ai-model-verifier 1.4.0 → 2.0.0

package/dist/rules/probe-ladder.js

@@ -0,0 +1,156 @@
+/**
+ * The behavioural ladder over the four probes, one rule per rung, in the order
+ * the table lists them. Each rule reads every probe's outcome, which is why the
+ * probes are data and the rules are here.
+ */
+import { detectSubstitution } from "../models/substitution";
+import { detectServedModelMismatch, detectTierMismatch } from "../models/tiers";
+import { defineRule } from "./types";
+const QUORUM = 3;
+const labelsWith = (probes, sig) => probes
+    .filter((r) => r.signal === sig)
+    .map((r) => r.label)
+    .join(", ");
+const signalRule = (id, signal, prefix) => defineRule({
+    id,
+    layer: "probe",
+    needs: ["probes"],
+    applies: () => true,
+    judge: async (ctx) => {
+        const labels = labelsWith(await ctx.evidence("probes"), signal);
+        return labels
+            ? { severity: "fail", reason: `${prefix}: ${labels}` }
+            : null;
+    },
+});
+export const codingToolRule = signalRule("coding-tool", "coding-tool", "coding-tool-refusal");
+export const scamRule = signalRule("scam", "scam", "scam-page");
+export const cjkLeakRule = signalRule("cjk-leak", "cjk-leak", "cjk-language-leak");
+/** Two probes whose reply never carried their nonce: the proxy mixes responses. */
+export const muxRule = defineRule({
+    id: "mux",
+    layer: "probe",
+    needs: ["probes"],
+    applies: () => true,
+    judge: async (ctx) => {
+        const labels = (await ctx.evidence("probes"))
+            .filter((r) => r.muxFailure)
+            .map((r) => r.label);
+        return labels.length >= 2
+            ? {
+                severity: "fail",
+                reason: `unsafe-proxy: response-mixing on ${labels.join(", ")}`,
+            }
+            : null;
+    },
+});
+/**
+ * A foreign vendor on EITHER identity probe is a hard fail: a real model never
+ * names a competitor as its maker. The reason lists every foreign probe.
+ */
+export const foreignRule = defineRule({
+    id: "foreign",
+    layer: "probe",
+    needs: ["probes"],
+    applies: () => true,
+    judge: async (ctx) => {
+        const probes = await ctx.evidence("probes");
+        const onIdentity = probes.some((r) => r.signal === "foreign" &&
+            (r.label === "model-name" || r.label === "identity"));
+        return onIdentity
+            ? {
+                severity: "fail",
+                reason: `foreign-identity: ${labelsWith(probes, "foreign")}`,
+            }
+            : null;
+    },
+});
+/** The reply's own `model` field names another tier than the one requested. */
+export const servedModelMismatchRule = defineRule({
+    id: "served-model-mismatch",
+    layer: "probe",
+    needs: ["probes"],
+    applies: (ctx) => ctx.wire.tiers !== null,
+    judge: async (ctx) => {
+        const probes = await ctx.evidence("probes");
+        const served = detectServedModelMismatch(ctx.model, probes.map((r) => r.detectedModel), ctx.wire.tiers ?? []);
+        return served
+            ? {
+                severity: "fail",
+                reason: `served-model-mismatch: requested ${ctx.model}, response model ${served}`,
+                data: { served },
+            }
+            : null;
+    },
+});
+/** The reply's `model` field is not the requested model at all (lenient spelling). */
+export const substitutedRule = defineRule({
+    id: "substituted",
+    layer: "probe",
+    needs: ["probes"],
+    applies: () => true,
+    judge: async (ctx) => {
+        const probes = await ctx.evidence("probes");
+        const detected = probes.find((r) => r.detectedModel)?.detectedModel ?? null;
+        const served = detectSubstitution(ctx.model, detected);
+        return served
+            ? { severity: "fail", reason: `substituted:${served}`, data: { served } }
+            : null;
+    },
+});
+/**
+ * Past the hard signals, the remaining failures are quality or transport. A
+ * shortfall driven only by transient upstream errors is inconclusive: a real
+ * lane that got rate limited mid-probe stays eligible and is probed again.
+ */
+export const quorumRule = defineRule({
+    id: "quorum",
+    layer: "probe",
+    needs: ["probes"],
+    applies: () => true,
+    judge: async (ctx) => {
+        const probes = await ctx.evidence("probes");
+        const failed = probes.filter((r) => !r.pass);
+        if (probes.length - failed.length >= QUORUM)
+            return null;
+        const labels = failed.map((r) => r.label).join(", ");
+        const transientFails = failed.filter((r) => r.transient).length;
+        const nonTransientFails = failed.filter((r) => !r.transient && r.signal !== null).length;
+        if (transientFails > 0 && nonTransientFails === 0)
+            return { severity: "inconclusive", reason: `transient: ${labels}` };
+        const blank = failed.filter((r) => r.signal === "blank").length;
+        if (blank === failed.length)
+            return { severity: "suspect", reason: `blank-response: ${labels}` };
+        const mux = failed.filter((r) => r.muxFailure).map((r) => r.label);
+        if (mux.length > 0)
+            return {
+                severity: "suspect",
+                reason: `unsafe-proxy: response-mixing on ${mux.join(", ")} (with ${labels} failing)`,
+            };
+        return { severity: "suspect", reason: `failed: ${labels}` };
+    },
+});
+/**
+ * What the model calls itself is coachable and unstable: a kiro pool's system
+ * prompt makes a real opus-4-6 answer "Claude 3.7 Sonnet" on one probe and
+ * refuse on the next. Logged only; the reply's model field and the tokenizer
+ * fingerprint are the evidence that fails a lane.
+ */
+export const tierSelfReportRule = defineRule({
+    id: "tier-self-report",
+    layer: "note",
+    needs: ["probes"],
+    applies: (ctx) => ctx.wire.tiers !== null,
+    judge: async (ctx) => {
+        const probes = await ctx.evidence("probes");
+        const claimed = detectTierMismatch(ctx.model, probes.find((r) => r.label === "model-name")?.text, ctx.wire.tiers ?? []);
+        return claimed
+            ? {
+                severity: "note",
+                reason: `tier-self-report: claims ${claimed}`,
+                data: { claimed },
+            }
+            : null;
+    },
+});
+//# sourceMappingURL=probe-ladder.js.map

