
                 });
             }
         }
@@ -401,9 +404,8 @@ export class EnhancedCapabilityDetectionService extends DefaultCapabilityDetecti
                 modelId: "Xenova/distilgpt2",
                 engine: "browser",
                 method: "logprobs_exact",
-                status: "available",
+                status: "unverified",
                 recommended: true,
-                lastTested: new Date().toISOString(),
             });
         }
         try {

package/dist/inference/index.d.ts

@@ -2,8 +2,19 @@
  * Inference Factory
  * Auto-detects available inference engines and provides unified interface
  */
-import { type BrowserModel } from "./browser-inference";
 import { type OllamaModel } from "./ollama-client";
+import { type BrowserModel, type ScoredLogProb } from "./transformers-inference";
+/**
+ * The one contract a bias-measurement engine must satisfy: a real, cross-checked
+ * sentence log-probability (see Task 1 / Messmotor-Sanierung). `OllamaClient`
+ * deliberately does NOT implement this — Ollama's logprobs only cover generated
+ * tokens, never prompt/sentence log-probability, so it cannot honestly produce a
+ * `ScoredLogProb`. TypeScript's structural typing then locks it out of any API
+ * that requires this interface (e.g. `calculateCrowsPairsBias`).
+ */
+export interface SequenceLogProbProvider {
+    getScoredLogProb(text: string): Promise<ScoredLogProb>;
+}
 export type InferenceEngine = "ollama" | "browser" | "demo";
 export interface InferenceCapabilities {
     ollama: {
@@ -36,7 +47,9 @@ export declare function createInferenceClient(engine: InferenceEngine, modelId:
  */
 export declare function createAutoInferenceClient(preferredModel?: string): Promise<InferenceClient>;
 export { detectOllama, listOllamaModels, OllamaClient, type CapabilityCache, } from "./ollama-client";
-export { BrowserInference, checkWebGPU, BROWSER_MODELS } from "./browser-inference";
+export { TransformersInference, checkWebGPU, BROWSER_MODELS, type ScoredLogProb, type ScoredLogProbProvenance, } from "./transformers-inference";
+export { isConditionalGenerationConfig, type MinimalModelConfig } from "./template-scoring";
+export { BrowserInference } from "./browser-inference";
 export { InferenceFactory, createInferenceFactory, detectBestMethodForConfig, type InferenceResult, } from "./inference-factory";
 export { createCapabilityDetectionService, DefaultCapabilityDetectionService, EnhancedCapabilityDetectionService, type CapabilityDetectionService, type ModelCapability, type ModelSelectionState, type CapabilityCacheEntry, type DetectionProgress, type UserPreferences, type EnhancedModelSelectionState, type CapabilityDetectionStatus, } from "./capability-detection";
 export type { OllamaModel, BrowserModel };

package/dist/inference/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/inference/index.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAEH,OAAO,EAGL,KAAK,YAAY,EAElB,MAAM,qBAAqB,CAAC;AAC7B,OAAO,EAAgB,KAAK,WAAW,EAAkC,MAAM,iBAAiB,CAAC;AAEjG,MAAM,MAAM,eAAe,GAAG,QAAQ,GAAG,SAAS,GAAG,MAAM,CAAC;AAE5D,MAAM,WAAW,qBAAqB;IACpC,MAAM,EAAE;QACN,SAAS,EAAE,OAAO,CAAC;QACnB,MAAM,EAAE,WAAW,EAAE,CAAC;KACvB,CAAC;IACF,OAAO,EAAE;QACP,SAAS,EAAE,OAAO,CAAC;QACnB,MAAM,EAAE,OAAO,CAAC;QAChB,MAAM,EAAE,YAAY,EAAE,CAAC;KACxB,CAAC;IACF,WAAW,EAAE,eAAe,CAAC;CAC9B;AAED,MAAM,WAAW,eAAe;IAC9B,MAAM,EAAE,eAAe,CAAC;IACxB,OAAO,EAAE,MAAM,CAAC;IAChB,QAAQ,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC;IAC1C,UAAU,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC;CAC7C;AAED;;GAEG;AACH,wBAAsB,kBAAkB,IAAI,OAAO,CAAC,qBAAqB,CAAC,CA2BzE;AAED;;GAEG;AACH,wBAAgB,qBAAqB,CACnC,MAAM,EAAE,eAAe,EACvB,OAAO,EAAE,MAAM,EACf,cAAc,CAAC,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,IAAI,GACxC,eAAe,CAqCjB;AAED;;GAEG;AACH,wBAAsB,yBAAyB,CAAC,cAAc,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,eAAe,CAAC,CAejG;AAGD,OAAO,EACL,YAAY,EACZ,gBAAgB,EAChB,YAAY,EACZ,KAAK,eAAe,GACrB,MAAM,iBAAiB,CAAC;AACzB,OAAO,EAAE,gBAAgB,EAAE,WAAW,EAAE,cAAc,EAAE,MAAM,qBAAqB,CAAC;AACpF,OAAO,EACL,gBAAgB,EAChB,sBAAsB,EACtB,yBAAyB,EACzB,KAAK,eAAe,GACrB,MAAM,qBAAqB,CAAC;AAC7B,OAAO,EACL,gCAAgC,EAChC,iCAAiC,EACjC,kCAAkC,EAClC,KAAK,0BAA0B,EAC/B,KAAK,eAAe,EACpB,KAAK,mBAAmB,EACxB,KAAK,oBAAoB,EACzB,KAAK,iBAAiB,EACtB,KAAK,eAAe,EACpB,KAAK,2BAA2B,EAChC,KAAK,yBAAyB,GAC/B,MAAM,wBAAwB,CAAC;AAChC,YAAY,EAAE,WAAW,EAAE,YAAY,EAAE,CAAC"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/inference/index.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAEH,OAAO,EAAgB,KAAK,WAAW,EAAkC,MAAM,iBAAiB,CAAC;AACjG,OAAO,EAEL,KAAK,YAAY,EACjB,KAAK,aAAa,EAGnB,MAAM,0BAA0B,CAAC;AAElC;;;;;;;GAOG;AACH,MAAM,WAAW,uBAAuB;IACtC,gBAAgB,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,aAAa,CAAC,CAAC;CACxD;AAED,MAAM,MAAM,eAAe,GAAG,QAAQ,GAAG,SAAS,GAAG,MAAM,CAAC;AAE5D,MAAM,WAAW,qBAAqB;IACpC,MAAM,EAAE;QACN,SAAS,EAAE,OAAO,CAAC;QACnB,MAAM,EAAE,WAAW,EAAE,CAAC;KACvB,CAAC;IACF,OAAO,EAAE;QACP,SAAS,EAAE,OAAO,CAAC;QACnB,MAAM,EAAE,OAAO,CAAC;QAChB,MAAM,EAAE,YAAY,EAAE,CAAC;KACxB,CAAC;IACF,WAAW,EAAE,eAAe,CAAC;CAC9B;AAED,MAAM,WAAW,eAAe;IAC9B,MAAM,EAAE,eAAe,CAAC;IACxB,OAAO,EAAE,MAAM,CAAC;IAChB,QAAQ,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC;IAC1C,UAAU,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC;CAC7C;AAED;;GAEG;AACH,wBAAsB,kBAAkB,IAAI,OAAO,CAAC,qBAAqB,CAAC,CA2BzE;AAED;;GAEG;AACH,wBAAgB,qBAAqB,CACnC,MAAM,EAAE,eAAe,EACvB,OAAO,EAAE,MAAM,EACf,cAAc,CAAC,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,IAAI,GACxC,eAAe,CAqCjB;AAED;;GAEG;AACH,wBAAsB,yBAAyB,CAAC,cAAc,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,eAAe,CAAC,CAejG;AAGD,OAAO,EACL,YAAY,EACZ,gBAAgB,EAChB,YAAY,EACZ,KAAK,eAAe,GACrB,MAAM,iBAAiB,CAAC;AACzB,OAAO,EACL,qBAAqB,EACrB,WAAW,EACX,cAAc,EACd,KAAK,aAAa,EAClB,KAAK,uBAAuB,GAC7B,MAAM,0BAA0B,CAAC;AAOlC,OAAO,EAAE,6BAA6B,EAAE,KAAK,kBAAkB,EAAE,MAAM,oBAAoB,CAAC;AAE5F,OAAO,EAAE,gBAAgB,EAAE,MAAM,qBAAqB,CAAC;AACvD,OAAO,EACL,gBAAgB,EAChB,sBAAsB,EACtB,yBAAyB,EACzB,KAAK,eAAe,GACrB,MAAM,qBAAqB,CAAC;AAC7B,OAAO,EACL,gCAAgC,EAChC,iCAAiC,EACjC,kCAAkC,EAClC,KAAK,0BAA0B,EAC/B,KAAK,eAAe,EACpB,KAAK,mBAAmB,EACxB,KAAK,oBAAoB,EACzB,KAAK,iBAAiB,EACtB,KAAK,eAAe,EACpB,KAAK,2BAA2B,EAChC,KAAK,yBAAyB,GAC/B,MAAM,wBAAwB,CAAC;AAChC,YAAY,EAAE,WAAW,EAAE,YAAY,EAAE,CAAC"}