package/dist/rules/probe-ladder.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"probe-ladder.js","sourceRoot":"","sources":["../../src/rules/probe-ladder.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAIH,OAAO,EAAE,kBAAkB,EAAE,MAAM,wBAAwB,CAAC;AAC5D,OAAO,EAAE,yBAAyB,EAAE,kBAAkB,EAAE,MAAM,iBAAiB,CAAC;AAChF,OAAO,EAAE,UAAU,EAAE,MAAM,SAAS,CAAC;AAErC,MAAM,MAAM,GAAG,CAAC,CAAC;AAEjB,MAAM,UAAU,GAAG,CAAC,MAA4B,EAAE,GAAgB,EAAE,EAAE,CACpE,MAAM;KACH,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,MAAM,KAAK,GAAG,CAAC;KAC/B,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,CAAC;KACnB,IAAI,CAAC,IAAI,CAAC,CAAC;AAEhB,MAAM,UAAU,GAAG,CACjB,EAAM,EACN,MAAmB,EACnB,MAAc,EACd,EAAE,CACF,UAAU,CAAC;IACT,EAAE;IACF,KAAK,EAAE,OAAO;IACd,KAAK,EAAE,CAAC,QAAQ,CAAC;IACjB,OAAO,EAAE,GAAG,EAAE,CAAC,IAAI;IACnB,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,EAAE;QACnB,MAAM,MAAM,GAAG,UAAU,CAAC,MAAM,GAAG,CAAC,QAAQ,CAAC,QAAQ,CAAC,EAAE,MAAM,CAAC,CAAC;QAChE,OAAO,MAAM;YACX,CAAC,CAAC,EAAE,QAAQ,EAAE,MAAM,EAAE,MAAM,EAAE,GAAG,MAAM,KAAK,MAAM,EAAE,EAAE;YACtD,CAAC,CAAC,IAAI,CAAC;IACX,CAAC;CACF,CAAC,CAAC;AAEL,MAAM,CAAC,MAAM,cAAc,GAAG,UAAU,CACtC,aAAa,EACb,aAAa,EACb,qBAAqB,CACtB,CAAC;AACF,MAAM,CAAC,MAAM,QAAQ,GAAG,UAAU,CAAC,MAAM,EAAE,MAAM,EAAE,WAAW,CAAC,CAAC;AAChE,MAAM,CAAC,MAAM,WAAW,GAAG,UAAU,CACnC,UAAU,EACV,UAAU,EACV,mBAAmB,CACpB,CAAC;AAEF,mFAAmF;AACnF,MAAM,CAAC,MAAM,OAAO,GAAG,UAAU,CAAC;IAChC,EAAE,EAAE,KAAK;IACT,KAAK,EAAE,OAAO;IACd,KAAK,EAAE,CAAC,QAAQ,CAAC;IACjB,OAAO,EAAE,GAAG,EAAE,CAAC,IAAI;IACnB,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,EAAE;QACnB,MAAM,MAAM,GAAG,CAAC,MAAM,GAAG,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC;aAC1C,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,UAAU,CAAC;aAC3B,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC;QACvB,OAAO,MAAM,CAAC,MAAM,IAAI,CAAC;YACvB,CAAC,CAAC;gBACE,QAAQ,EAAE,MAAM;gBAChB,MAAM,EAAE,oCAAoC,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE;aAChE;YACH,CAAC,CAAC,IAAI,CAAC;IACX,CAAC;CACF,CAAC,CAAC;AAEH;;;GAGG;AACH,MAAM,CAAC,MAAM,WAAW,GAAG,UAAU,CAAC;IACpC,EAAE,EAAE,SAAS;IACb,KAAK,EAAE,OAAO;IACd,KAAK,EAAE,CAAC,QAAQ,CAAC;IACjB,OAAO,EAAE,GAAG,EAAE,CAAC,IAAI;IACnB,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,EAAE;QACnB,MAAM,MAAM,GAAG,MAAM,GAAG,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC;QAC5C,MAAM,UAAU,GAAG,MAAM,CAAC,IAAI,CAC5B,CAAC,CAAC,EAAE,EAAE,CACJ,CAAC,CAAC,MAAM,KAAK,SAAS;YACtB,CAAC,CAAC,CAAC,KAAK,KAAK,YAAY,IAAI,CAAC,CAAC,KAAK,KAAK,UAAU,CAAC,CACvD,CAAC;QACF,OAAO,UAAU;YACf,CAAC,CAAC;gBACE,QAAQ,EAAE,MAAM;gBAChB,MAAM,EAAE,qBAAqB,UAAU,CAAC,MAAM,EAAE,SAAS,CAAC,EAAE;aAC7D;YACH,CAAC,CAAC,IAAI,CAAC;IACX,CAAC;CACF,CAAC,CAAC;AAEH,+EAA+E;AAC/E,MAAM,CAAC,MAAM,uBAAuB,GAAG,UAAU,CAAC;IAChD,EAAE,EAAE,uBAAuB;IAC3B,KAAK,EAAE,OAAO;IACd,KAAK,EAAE,CAAC,QAAQ,CAAC;IACjB,OAAO,EAAE,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,IAAI,CAAC,KAAK,KAAK,IAAI;IACzC,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,EAAE;QACnB,MAAM,MAAM,GAAG,MAAM,GAAG,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC;QAC5C,MAAM,MAAM,GAAG,yBAAyB,CACtC,GAAG,CAAC,KAAK,EACT,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,aAAa,CAAC,EAClC,GAAG,CAAC,IAAI,CAAC,KAAK,IAAI,EAAE,CACrB,CAAC;QACF,OAAO,MAAM;YACX,CAAC,CAAC;gBACE,QAAQ,EAAE,MAAM;gBAChB,MAAM,EAAE,oCAAoC,GAAG,CAAC,KAAK,oBAAoB,MAAM,EAAE;gBACjF,IAAI,EAAE,EAAE,MAAM,EAAE;aACjB;YACH,CAAC,CAAC,IAAI,CAAC;IACX,CAAC;CACF,CAAC,CAAC;AAEH,sFAAsF;AACtF,MAAM,CAAC,MAAM,eAAe,GAAG,UAAU,CAAC;IACxC,EAAE,EAAE,aAAa;IACjB,KAAK,EAAE,OAAO;IACd,KAAK,EAAE,CAAC,QAAQ,CAAC;IACjB,OAAO,EAAE,GAAG,EAAE,CAAC,IAAI;IACnB,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,EAAE;QACnB,MAAM,MAAM,GAAG,MAAM,GAAG,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC;QAC5C,MAAM,QAAQ,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,aAAa,CAAC,EAAE,aAAa,IAAI,IAAI,CAAC;QAC5E,MAAM,MAAM,GAAG,kBAAkB,CAAC,GAAG,CAAC,KAAK,EAAE,QAAQ,CAAC,CAAC;QACvD,OAAO,MAAM;YACX,CAAC,CAAC,EAAE,QAAQ,EAAE,MAAM,EAAE,MAAM,EAAE,eAAe,MAAM,EAAE,EAAE,IAAI,EAAE,EAAE,MAAM,EAAE,EAAE;YACzE,CAAC,CAAC,IAAI,CAAC;IACX,CAAC;CACF,CAAC,CAAC;AAEH;;;;GAIG;AACH,MAAM,CAAC,MAAM,UAAU,GAAG,UAAU,CAAC;IACnC,EAAE,EAAE,QAAQ;IACZ,KAAK,EAAE,OAAO;IACd,KAAK,EAAE,CAAC,QAAQ,CAAC;IACjB,OAAO,EAAE,GAAG,EAAE,CAAC,IAAI;IACnB,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,EAAE;QACnB,MAAM,MAAM,GAAG,MAAM,GAAG,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC;QAC5C,MAAM,MAAM,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC;QAC7C,IAAI,MAAM,CAAC,MAAM,GAAG,MAAM,CAAC,MAAM,IAAI,MAAM;YAAE,OAAO,IAAI,CAAC;QACzD,MAAM,MAAM,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACrD,MAAM,cAAc,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,MAAM,CAAC;QAChE,MAAM,iBAAiB,GAAG,MAAM,CAAC,MAAM,CACrC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,SAAS,IAAI,CAAC,CAAC,MAAM,KAAK,IAAI,CACzC,CAAC,MAAM,CAAC;QACT,IAAI,cAAc,GAAG,CAAC,IAAI,iBAAiB,KAAK,CAAC;YAC/C,OAAO,EAAE,QAAQ,EAAE,cAAc,EAAE,MAAM,EAAE,cAAc,MAAM,EAAE,EAAE,CAAC;QACtE,MAAM,KAAK,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,MAAM,KAAK,OAAO,CAAC,CAAC,MAAM,CAAC;QAChE,IAAI,KAAK,KAAK,MAAM,CAAC,MAAM;YACzB,OAAO,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,EAAE,mBAAmB,MAAM,EAAE,EAAE,CAAC;QACtE,MAAM,GAAG,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC;QACnE,IAAI,GAAG,CAAC,MAAM,GAAG,CAAC;YAChB,OAAO;gBACL,QAAQ,EAAE,SAAS;gBACnB,MAAM,EAAE,oCAAoC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,UAAU,MAAM,WAAW;aACtF,CAAC;QACJ,OAAO,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,EAAE,WAAW,MAAM,EAAE,EAAE,CAAC;IAC9D,CAAC;CACF,CAAC,CAAC;AAEH;;;;;GAKG;AACH,MAAM,CAAC,MAAM,kBAAkB,GAAG,UAAU,CAAC;IAC3C,EAAE,EAAE,kBAAkB;IACtB,KAAK,EAAE,MAAM;IACb,KAAK,EAAE,CAAC,QAAQ,CAAC;IACjB,OAAO,EAAE,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,IAAI,CAAC,KAAK,KAAK,IAAI;IACzC,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,EAAE;QACnB,MAAM,MAAM,GAAG,MAAM,GAAG,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC;QAC5C,MAAM,OAAO,GAAG,kBAAkB,CAChC,GAAG,CAAC,KAAK,EACT,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,KAAK,YAAY,CAAC,EAAE,IAAI,EAClD,GAAG,CAAC,IAAI,CAAC,KAAK,IAAI,EAAE,CACrB,CAAC;QACF,OAAO,OAAO;YACZ,CAAC,CAAC;gBACE,QAAQ,EAAE,MAAM;gBAChB,MAAM,EAAE,4BAA4B,OAAO,EAAE;gBAC7C,IAAI,EAAE,EAAE,OAAO,EAAE;aAClB;YACH,CAAC,CAAC,IAAI,CAAC;IACX,CAAC;CACF,CAAC,CAAC"}