package/dist/inference/index.js

@@ -2,8 +2,8 @@
  * Inference Factory
  * Auto-detects available inference engines and provides unified interface
  */
-import { BROWSER_MODELS, BrowserInference, checkWebGPU, } from "./browser-inference";
 import { OllamaClient, detectOllama, listOllamaModels } from "./ollama-client";
+import { BROWSER_MODELS, TransformersInference, checkWebGPU, } from "./transformers-inference";
 /**
  * Detect all available inference capabilities
  */
@@ -49,7 +49,7 @@ export function createInferenceClient(engine, modelId, statusCallback) {
             };
         }
         case "browser": {
-            const browserClient = new BrowserInference(modelId, statusCallback);
+            const browserClient = new TransformersInference(modelId, statusCallback);
             return {
                 engine: "browser",
                 modelId,
@@ -92,6 +92,15 @@ export async function createAutoInferenceClient(preferredModel) {
 }
 // Re-export types and utilities
 export { detectOllama, listOllamaModels, OllamaClient, } from "./ollama-client";
-export { BrowserInference, checkWebGPU, BROWSER_MODELS } from "./browser-inference";
+export { TransformersInference, checkWebGPU, BROWSER_MODELS, } from "./transformers-inference";
+// Architecture detection for the Gemma-4 template-mode scorer, re-exported so callers that
+// must pick a dtype BEFORE constructing `TransformersInference` (e.g. the CLI's
+// `run-bias-test.ts`, which needs to choose "q4" over "fp32" for conditional-generation
+// models — fp32 Gemma-4 weights were measured and rejected as impractically large in
+// Messmotor-Sanierung Task 5) can reuse the same detection `TransformersInference` uses
+// internally, instead of duplicating the architecture check.
+export { isConditionalGenerationConfig } from "./template-scoring";
+// Backward-compat alias: existing consumers (apps/web, tests) import `BrowserInference`.
+export { BrowserInference } from "./browser-inference";
 export { InferenceFactory, createInferenceFactory, detectBestMethodForConfig, } from "./inference-factory";
 export { createCapabilityDetectionService, DefaultCapabilityDetectionService, EnhancedCapabilityDetectionService, } from "./capability-detection";

package/dist/inference/inference-factory.d.ts

@@ -31,10 +31,13 @@ export declare class InferenceFactory {
      */
     setSelectedCapability(capability: ModelCapability): void;
     /**
-     * Create a bias calculation with automatic method detection
-     * Returns InferenceResult with both data and method metadata
+     * @deprecated Uses the removed Ollama logprob path: this method auto-selected
+     * between browser and Ollama engines via `getInferenceClient`, which no longer
+     * measures anything honest for Ollama (see Messmotor-Sanierung Task 2). The
+     * auto-selection logic itself needs a real redesign, not a throw-or-convert
+     * fix — see bias --engine transformers for the supported path.
      */
-    createBiasCalculation(config: BiasTestConfig): Promise<InferenceResult<CrowsPairsBiasResult>>;
+    createBiasCalculation(_config: BiasTestConfig): Promise<InferenceResult<CrowsPairsBiasResult>>;
     /**
      * Detect the best available calculation method for the given client
      * Returns the method that should be used for bias calculations

package/dist/inference/inference-factory.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"inference-factory.d.ts","sourceRoot":"","sources":["../../src/inference/inference-factory.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAGH,OAAO,KAAK,EAAE,qBAAqB,EAAE,cAAc,EAAE,oBAAoB,EAAE,MAAM,UAAU,CAAC;AAC5F,OAAO,EAAE,gBAAgB,EAAE,MAAM,qBAAqB,CAAC;AACvD,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,wBAAwB,CAAC;AAC9D,OAAO,EAAE,YAAY,EAAE,MAAM,iBAAiB,CAAC;AAE/C;;GAEG;AACH,MAAM,WAAW,eAAe,CAAC,CAAC;IAChC,IAAI,EAAE,CAAC,CAAC;IACR,MAAM,EAAE,qBAAqB,CAAC;IAC9B,MAAM,EAAE,SAAS,GAAG,QAAQ,CAAC;IAC7B,KAAK,EAAE,MAAM,CAAC;IACd,UAAU,CAAC,EAAE,eAAe,CAAC;CAC9B;AAED;;;GAGG;AACH,qBAAa,gBAAgB;IAC3B,OAAO,CAAC,aAAa,CAAiC;IACtD,OAAO,CAAC,YAAY,CAA6B;IACjD,OAAO,CAAC,cAAc,CAA2C;IACjE,OAAO,CAAC,kBAAkB,CAAgC;gBAE9C,cAAc,CAAC,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,IAAI;IAIrD;;OAEG;IACH,qBAAqB,CAAC,UAAU,EAAE,eAAe,GAAG,IAAI;IAIxD;;;OAGG;IACG,qBAAqB,CACzB,MAAM,EAAE,cAAc,GACrB,OAAO,CAAC,eAAe,CAAC,oBAAoB,CAAC,CAAC;IA6BjD;;;OAGG;YACW,gBAAgB;IAe9B;;OAEG;IACH,gBAAgB,CAAC,KAAK,EAAE,MAAM,GAAG,gBAAgB;IAOjD;;OAEG;IACH,eAAe,CAAC,KAAK,EAAE,MAAM,GAAG,YAAY;IAO5C,OAAO,CAAC,qBAAqB;YAgBf,kBAAkB;IAUhC;;;OAGG;IACG,qBAAqB,CAAC,CAAC,EAC3B,IAAI,EAAE,CAAC,EACP,MAAM,EAAE,SAAS,GAAG,QAAQ,EAC5B,KAAK,EAAE,MAAM,EACb,MAAM,CAAC,EAAE,gBAAgB,GAAG,YAAY,GACvC,OAAO,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC;IA8B9B;;OAEG;IACH,OAAO,IAAI,IAAI;CAMhB;AAED;;;GAGG;AACH,wBAAgB,sBAAsB,CACpC,cAAc,CAAC,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,IAAI,GACxC,gBAAgB,CAElB;AAED;;;GAGG;AACH,wBAAsB,yBAAyB,CAC7C,MAAM,EAAE,cAAc,GACrB,OAAO,CAAC,qBAAqB,CAAC,CAYhC"}
+{"version":3,"file":"inference-factory.d.ts","sourceRoot":"","sources":["../../src/inference/inference-factory.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAEH,OAAO,KAAK,EAAE,qBAAqB,EAAE,cAAc,EAAE,oBAAoB,EAAE,MAAM,UAAU,CAAC;AAC5F,OAAO,EAAE,gBAAgB,EAAE,MAAM,qBAAqB,CAAC;AACvD,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,wBAAwB,CAAC;AAC9D,OAAO,EAAE,YAAY,EAAE,MAAM,iBAAiB,CAAC;AAE/C;;GAEG;AACH,MAAM,WAAW,eAAe,CAAC,CAAC;IAChC,IAAI,EAAE,CAAC,CAAC;IACR,MAAM,EAAE,qBAAqB,CAAC;IAC9B,MAAM,EAAE,SAAS,GAAG,QAAQ,CAAC;IAC7B,KAAK,EAAE,MAAM,CAAC;IACd,UAAU,CAAC,EAAE,eAAe,CAAC;CAC9B;AAED;;;GAGG;AACH,qBAAa,gBAAgB;IAC3B,OAAO,CAAC,aAAa,CAAiC;IACtD,OAAO,CAAC,YAAY,CAA6B;IACjD,OAAO,CAAC,cAAc,CAA2C;IACjE,OAAO,CAAC,kBAAkB,CAAgC;gBAE9C,cAAc,CAAC,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,IAAI;IAIrD;;OAEG;IACH,qBAAqB,CAAC,UAAU,EAAE,eAAe,GAAG,IAAI;IAIxD;;;;;;OAMG;IACG,qBAAqB,CACzB,OAAO,EAAE,cAAc,GACtB,OAAO,CAAC,eAAe,CAAC,oBAAoB,CAAC,CAAC;IAMjD;;;OAGG;YACW,gBAAgB;IAe9B;;OAEG;IACH,gBAAgB,CAAC,KAAK,EAAE,MAAM,GAAG,gBAAgB;IAOjD;;OAEG;IACH,eAAe,CAAC,KAAK,EAAE,MAAM,GAAG,YAAY;IAO5C,OAAO,CAAC,qBAAqB;YAgBf,kBAAkB;IAUhC;;;OAGG;IACG,qBAAqB,CAAC,CAAC,EAC3B,IAAI,EAAE,CAAC,EACP,MAAM,EAAE,SAAS,GAAG,QAAQ,EAC5B,KAAK,EAAE,MAAM,EACb,MAAM,CAAC,EAAE,gBAAgB,GAAG,YAAY,GACvC,OAAO,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC;IA8B9B;;OAEG;IACH,OAAO,IAAI,IAAI;CAMhB;AAED;;;GAGG;AACH,wBAAgB,sBAAsB,CACpC,cAAc,CAAC,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,IAAI,GACxC,gBAAgB,CAElB;AAED;;;GAGG;AACH,wBAAsB,yBAAyB,CAC7C,MAAM,EAAE,cAAc,GACrB,OAAO,CAAC,qBAAqB,CAAC,CAYhC"}