package/dist/{detectors → rules}/response-metadata.d.ts

@@ -1,21 +1,16 @@
 /**
  * Passive response metadata: what the envelope reveals about who really
- * answered, read from responses the run already made.
+ * answered, read from replies the run already made.
  *
- * Costs nothing (no extra requests) and asserts nothing. Everything here is
- * reported as observation, because each signal has an honest explanation as
- * well as a dishonest one:
- *
- *  - a `chatcmpl-` id inside an Anthropic-shaped reply means a translation
- *    layer, which is what our own gateway is;
- *  - OpenAI-style `usage` keys in a Messages reply mean the same;
- *  - a missing id means a sloppy relay, not a fake model.
- *
- * The value is in the aggregate: over many lanes, "which of my Claude lanes
- * answer with a chatcmpl id" is a question worth being able to ask.
+ * Costs nothing and asserts nothing. Everything here is reported as
+ * observation, because each signal has an honest explanation as well as a
+ * dishonest one: a `chatcmpl-` id inside an Anthropic-shaped reply means a
+ * translation layer, which is what our own gateway is; a missing id means a
+ * sloppy relay, not a fake model. The value is in the aggregate: over many
+ * lanes, "which of my Claude lanes answer with a chatcmpl id" is a question
+ * worth being able to ask.
  */
-/** Vendor whose response shape the envelope actually looks like. */
-export type EnvelopeShape = "anthropic" | "openai" | "gemini" | "unknown";
+import type { EnvelopeShape } from "../vendors/types";
 export type ResponseMetadata = {
     /** Shape the payload matches, regardless of which API was called. */
     shape: EnvelopeShape;
@@ -33,4 +28,6 @@ export type ResponseMetadata = {
     notes: string[];
 };
 export declare function readResponseMetadata(data: unknown, expected: EnvelopeShape): ResponseMetadata;
+/** Free metadata from the richest probe reply, read as the wire's shape. */
+export declare const envelopeRule: import("./types").Rule<"envelope", "probes", "note">;
 //# sourceMappingURL=response-metadata.d.ts.map

package/dist/rules/response-metadata.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"response-metadata.d.ts","sourceRoot":"","sources":["../../src/rules/response-metadata.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAGH,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,kBAAkB,CAAC;AAGtD,MAAM,MAAM,gBAAgB,GAAG;IAC7B,qEAAqE;IACrE,KAAK,EAAE,aAAa,CAAC;IACrB,6CAA6C;IAC7C,QAAQ,EAAE,aAAa,CAAC;IACxB,iEAAiE;IACjE,UAAU,EAAE,OAAO,CAAC;IACpB,4EAA4E;IAC5E,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAC;IACxB,yEAAyE;IACzE,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,8EAA8E;IAC9E,gBAAgB,EAAE,MAAM,EAAE,CAAC;IAC3B,kEAAkE;IAClE,KAAK,EAAE,MAAM,EAAE,CAAC;CACjB,CAAC;AAmDF,wBAAgB,oBAAoB,CAClC,IAAI,EAAE,OAAO,EACb,QAAQ,EAAE,aAAa,GACtB,gBAAgB,CAyClB;AAED,4EAA4E;AAC5E,eAAO,MAAM,YAAY,sDAiBvB,CAAC"}

package/dist/{detectors → rules}/response-metadata.js

@@ -1,19 +1,17 @@
 /**
  * Passive response metadata: what the envelope reveals about who really
- * answered, read from responses the run already made.
+ * answered, read from replies the run already made.
  *
- * Costs nothing (no extra requests) and asserts nothing. Everything here is
- * reported as observation, because each signal has an honest explanation as
- * well as a dishonest one:
- *
- *  - a `chatcmpl-` id inside an Anthropic-shaped reply means a translation
- *    layer, which is what our own gateway is;
- *  - OpenAI-style `usage` keys in a Messages reply mean the same;
- *  - a missing id means a sloppy relay, not a fake model.
- *
- * The value is in the aggregate: over many lanes, "which of my Claude lanes
- * answer with a chatcmpl id" is a question worth being able to ask.
+ * Costs nothing and asserts nothing. Everything here is reported as
+ * observation, because each signal has an honest explanation as well as a
+ * dishonest one: a `chatcmpl-` id inside an Anthropic-shaped reply means a
+ * translation layer, which is what our own gateway is; a missing id means a
+ * sloppy relay, not a fake model. The value is in the aggregate: over many
+ * lanes, "which of my Claude lanes answer with a chatcmpl id" is a question
+ * worth being able to ask.
  */
+import { richestProbe } from "../engine/probe-runner";
+import { defineRule } from "./types";
 const ANTHROPIC_USAGE = new Set([
     "input_tokens",
     "output_tokens",
@@ -42,11 +40,14 @@ function idPrefixOf(id) {
     return cut > 0 ? id.slice(0, cut + 1) : "(none)";
 }
 function shapeOf(data, usageKeys) {
-    if (data["candidates"] !== undefined || usageKeys.some((k) => GEMINI_USAGE.has(k)))
+    if (data["candidates"] !== undefined ||
+        usageKeys.some((k) => GEMINI_USAGE.has(k)))
         return "gemini";
-    if (data["choices"] !== undefined || usageKeys.some((k) => OPENAI_USAGE.has(k)))
+    if (data["choices"] !== undefined ||
+        usageKeys.some((k) => OPENAI_USAGE.has(k)))
         return "openai";
-    if (data["content"] !== undefined || usageKeys.some((k) => ANTHROPIC_USAGE.has(k)))
+    if (data["content"] !== undefined ||
+        usageKeys.some((k) => ANTHROPIC_USAGE.has(k)))
         return "anthropic";
     return "unknown";
 }
@@ -88,4 +89,20 @@ export function readResponseMetadata(data, expected) {
         notes,
     };
 }
+/** Free metadata from the richest probe reply, read as the wire's shape. */
+export const envelopeRule = defineRule({
+    id: "envelope",
+    layer: "note",
+    needs: ["probes"],
+    applies: () => true,
+    judge: async () => null,
+    report: async (ctx) => {
+        const richest = richestProbe(await ctx.evidence("probes"));
+        return richest
+            ? {
+                responseMetadata: readResponseMetadata(richest.raw, ctx.wire.envelope),
+            }
+            : {};
+    },
+});
 //# sourceMappingURL=response-metadata.js.map