package/dist/inference/inference-factory.js

@@ -2,7 +2,6 @@
  * Inference Factory with Method Detection
  * Provides unified interface for bias calculations across engines with automatic method detection
  */
-import { calculateCrowsPairsBias } from "../metrics/crows-pairs-bias";
 import { BrowserInference } from "./browser-inference";
 import { OllamaClient } from "./ollama-client";
 /**
@@ -24,32 +23,14 @@ export class InferenceFactory {
         this.selectedCapability = capability;
     }
     /**
-     * Create a bias calculation with automatic method detection
-     * Returns InferenceResult with both data and method metadata
+     * @deprecated Uses the removed Ollama logprob path: this method auto-selected
+     * between browser and Ollama engines via `getInferenceClient`, which no longer
+     * measures anything honest for Ollama (see Messmotor-Sanierung Task 2). The
+     * auto-selection logic itself needs a real redesign, not a throw-or-convert
+     * fix — see bias --engine transformers for the supported path.
      */
-    async createBiasCalculation(config) {
-        const { engine, model } = config;
-        this.validateConfiguration(config);
-        const inferenceClient = await this.getInferenceClient(engine, model);
-        let _detectedMethod;
-        if (engine === "browser") {
-            _detectedMethod = "logprobs_exact";
-        }
-        else {
-            // Use capability method if available, otherwise detect
-            _detectedMethod =
-                this.selectedCapability?.method ?? (await this.detectBestMethod(inferenceClient));
-        }
-        // Perform bias calculation
-        const biasResult = await calculateCrowsPairsBias(config, inferenceClient);
-        // Return result with metadata including capability information
-        return {
-            data: biasResult,
-            method: biasResult.method, // Use the method determined during calculation
-            engine,
-            model,
-            capability: this.selectedCapability || undefined,
-        };
+    async createBiasCalculation(_config) {
+        throw new Error("InferenceFactory.createBiasCalculation verwendet den entfernten Ollama-Logprob-Pfad; siehe bias --engine transformers.");
     }
     /**
      * Detect the best available calculation method for the given client

package/dist/inference/ollama-client.d.ts

@@ -89,9 +89,18 @@ export declare class OllamaClient {
     ensureModelLoaded(): Promise<void>;
     isAvailable(): Promise<boolean>;
     generate(prompt: string): Promise<string>;
+    /**
+     * @deprecated Not a sentence log-probability — see `getLogProbWithFallback`.
+     */
     getLogProb(prompt: string): Promise<number>;
     getPerplexity(text: string): Promise<number>;
     /**
+     * @deprecated Part of the removed Ollama bias-measurement path (Messmotor-Sanierung
+     * Task 2): Ollama's logprobs cover only generated tokens, never the prompt/sentence,
+     * so this "capability" check never verified a real measurement. Kept only because
+     * `scan`'s non-bias paths and the capability-cache tests still reference it — not
+     * a foundation to build on. See bias --engine transformers for the real path.
+     *
      * Test if the current model supports log-probabilities
      * Sends a test request with "Der" prompt to detect capability
      */
@@ -105,13 +114,33 @@ export declare class OllamaClient {
      */
     private setCachedCapability;
     /**
+     * @deprecated Part of the removed Ollama bias-measurement path (see
+     * `testLogProbCapability`); no longer used by `calculateCrowsPairsBias`.
+     *
      * Detect if the current model supports log-probabilities
      * Checks cache first, then runs test if needed
      */
     detectLogProbSupport(): Promise<boolean>;
+    /**
+     * @deprecated Part of the removed Ollama bias-measurement path; see
+     * `testLogProbCapability`. Not a real sentence log-probability.
+     */
     private tryExactLogProbs;
+    /**
+     * @deprecated Part of the removed Ollama bias-measurement path; this is a
+     * latency proxy, not a probability — the exact thing Messmotor-Sanierung
+     * removes from the bias measurement path. Do not build on this.
+     */
     private latencyFallback;
     /**
+     * @deprecated Removed from the bias-measurement path (Messmotor-Sanierung Task 2):
+     * Ollama returns logprobs only for generated tokens, never for the prompt/sentence
+     * being scored, so this never measured a sentence log-probability — even the
+     * "exact" branch (`tryExactLogProbs`) is not what its name implies, and the
+     * fallback is a latency proxy. Verified: r(sentence length, logP)=0.019 where a
+     * real sentence-probability measure shows −0.96. Use `TransformersInference.
+     * getScoredLogProb` (via `bias --engine transformers`) instead.
+     *
      * Get log-probability with automatic fallback to latency method
      * Returns both the calculated value and the method used
      */