package/dist/rules/response-metadata.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"response-metadata.js","sourceRoot":"","sources":["../../src/rules/response-metadata.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAEH,OAAO,EAAE,YAAY,EAAE,MAAM,wBAAwB,CAAC;AAEtD,OAAO,EAAE,UAAU,EAAE,MAAM,SAAS,CAAC;AAmBrC,MAAM,eAAe,GAAG,IAAI,GAAG,CAAC;IAC9B,cAAc;IACd,eAAe;IACf,yBAAyB;IACzB,6BAA6B;CAC9B,CAAC,CAAC;AACH,MAAM,YAAY,GAAG,IAAI,GAAG,CAAC;IAC3B,eAAe;IACf,mBAAmB;IACnB,cAAc;CACf,CAAC,CAAC;AACH,MAAM,YAAY,GAAG,IAAI,GAAG,CAAC;IAC3B,kBAAkB;IAClB,sBAAsB;IACtB,iBAAiB;IACjB,oBAAoB;CACrB,CAAC,CAAC;AAEH,SAAS,UAAU,CAAC,EAAW;IAC7B,IAAI,OAAO,EAAE,KAAK,QAAQ,IAAI,EAAE,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,IAAI,CAAC;IAC3D,KAAK,MAAM,CAAC,IAAI,CAAC,MAAM,EAAE,WAAW,EAAE,OAAO,EAAE,OAAO,EAAE,QAAQ,CAAC;QAC/D,IAAI,EAAE,CAAC,UAAU,CAAC,CAAC,CAAC;YAAE,OAAO,CAAC,CAAC;IACjC,sEAAsE;IACtE,MAAM,GAAG,GAAG,EAAE,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;IAC9B,OAAO,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC;AACnD,CAAC;AAED,SAAS,OAAO,CACd,IAA6B,EAC7B,SAAmB;IAEnB,IACE,IAAI,CAAC,YAAY,CAAC,KAAK,SAAS;QAChC,SAAS,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,YAAY,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;QAE1C,OAAO,QAAQ,CAAC;IAClB,IACE,IAAI,CAAC,SAAS,CAAC,KAAK,SAAS;QAC7B,SAAS,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,YAAY,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;QAE1C,OAAO,QAAQ,CAAC;IAClB,IACE,IAAI,CAAC,SAAS,CAAC,KAAK,SAAS;QAC7B,SAAS,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,eAAe,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;QAE7C,OAAO,WAAW,CAAC;IACrB,OAAO,SAAS,CAAC;AACnB,CAAC;AAED,MAAM,UAAU,oBAAoB,CAClC,IAAa,EACb,QAAuB;IAEvB,MAAM,GAAG,GAAG,CAAC,IAAI,IAAI,EAAE,CAA4B,CAAC;IACpD,MAAM,KAAK,GAAG,CAAC,GAAG,CAAC,OAAO,CAAC,IAAI,EAAE,CAA4B,CAAC;IAC9D,MAAM,SAAS,GAAG,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,IAAI,EAAE,CAAC;IAC5C,MAAM,KAAK,GAAG,OAAO,CAAC,GAAG,EAAE,SAAS,CAAC,CAAC;IACtC,MAAM,QAAQ,GAAG,UAAU,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC;IACvC,MAAM,KAAK,GAAa,EAAE,CAAC;IAE3B,wEAAwE;IACxE,MAAM,gBAAgB,GAAG,SAAS,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE;QAC9C,IAAI,CAAC,CAAC,UAAU,CAAC,SAAS,CAAC;YAAE,OAAO,KAAK,KAAK,WAAW,CAAC;QAC1D,IAAI,CAAC,CAAC,UAAU,CAAC,SAAS,CAAC;YAAE,OAAO,KAAK,KAAK,QAAQ,CAAC;QACvD,IAAI,eAAe,CAAC,GAAG,CAAC,CAAC,CAAC;YAAE,OAAO,KAAK,KAAK,WAAW,CAAC;QACzD,IAAI,YAAY,CAAC,GAAG,CAAC,CAAC,CAAC;YAAE,OAAO,KAAK,KAAK,QAAQ,CAAC;QACnD,IAAI,YAAY,CAAC,GAAG,CAAC,CAAC,CAAC;YAAE,OAAO,KAAK,KAAK,QAAQ,CAAC;QACnD,OAAO,KAAK,CAAC;IACf,CAAC,CAAC,CAAC;IAEH,MAAM,UAAU,GAAG,KAAK,KAAK,SAAS,IAAI,KAAK,KAAK,QAAQ,CAAC;IAC7D,IAAI,UAAU;QACZ,KAAK,CAAC,IAAI,CACR,eAAe,KAAK,iBAAiB,QAAQ,iDAAiD,CAC/F,CAAC;IACJ,IAAI,QAAQ,KAAK,WAAW,IAAI,QAAQ,IAAI,QAAQ,KAAK,MAAM;QAC7D,KAAK,CAAC,IAAI,CACR,cAAc,QAAQ,oEAAoE,CAC3F,CAAC;IACJ,IAAI,gBAAgB,CAAC,MAAM,GAAG,CAAC;QAC7B,KAAK,CAAC,IAAI,CACR,iBAAiB,gBAAgB,CAAC,IAAI,CAAC,IAAI,CAAC,wCAAwC,CACrF,CAAC;IAEJ,OAAO;QACL,KAAK;QACL,QAAQ;QACR,UAAU;QACV,QAAQ;QACR,SAAS;QACT,gBAAgB;QAChB,KAAK;KACN,CAAC;AACJ,CAAC;AAED,4EAA4E;AAC5E,MAAM,CAAC,MAAM,YAAY,GAAG,UAAU,CAAC;IACrC,EAAE,EAAE,UAAU;IACd,KAAK,EAAE,MAAM;IACb,KAAK,EAAE,CAAC,QAAQ,CAAC;IACjB,OAAO,EAAE,GAAG,EAAE,CAAC,IAAI;IACnB,KAAK,EAAE,KAAK,IAAI,EAAE,CAAC,IAAI;IACvB,MAAM,EAAE,KAAK,EAAE,GAAG,EAAE,EAAE;QACpB,MAAM,OAAO,GAAG,YAAY,CAAC,MAAM,GAAG,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC,CAAC;QAC3D,OAAO,OAAO;YACZ,CAAC,CAAC;gBACE,gBAAgB,EAAE,oBAAoB,CACpC,OAAO,CAAC,GAAG,EACX,GAAG,CAAC,IAAI,CAAC,QAAQ,CAClB;aACF;YACH,CAAC,CAAC,EAAE,CAAC;IACT,CAAC;CACF,CAAC,CAAC"}

package/dist/rules/table.d.ts

@@ -0,0 +1,27 @@
+import type { Rule } from "./types";
+/**
+ * Precedence is this order: the first finding that is not a note decides the
+ * verdict. Evidence rules sit above the probe ladder because they read facts a
+ * coached reply cannot fake; note rules only report.
+ */
+export declare const RULES: readonly [Rule<"thinking-floor", "floorReply", "evidence">, Rule<"tokenizer-fingerprint", "fixedText", "evidence">, Rule<"coding-tool", "probes", "probe">, Rule<"scam", "probes", "probe">, Rule<"cjk-leak", "probes", "probe">, Rule<"mux", "probes", "probe">, Rule<"foreign", "probes", "probe">, Rule<"served-model-mismatch", "probes", "probe">, Rule<"substituted", "probes", "probe">, Rule<"quorum", "probes", "probe">, Rule<"tier-self-report", "probes", "note">, Rule<"signature", "thinkingReply", "note">, Rule<"token-truth", "countTokens" | "fixedText", "note">, Rule<"envelope", "probes", "note">, Rule<"throughput", "probes", "note">];
+type RuleEntry = (typeof RULES)[number];
+export type RuleId = RuleEntry["id"];
+/** Rules that can decide a verdict; a UI naming every rule keys off this. */
+export type VerdictRuleId = Extract<RuleEntry, {
+    layer: "evidence" | "probe";
+}>["id"];
+export declare const DETECTION_RULES: readonly VerdictRuleId[];
+export declare const DETECTION_EXCEPTIONS: readonly ["version", "transient", "cloud-host", "reshaping", "threshold"];
+export type DetectionExceptionId = (typeof DETECTION_EXCEPTIONS)[number];
+/** The rule a probe's signal feeds, for linking a probe row to a rule. */
+export declare const RULE_FOR_SIGNAL: {
+    readonly "coding-tool": "coding-tool";
+    readonly scam: "scam";
+    readonly "cjk-leak": "cjk-leak";
+    readonly foreign: "foreign";
+    readonly "cloud-host": null;
+    readonly blank: "quorum";
+};
+export {};
+//# sourceMappingURL=table.d.ts.map

package/dist/rules/table.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"table.d.ts","sourceRoot":"","sources":["../../src/rules/table.ts"],"names":[],"mappings":"AAkBA,OAAO,KAAK,EAAE,IAAI,EAAE,MAAM,SAAS,CAAC;AAEpC;;;;GAIG;AACH,eAAO,MAAM,KAAK,goBAgBkB,CAAC;AAErC,KAAK,SAAS,GAAG,CAAC,OAAO,KAAK,CAAC,CAAC,MAAM,CAAC,CAAC;AACxC,MAAM,MAAM,MAAM,GAAG,SAAS,CAAC,IAAI,CAAC,CAAC;AACrC,6EAA6E;AAC7E,MAAM,MAAM,aAAa,GAAG,OAAO,CACjC,SAAS,EACT;IAAE,KAAK,EAAE,UAAU,GAAG,OAAO,CAAA;CAAE,CAChC,CAAC,IAAI,CAAC,CAAC;AAER,eAAO,MAAM,eAAe,EAAE,SAAS,aAAa,EAGlC,CAAC;AAEnB,eAAO,MAAM,oBAAoB,YAC/B,SAAS,EACT,WAAW,EACX,YAAY,EACZ,WAAW,EACX,WAAW,CACH,CAAC;AACX,MAAM,MAAM,oBAAoB,GAAG,CAAC,OAAO,oBAAoB,CAAC,CAAC,MAAM,CAAC,CAAC;AAEzE,0EAA0E;AAC1E,eAAO,MAAM,eAAe;4BACX,aAAa;mBACtB,MAAM;yBACA,UAAU;sBACb,SAAS;;oBAEX,QAAQ;CAC0D,CAAC"}

package/dist/rules/table.js

@@ -0,0 +1,47 @@
+import { cjkLeakRule, codingToolRule, foreignRule, muxRule, quorumRule, scamRule, servedModelMismatchRule, substitutedRule, tierSelfReportRule, } from "./probe-ladder";
+import { envelopeRule } from "./response-metadata";
+import { signatureRule } from "./thinking-signature";
+import { thinkingFloorRule } from "./thinking-floor";
+import { throughputRule } from "./throughput";
+import { tokenTruthRule } from "./token-truth";
+import { tokenizerFingerprintRule } from "./tokenizer-fingerprint";
+/**
+ * Precedence is this order: the first finding that is not a note decides the
+ * verdict. Evidence rules sit above the probe ladder because they read facts a
+ * coached reply cannot fake; note rules only report.
+ */
+export const RULES = [
+    thinkingFloorRule,
+    tokenizerFingerprintRule,
+    codingToolRule,
+    scamRule,
+    cjkLeakRule,
+    muxRule,
+    foreignRule,
+    servedModelMismatchRule,
+    substitutedRule,
+    quorumRule,
+    tierSelfReportRule,
+    signatureRule,
+    tokenTruthRule,
+    envelopeRule,
+    throughputRule,
+];
+export const DETECTION_RULES = RULES.filter((r) => r.layer !== "note").map((r) => r.id);
+export const DETECTION_EXCEPTIONS = [
+    "version",
+    "transient",
+    "cloud-host",
+    "reshaping",
+    "threshold",
+];
+/** The rule a probe's signal feeds, for linking a probe row to a rule. */
+export const RULE_FOR_SIGNAL = {
+    "coding-tool": "coding-tool",
+    scam: "scam",
+    "cjk-leak": "cjk-leak",
+    foreign: "foreign",
+    "cloud-host": null,
+    blank: "quorum",
+};
+//# sourceMappingURL=table.js.map