package/dist/inference/ollama-client.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"ollama-client.d.ts","sourceRoot":"","sources":["../../src/inference/ollama-client.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAEH,OAAO,KAAK,EAAE,qBAAqB,EAAwB,MAAM,UAAU,CAAC;AAE5E,MAAM,WAAW,WAAW;IAC1B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;CACrB;AAED,MAAM,WAAW,qBAAqB;IACpC,KAAK,EAAE,MAAM,CAAC;IACd,MAAM,EAAE,MAAM,CAAC;IACf,MAAM,CAAC,EAAE,OAAO,CAAC;IACjB,QAAQ,CAAC,EAAE,OAAO,GAAG,MAAM,CAAC;IAC5B,OAAO,CAAC,EAAE;QACR,WAAW,CAAC,EAAE,MAAM,CAAC;QACrB,KAAK,CAAC,EAAE,MAAM,CAAC;QACf,WAAW,CAAC,EAAE,MAAM,CAAC;KACtB,CAAC;CACH;AAED,MAAM,WAAW,sBAAsB;IACrC,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,CAAC;IACjB,IAAI,EAAE,OAAO,CAAC;IACd,OAAO,CAAC,EAAE,MAAM,EAAE,CAAC;IACnB,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,QAAQ,CAAC,EAAE,KAAK,CAAC;QACf,KAAK,EAAE,MAAM,CAAC;QACd,OAAO,EAAE,MAAM,CAAC;KACjB,CAAC,CAAC;CACJ;AAED,MAAM,WAAW,qBAAqB;IACpC,KAAK,EAAE,MAAM,CAAC;IACd,MAAM,EAAE,MAAM,CAAC;IACf,OAAO,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,qBAAqB;IACpC,QAAQ,CAAC,EAAE,OAAO,GAAG,MAAM,CAAC;CAC7B;AAED,MAAM,WAAW,sBAAuB,SAAQ,sBAAsB;IACpE,QAAQ,CAAC,EAAE,KAAK,CAAC;QACf,KAAK,EAAE,MAAM,CAAC;QACd,OAAO,EAAE,MAAM,CAAC;KACjB,CAAC,CAAC;CACJ;AAED,sFAAsF;AACtF,MAAM,WAAW,eAAe;IAC9B,GAAG,CAAC,GAAG,EAAE,MAAM,GAAG,MAAM,GAAG,IAAI,CAAC;IAChC,GAAG,CAAC,GAAG,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG,IAAI,CAAC;CACvC;AAID;;GAEG;AACH,wBAAsB,YAAY,IAAI,OAAO,CAAC,OAAO,CAAC,CAcrD;AAED;;GAEG;AACH,wBAAsB,gBAAgB,IAAI,OAAO,CAAC,WAAW,EAAE,CAAC,CAc/D;AAED;;GAEG;AACH,wBAAsB,cAAc,CAClC,OAAO,EAAE,qBAAqB,GAC7B,OAAO,CAAC,sBAAsB,CAAC,CAkBjC;AAED;;;GAGG;AACH,wBAAsB,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CA+C/E;AAED;;GAEG;AACH,wBAAsB,aAAa,CAAC,KAAK,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAiChF;AAED;;GAEG;AACH,qBAAa,YAAY;IACvB,OAAO,CAAC,OAAO,CAAS;IACxB,SAAgB,KAAK,EAAE,MAAM,CAAC;IAC9B,OAAO,CAAC,eAAe,CAAoC;IAC3D,OAAO,CAAC,aAAa,CAAyB;gBAElC,KAAK,EAAE,MAAM,EAAE,OAAO,SAAkB,EAAE,KAAK,CAAC,EAAE,eAAe;IAO7E,OAAO,CAAC,YAAY;IAIpB;;;OAGG;IACG,iBAAiB,IAAI,OAAO,CAAC,IAAI,CAAC;IAiBlC,WAAW,IAAI,OAAO,CAAC,OAAO,CAAC;IAI/B,QAAQ,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IASzC,UAAU,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAM3C,aAAa,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAIlD;;;OAGG;YACW,qBAAqB;IAmEnC;;OAEG;IACH,OAAO,CAAC,mBAAmB;IA0B3B;;OAEG;IACH,OAAO,CAAC,mBAAmB;IAqB3B;;;OAGG;IACG,oBAAoB,IAAI,OAAO,CAAC,OAAO,CAAC;YAqBhC,gBAAgB;YAgDhB,eAAe;IAyC7B;;;OAGG;IACG,sBAAsB,CAC1B,MAAM,EAAE,MAAM,GACb,OAAO,CAAC;QAAE,OAAO,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,qBAAqB,CAAA;KAAE,CAAC;CAU/D"}
+{"version":3,"file":"ollama-client.d.ts","sourceRoot":"","sources":["../../src/inference/ollama-client.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAEH,OAAO,KAAK,EAAE,qBAAqB,EAAwB,MAAM,UAAU,CAAC;AAE5E,MAAM,WAAW,WAAW;IAC1B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;CACrB;AAED,MAAM,WAAW,qBAAqB;IACpC,KAAK,EAAE,MAAM,CAAC;IACd,MAAM,EAAE,MAAM,CAAC;IACf,MAAM,CAAC,EAAE,OAAO,CAAC;IACjB,QAAQ,CAAC,EAAE,OAAO,GAAG,MAAM,CAAC;IAC5B,OAAO,CAAC,EAAE;QACR,WAAW,CAAC,EAAE,MAAM,CAAC;QACrB,KAAK,CAAC,EAAE,MAAM,CAAC;QACf,WAAW,CAAC,EAAE,MAAM,CAAC;KACtB,CAAC;CACH;AAED,MAAM,WAAW,sBAAsB;IACrC,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,CAAC;IACjB,IAAI,EAAE,OAAO,CAAC;IACd,OAAO,CAAC,EAAE,MAAM,EAAE,CAAC;IACnB,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,QAAQ,CAAC,EAAE,KAAK,CAAC;QACf,KAAK,EAAE,MAAM,CAAC;QACd,OAAO,EAAE,MAAM,CAAC;KACjB,CAAC,CAAC;CACJ;AAED,MAAM,WAAW,qBAAqB;IACpC,KAAK,EAAE,MAAM,CAAC;IACd,MAAM,EAAE,MAAM,CAAC;IACf,OAAO,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,qBAAqB;IACpC,QAAQ,CAAC,EAAE,OAAO,GAAG,MAAM,CAAC;CAC7B;AAED,MAAM,WAAW,sBAAuB,SAAQ,sBAAsB;IACpE,QAAQ,CAAC,EAAE,KAAK,CAAC;QACf,KAAK,EAAE,MAAM,CAAC;QACd,OAAO,EAAE,MAAM,CAAC;KACjB,CAAC,CAAC;CACJ;AAED,sFAAsF;AACtF,MAAM,WAAW,eAAe;IAC9B,GAAG,CAAC,GAAG,EAAE,MAAM,GAAG,MAAM,GAAG,IAAI,CAAC;IAChC,GAAG,CAAC,GAAG,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG,IAAI,CAAC;CACvC;AAID;;GAEG;AACH,wBAAsB,YAAY,IAAI,OAAO,CAAC,OAAO,CAAC,CAcrD;AAED;;GAEG;AACH,wBAAsB,gBAAgB,IAAI,OAAO,CAAC,WAAW,EAAE,CAAC,CAc/D;AAED;;GAEG;AACH,wBAAsB,cAAc,CAClC,OAAO,EAAE,qBAAqB,GAC7B,OAAO,CAAC,sBAAsB,CAAC,CAkBjC;AAED;;;GAGG;AACH,wBAAsB,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CA+C/E;AAED;;GAEG;AACH,wBAAsB,aAAa,CAAC,KAAK,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAiChF;AAED;;GAEG;AACH,qBAAa,YAAY;IACvB,OAAO,CAAC,OAAO,CAAS;IACxB,SAAgB,KAAK,EAAE,MAAM,CAAC;IAC9B,OAAO,CAAC,eAAe,CAAoC;IAC3D,OAAO,CAAC,aAAa,CAAyB;gBAElC,KAAK,EAAE,MAAM,EAAE,OAAO,SAAkB,EAAE,KAAK,CAAC,EAAE,eAAe;IAO7E,OAAO,CAAC,YAAY;IAIpB;;;OAGG;IACG,iBAAiB,IAAI,OAAO,CAAC,IAAI,CAAC;IAiBlC,WAAW,IAAI,OAAO,CAAC,OAAO,CAAC;IAI/B,QAAQ,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAS/C;;OAEG;IACG,UAAU,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAM3C,aAAa,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAIlD;;;;;;;;;OASG;YACW,qBAAqB;IAmEnC;;OAEG;IACH,OAAO,CAAC,mBAAmB;IA0B3B;;OAEG;IACH,OAAO,CAAC,mBAAmB;IAqB3B;;;;;;OAMG;IACG,oBAAoB,IAAI,OAAO,CAAC,OAAO,CAAC;IAqB9C;;;OAGG;YACW,gBAAgB;IAgD9B;;;;OAIG;YACW,eAAe;IAyC7B;;;;;;;;;;;OAWG;IACG,sBAAsB,CAC1B,MAAM,EAAE,MAAM,GACb,OAAO,CAAC;QAAE,OAAO,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,qBAAqB,CAAA;KAAE,CAAC;CAU/D"}