package/dist/rules/table.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"table.js","sourceRoot":"","sources":["../../src/rules/table.ts"],"names":[],"mappings":"AACA,OAAO,EACL,WAAW,EACX,cAAc,EACd,WAAW,EACX,OAAO,EACP,UAAU,EACV,QAAQ,EACR,uBAAuB,EACvB,eAAe,EACf,kBAAkB,GACnB,MAAM,gBAAgB,CAAC;AACxB,OAAO,EAAE,YAAY,EAAE,MAAM,qBAAqB,CAAC;AACnD,OAAO,EAAE,aAAa,EAAE,MAAM,sBAAsB,CAAC;AACrD,OAAO,EAAE,iBAAiB,EAAE,MAAM,kBAAkB,CAAC;AACrD,OAAO,EAAE,cAAc,EAAE,MAAM,cAAc,CAAC;AAC9C,OAAO,EAAE,cAAc,EAAE,MAAM,eAAe,CAAC;AAC/C,OAAO,EAAE,wBAAwB,EAAE,MAAM,yBAAyB,CAAC;AAGnE;;;;GAIG;AACH,MAAM,CAAC,MAAM,KAAK,GAAG;IACnB,iBAAiB;IACjB,wBAAwB;IACxB,cAAc;IACd,QAAQ;IACR,WAAW;IACX,OAAO;IACP,WAAW;IACX,uBAAuB;IACvB,eAAe;IACf,UAAU;IACV,kBAAkB;IAClB,aAAa;IACb,cAAc;IACd,YAAY;IACZ,cAAc;CACoB,CAAC;AAUrC,MAAM,CAAC,MAAM,eAAe,GAA6B,KAAK,CAAC,MAAM,CACnE,CAAC,CAAC,EAA4D,EAAE,CAC9D,CAAC,CAAC,KAAK,KAAK,MAAM,CACrB,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;AAEnB,MAAM,CAAC,MAAM,oBAAoB,GAAG;IAClC,SAAS;IACT,WAAW;IACX,YAAY;IACZ,WAAW;IACX,WAAW;CACH,CAAC;AAGX,0EAA0E;AAC1E,MAAM,CAAC,MAAM,eAAe,GAAG;IAC7B,aAAa,EAAE,aAAa;IAC5B,IAAI,EAAE,MAAM;IACZ,UAAU,EAAE,UAAU;IACtB,OAAO,EAAE,SAAS;IAClB,YAAY,EAAE,IAAI;IAClB,KAAK,EAAE,QAAQ;CAC0D,CAAC"}

package/dist/{detectors → rules}/thinking-floor.d.ts

@@ -1,5 +1,5 @@
 /**
- * Thinking-floor check for models that cannot switch reasoning off.
+ * Thinking floor for models that cannot switch reasoning off.
  *
  * Gemini 2.5 Pro has a thinking budget floor (128 tokens): even "Reply with
  * only the word ok." produces hidden thought tokens, reported as
@@ -10,24 +10,17 @@
  * prompt was 12, the native thoughts count 99 to 158.
  *
  * A flash-class backend sold under the pro name answers the same prompt with
- * `completion_tokens: 1` (a6 merchant 3623, 2026-09-09), and it still echoes
- * `model: gemini-2.5-pro`, so the envelope substitution check never fires.
- * The token count is the only free evidence left.
- *
- * Not applied to flash tiers: those may legitimately run with thinking off.
+ * `completion_tokens: 1`, and it still echoes `model: gemini-2.5-pro`, so the
+ * substitution rule never fires. The token count is the only free evidence left.
  */
-import type { TransportFn } from "../transport";
+import type { ReasoningUsage } from "../vendors/types";
+export declare const DEFAULT_MIN_COMPLETION_TOKENS = 10;
 export type ThinkingFloorOptions = {
-    /**
-     * Model id prefixes that must always think. Compared after lowercasing and
-     * mapping `.`/`_` to `-`, provider prefix stripped.
-     */
+    /** Id globs added to the always-thinks set (normalised, prefix stripped). */
     models?: readonly string[];
     /** Below this many completion tokens with no reasoning count, the reply did not think. */
     minCompletionTokens?: number;
 };
-export declare const DEFAULT_ALWAYS_THINKS: readonly string[];
-export declare const DEFAULT_MIN_COMPLETION_TOKENS = 10;
 export type ThinkingFloorState = 
 /** Reasoning tokens reported, or the completion count is above the floor. */
 "thought"
@@ -43,23 +36,15 @@ export type ThinkingFloorResult = {
     reasoningTokens: number | null;
     reason?: string;
 };
+export declare const alwaysThinksEntries: (models?: readonly string[]) => {
+    match: string;
+    alwaysThinks: boolean;
+}[];
 export declare function mustAlwaysThink(model: string, options?: ThinkingFloorOptions): boolean;
 /** Completion and reasoning counts from an OpenAI `usage` or a Gemini `usageMetadata`. */
-export declare function readThinkingUsage(data: unknown): {
-    completion: number | null;
-    reasoning: number | null;
-} | null;
+export declare function readThinkingUsage(data: unknown): ReasoningUsage | null;
+export declare function judgeReasoningUsage(usage: ReasoningUsage | null, floor: number): ThinkingFloorResult;
 /** Judge one reply to the short probe. Pure: pass any response body already collected. */
-export declare function checkThinkingFloor(model: string, data: unknown, options?: ThinkingFloorOptions): ThinkingFloorResult;
-/** One extra generation: the short prompt through the endpoint, then `checkThinkingFloor`. */
-export declare function probeThinkingFloor(opts: {
-    transport: TransportFn;
-    baseUrl: string;
-    apiKey: string;
-    model: string;
-    /** Wire shape of the endpoint. */
-    provider: "openai" | "gemini";
-    timeoutMs: number;
-    options?: ThinkingFloorOptions;
-}): Promise<ThinkingFloorResult>;
+export declare function judgeThinkingFloor(model: string, data: unknown, options?: ThinkingFloorOptions): ThinkingFloorResult;
+export declare const thinkingFloorRule: import("./types").Rule<"thinking-floor", "floorReply", "evidence">;
 //# sourceMappingURL=thinking-floor.d.ts.map