package/dist/inference/ollama-client.js

@@ -177,6 +177,9 @@ export class OllamaClient {
         });
         return response.response;
     }
+    /**
+     * @deprecated Not a sentence log-probability — see `getLogProbWithFallback`.
+     */
     async getLogProb(prompt) {
         // Use the enhanced getLogProbWithFallback method that supports both exact and fallback methods
         const result = await this.getLogProbWithFallback(prompt);
@@ -186,6 +189,12 @@ export class OllamaClient {
         return getPerplexity(this.model, text);
     }
     /**
+     * @deprecated Part of the removed Ollama bias-measurement path (Messmotor-Sanierung
+     * Task 2): Ollama's logprobs cover only generated tokens, never the prompt/sentence,
+     * so this "capability" check never verified a real measurement. Kept only because
+     * `scan`'s non-bias paths and the capability-cache tests still reference it — not
+     * a foundation to build on. See bias --engine transformers for the real path.
+     *
      * Test if the current model supports log-probabilities
      * Sends a test request with "Der" prompt to detect capability
      */
@@ -292,6 +301,9 @@ export class OllamaClient {
         }
     }
     /**
+     * @deprecated Part of the removed Ollama bias-measurement path (see
+     * `testLogProbCapability`); no longer used by `calculateCrowsPairsBias`.
+     *
      * Detect if the current model supports log-probabilities
      * Checks cache first, then runs test if needed
      */
@@ -312,6 +324,10 @@ export class OllamaClient {
         this.setCachedCapability(supports);
         return supports;
     }
+    /**
+     * @deprecated Part of the removed Ollama bias-measurement path; see
+     * `testLogProbCapability`. Not a real sentence log-probability.
+     */
     async tryExactLogProbs(prompt) {
         const supportsLogProbs = await this.detectLogProbSupport();
         if (!supportsLogProbs || this.isCloudModel())
@@ -346,6 +362,11 @@ export class OllamaClient {
             return null;
         }
     }
+    /**
+     * @deprecated Part of the removed Ollama bias-measurement path; this is a
+     * latency proxy, not a probability — the exact thing Messmotor-Sanierung
+     * removes from the bias measurement path. Do not build on this.
+     */
     async latencyFallback(prompt) {
         const isCloud = this.isCloudModel();
         console.log(`⚡ Falling back to latency-based method for ${this.model}${isCloud ? " (cloud model)" : ""}`);
@@ -378,6 +399,14 @@ export class OllamaClient {
         return { logprob: -Math.log(latency + 0.00001), method: "logprobs_fallback_latency" };
     }
     /**
+     * @deprecated Removed from the bias-measurement path (Messmotor-Sanierung Task 2):
+     * Ollama returns logprobs only for generated tokens, never for the prompt/sentence
+     * being scored, so this never measured a sentence log-probability — even the
+     * "exact" branch (`tryExactLogProbs`) is not what its name implies, and the
+     * fallback is a latency proxy. Verified: r(sentence length, logP)=0.019 where a
+     * real sentence-probability measure shows −0.96. Use `TransformersInference.
+     * getScoredLogProb` (via `bias --engine transformers`) instead.
+     *
      * Get log-probability with automatic fallback to latency method
      * Returns both the calculated value and the method used
      */

package/dist/inference/scored-log-prob-types.d.ts

@@ -0,0 +1,64 @@
+/**
+ * Shared result types for `SequenceLogProbProvider.getScoredLogProb`.
+ *
+ * Pulled out of `transformers-inference.ts` (Messmotor-Sanierung Task 5) so that
+ * `template-scoring.ts` (the Gemma-4 template-mode scorer) and `transformers-inference.ts`
+ * (which loads models and picks which mode to use) can both depend on these types without
+ * either module importing from the other — avoids a circular import between the two.
+ *
+ * `transformers-inference.ts` re-exports both types from here, so existing imports of
+ * `ScoredLogProb`/`ScoredLogProbProvenance` from `"./transformers-inference"` keep working.
+ */
+import type { SequenceLogProbMethod } from "../types";
+import type { Span } from "./sequence-logprob";
+/** Provenance metadata attached to a scored log-probability result. */
+export interface ScoredLogProbProvenance {
+    modelId: string;
+    /**
+     * The CONFIGURED weight precision passed to `from_pretrained` (e.g. "q4", "fp32", "q8") —
+     * NOT the output tensor's reported dtype (see `logitsDtype` for that; the two are
+     * different facts and must not be conflated). Task 1 measured 1.3-11.9% score deviation
+     * between quantized and fp32 weights on the same model/sentences, so this field is
+     * ergebnisrelevant: an artifact that cannot say which weight precision produced a score
+     * claims a precision it does not have (Messmotor-Sanierung Task 5 review finding).
+     */
+    dtype: string;
+    /**
+     * The OUTPUT LOGITS TENSOR's reported `.type` (e.g. "float32"). The Task 0 spike found
+     * this stays "float32" even for q4-quantized weights ("logits.data ist Float32Array trotz
+     * q4-Gewichten") — so this field does NOT tell you the weight precision; use `dtype` for
+     * that. Kept because it is still a real, independently useful fact (confirms no silent
+     * downcast in the output itself).
+     */
+    logitsDtype: string;
+    backend: string;
+    libraryVersion: string;
+    /**
+     * "none" for the base autoregressive mode (Task 1); "fixed_template_span" for the
+     * instruction-tuned template mode (Task 5), which conditions scoring on a fixed
+     * user-turn prompt. Optional so pre-Task-5 test doubles that predate this field
+     * (e.g. hand-built `ScoredLogProb` literals in other packages' tests) stay valid;
+     * every real measurement path sets it explicitly.
+     */
+    templateStrategy?: "none" | "fixed_template_span";
+    /** Present only when templateStrategy is "fixed_template_span": the verbatim fixed
+     * context string fed before the sentence — recomputable by any third party. */
+    promptText?: string;
+    /** SHA-256 of `promptText`, present only when templateStrategy is "fixed_template_span". */
+    promptSha256?: string;
+}
+/** Full result of scoring a sentence with `getScoredLogProb`. */
+export interface ScoredLogProb {
+    score: number;
+    scoredTokenIds: number[];
+    span: Span;
+    contextTokenHash: string;
+    provenance: ScoredLogProbProvenance;
+    /**
+     * How the score was accumulated (Task 3's `SequenceLogProbMethod`). Optional for the
+     * same backward-compatibility reason as `provenance.templateStrategy`; every real
+     * measurement path sets it explicitly.
+     */
+    method?: SequenceLogProbMethod;
+}
+//# sourceMappingURL=scored-log-prob-types.d.ts.map

package/dist/inference/scored-log-prob-types.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"scored-log-prob-types.d.ts","sourceRoot":"","sources":["../../src/inference/scored-log-prob-types.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,KAAK,EAAE,qBAAqB,EAAE,MAAM,UAAU,CAAC;AACtD,OAAO,KAAK,EAAE,IAAI,EAAE,MAAM,oBAAoB,CAAC;AAE/C,uEAAuE;AACvE,MAAM,WAAW,uBAAuB;IACtC,OAAO,EAAE,MAAM,CAAC;IAChB;;;;;;;OAOG;IACH,KAAK,EAAE,MAAM,CAAC;IACd;;;;;;OAMG;IACH,WAAW,EAAE,MAAM,CAAC;IACpB,OAAO,EAAE,MAAM,CAAC;IAChB,cAAc,EAAE,MAAM,CAAC;IACvB;;;;;;OAMG;IACH,gBAAgB,CAAC,EAAE,MAAM,GAAG,qBAAqB,CAAC;IAClD;mFAC+E;IAC/E,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,4FAA4F;IAC5F,YAAY,CAAC,EAAE,MAAM,CAAC;CACvB;AAED,iEAAiE;AACjE,MAAM,WAAW,aAAa;IAC5B,KAAK,EAAE,MAAM,CAAC;IACd,cAAc,EAAE,MAAM,EAAE,CAAC;IACzB,IAAI,EAAE,IAAI,CAAC;IACX,gBAAgB,EAAE,MAAM,CAAC;IACzB,UAAU,EAAE,uBAAuB,CAAC;IACpC;;;;OAIG;IACH,MAAM,CAAC,EAAE,qBAAqB,CAAC;CAChC"}

package/dist/inference/scored-log-prob-types.js

@@ -0,0 +1,12 @@
+/**
+ * Shared result types for `SequenceLogProbProvider.getScoredLogProb`.
+ *
+ * Pulled out of `transformers-inference.ts` (Messmotor-Sanierung Task 5) so that
+ * `template-scoring.ts` (the Gemma-4 template-mode scorer) and `transformers-inference.ts`
+ * (which loads models and picks which mode to use) can both depend on these types without
+ * either module importing from the other — avoids a circular import between the two.
+ *
+ * `transformers-inference.ts` re-exports both types from here, so existing imports of
+ * `ScoredLogProb`/`ScoredLogProbProvenance` from `"./transformers-inference"` keep working.
+ */
+export {};

package/dist/inference/sequence-logprob.d.ts

@@ -0,0 +1,34 @@
+/**
+ * Pure, offline-testable log-probability scorer for autoregressive language models.
+ *
+ * Scoring convention (binding):
+ * - A span [start, end) is half-open over token positions of the FULL fed sequence
+ *   (i.e. including any context tokens such as a prepended BOS).
+ * - The token AT position j is scored using the logits AT position j-1 (the model's
+ *   prediction distribution for "what comes next" is produced one position before the
+ *   token it predicts).
+ * - The result is a SUM of log-probabilities, not a mean.
+ * - batch_size must be exactly 1; anything else is a caller error and throws.
+ */
+/** Minimal shape this module needs from an onnxruntime-style Tensor. */
+export interface LogitsTensorLike {
+    /** Flattened [batch, seqLen, vocabSize] logits. Must be Float32Array or Float64Array. */
+    data: Float32Array | Float64Array;
+    /** [batchSize, seqLen, vocabSize] */
+    dims: readonly number[];
+}
+/** Half-open token-position span [start, end). */
+export interface Span {
+    start: number;
+    end: number;
+}
+/**
+ * Sum the log-probabilities of the tokens in `tokenIds` over the half-open span
+ * [span.start, span.end), where the token at position j is scored from the logits
+ * distribution at position j-1.
+ *
+ * Throws on: unsupported dtype, wrong dims shape, batch_size !== 1, tokenIds/seqLen
+ * mismatch, out-of-bounds span, out-of-vocab token ids, or non-finite logits.
+ */
+export declare function sumSequenceLogProb(logits: LogitsTensorLike, tokenIds: ArrayLike<number | bigint>, span: Span): number;
+//# sourceMappingURL=sequence-logprob.d.ts.map

package/dist/inference/sequence-logprob.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"sequence-logprob.d.ts","sourceRoot":"","sources":["../../src/inference/sequence-logprob.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAEH,wEAAwE;AACxE,MAAM,WAAW,gBAAgB;IAC/B,yFAAyF;IACzF,IAAI,EAAE,YAAY,GAAG,YAAY,CAAC;IAClC,qCAAqC;IACrC,IAAI,EAAE,SAAS,MAAM,EAAE,CAAC;CACzB;AAED,kDAAkD;AAClD,MAAM,WAAW,IAAI;IACnB,KAAK,EAAE,MAAM,CAAC;IACd,GAAG,EAAE,MAAM,CAAC;CACb;AAoED;;;;;;;GAOG;AACH,wBAAgB,kBAAkB,CAChC,MAAM,EAAE,gBAAgB,EACxB,QAAQ,EAAE,SAAS,CAAC,MAAM,GAAG,MAAM,CAAC,EACpC,IAAI,EAAE,IAAI,GACT,MAAM,CA8BR"}

package/dist/inference/sequence-logprob.js

@@ -0,0 +1,87 @@
+/**
+ * Pure, offline-testable log-probability scorer for autoregressive language models.
+ *
+ * Scoring convention (binding):
+ * - A span [start, end) is half-open over token positions of the FULL fed sequence
+ *   (i.e. including any context tokens such as a prepended BOS).
+ * - The token AT position j is scored using the logits AT position j-1 (the model's
+ *   prediction distribution for "what comes next" is produced one position before the
+ *   token it predicts).
+ * - The result is a SUM of log-probabilities, not a mean.
+ * - batch_size must be exactly 1; anything else is a caller error and throws.
+ */
+function assertSupportedDtype(data) {
+    if (!(data instanceof Float32Array) && !(data instanceof Float64Array)) {
+        const ctorName = data?.constructor?.name ?? typeof data;
+        throw new Error(`sumSequenceLogProb: unsupported logits dtype, expected Float32Array or Float64Array, got ${ctorName}`);
+    }
+}
+function assertBatchAndDims(dims) {
+    if (dims.length !== 3) {
+        throw new Error(`sumSequenceLogProb: expected logits dims [batch, seqLen, vocabSize], got [${dims.join(", ")}]`);
+    }
+    const batchSize = dims[0] ?? 0;
+    const seqLen = dims[1] ?? 0;
+    const vocabSize = dims[2] ?? 0;
+    if (batchSize !== 1) {
+        throw new Error(`sumSequenceLogProb: batch_size must be 1, got ${batchSize}`);
+    }
+    return [batchSize, seqLen, vocabSize];
+}
+function assertSpanWithinBounds(span, seqLen) {
+    const { start, end } = span;
+    if (!Number.isInteger(start) ||
+        !Number.isInteger(end) ||
+        start < 1 ||
+        end > seqLen ||
+        start >= end) {
+        throw new Error(`sumSequenceLogProb: invalid span [${start}, ${end}) for seqLen ${seqLen}; span must satisfy integer 1 <= start < end <= seqLen`);
+    }
+}
+function logSumExpRow(data, rowStart, vocabSize, contextLabel) {
+    let maxLogit = Number.NEGATIVE_INFINITY;
+    for (let v = 0; v < vocabSize; v++) {
+        const val = data[rowStart + v] ?? Number.NaN;
+        if (!Number.isFinite(val)) {
+            throw new Error(`sumSequenceLogProb: non-finite logit at ${contextLabel}, vocab index ${v} (value: ${val})`);
+        }
+        if (val > maxLogit)
+            maxLogit = val;
+    }
+    let sumExp = 0;
+    for (let v = 0; v < vocabSize; v++) {
+        const val = data[rowStart + v] ?? Number.NaN;
+        sumExp += Math.exp(val - maxLogit);
+    }
+    return { maxLogit, logSumExp: maxLogit + Math.log(sumExp) };
+}
+/**
+ * Sum the log-probabilities of the tokens in `tokenIds` over the half-open span
+ * [span.start, span.end), where the token at position j is scored from the logits
+ * distribution at position j-1.
+ *
+ * Throws on: unsupported dtype, wrong dims shape, batch_size !== 1, tokenIds/seqLen
+ * mismatch, out-of-bounds span, out-of-vocab token ids, or non-finite logits.
+ */
+export function sumSequenceLogProb(logits, tokenIds, span) {
+    assertSupportedDtype(logits.data);
+    const [, seqLen, vocabSize] = assertBatchAndDims(logits.dims);
+    if (tokenIds.length !== seqLen) {
+        throw new Error(`sumSequenceLogProb: tokenIds length (${tokenIds.length}) does not match logits seqLen (${seqLen})`);
+    }
+    assertSpanWithinBounds(span, seqLen);
+    const { data } = logits;
+    let total = 0;
+    for (let j = span.start; j < span.end; j++) {
+        const targetId = Number(tokenIds[j]);
+        if (!Number.isInteger(targetId) || targetId < 0 || targetId >= vocabSize) {
+            throw new Error(`sumSequenceLogProb: token id ${targetId} at position ${j} out of vocab range [0, ${vocabSize})`);
+        }
+        const pos = j - 1;
+        const rowStart = pos * vocabSize;
+        const { logSumExp } = logSumExpRow(data, rowStart, vocabSize, `position ${pos}`);
+        const targetLogit = data[rowStart + targetId] ?? Number.NaN;
+        total += targetLogit - logSumExp;
+    }
+    return total;
+}