package/dist/rules/thinking-floor.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"thinking-floor.d.ts","sourceRoot":"","sources":["../../src/rules/thinking-floor.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAIH,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,kBAAkB,CAAC;AAGvD,eAAO,MAAM,6BAA6B,KAAK,CAAC;AAEhD,MAAM,MAAM,oBAAoB,GAAG;IACjC,6EAA6E;IAC7E,MAAM,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;IAC3B,0FAA0F;IAC1F,mBAAmB,CAAC,EAAE,MAAM,CAAC;CAC9B,CAAC;AAEF,MAAM,MAAM,kBAAkB;AAC5B,6EAA6E;AAC3E,SAAS;AACX,2EAA2E;GACzE,aAAa;AACf,+CAA+C;GAC7C,YAAY;AACd,+CAA+C;GAC7C,SAAS,CAAC;AAEd,MAAM,MAAM,mBAAmB,GAAG;IAChC,KAAK,EAAE,kBAAkB,CAAC;IAC1B,gBAAgB,EAAE,MAAM,GAAG,IAAI,CAAC;IAChC,eAAe,EAAE,MAAM,GAAG,IAAI,CAAC;IAC/B,MAAM,CAAC,EAAE,MAAM,CAAC;CACjB,CAAC;AAEF,eAAO,MAAM,mBAAmB,YAAY,SAAS,MAAM,EAAE;;;GACkB,CAAC;AAEhF,wBAAgB,eAAe,CAC7B,KAAK,EAAE,MAAM,EACb,OAAO,CAAC,EAAE,oBAAoB,GAC7B,OAAO,CAGT;AAED,0FAA0F;AAC1F,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,OAAO,GAAG,cAAc,GAAG,IAAI,CAWtE;AAED,wBAAgB,mBAAmB,CACjC,KAAK,EAAE,cAAc,GAAG,IAAI,EAC5B,KAAK,EAAE,MAAM,GACZ,mBAAmB,CA0CrB;AAQD,0FAA0F;AAC1F,wBAAgB,kBAAkB,CAChC,KAAK,EAAE,MAAM,EACb,IAAI,EAAE,OAAO,EACb,OAAO,CAAC,EAAE,oBAAoB,GAC7B,mBAAmB,CAMrB;AAqBD,eAAO,MAAM,iBAAiB,oEAa5B,CAAC"}

package/dist/rules/thinking-floor.js

@@ -0,0 +1,119 @@
+/**
+ * Thinking floor for models that cannot switch reasoning off.
+ *
+ * Gemini 2.5 Pro has a thinking budget floor (128 tokens): even "Reply with
+ * only the word ok." produces hidden thought tokens, reported as
+ * `completion_tokens_details.reasoning_tokens` on an OpenAI-shaped hop, folded
+ * into `completion_tokens` when the relay strips the breakdown, or as
+ * `usageMetadata.thoughtsTokenCount` on the native endpoint. Measured across 66
+ * historical probes of genuine 2.5 Pro: the lowest completion count for that
+ * prompt was 12, the native thoughts count 99 to 158.
+ *
+ * A flash-class backend sold under the pro name answers the same prompt with
+ * `completion_tokens: 1`, and it still echoes `model: gemini-2.5-pro`, so the
+ * substitution rule never fires. The token count is the only free evidence left.
+ */
+import { intOf, rec } from "../internal/utils";
+import { normalizeModelId, resolveModelFacts } from "../models/facts";
+import { defineRule } from "./types";
+export const DEFAULT_MIN_COMPLETION_TOKENS = 10;
+export const alwaysThinksEntries = (models = []) => models.map((p) => ({ match: `${normalizeModelId(p)}*`, alwaysThinks: true }));
+export function mustAlwaysThink(model, options) {
+    return resolveModelFacts(model, alwaysThinksEntries(options?.models))
+        .alwaysThinks;
+}
+/** Completion and reasoning counts from an OpenAI `usage` or a Gemini `usageMetadata`. */
+export function readThinkingUsage(data) {
+    const d = rec(data);
+    const u = rec(d?.["usage"] ?? d?.["usageMetadata"]);
+    if (!u)
+        return null;
+    const details = rec(u["completion_tokens_details"]);
+    return {
+        reasoning: intOf(details?.["reasoning_tokens"]) ?? intOf(u["thoughtsTokenCount"]),
+        completion: intOf(u["completion_tokens"]) ?? intOf(u["candidatesTokenCount"]),
+    };
+}
+export function judgeReasoningUsage(usage, floor) {
+    if (!usage)
+        return {
+            state: "unmeasured",
+            completionTokens: null,
+            reasoningTokens: null,
+            reason: "no usage in reply",
+        };
+    if (usage.reasoning !== null && usage.reasoning > 0)
+        return {
+            state: "thought",
+            completionTokens: usage.completion,
+            reasoningTokens: usage.reasoning,
+        };
+    if (usage.completion === null)
+        return {
+            state: "unmeasured",
+            completionTokens: null,
+            reasoningTokens: usage.reasoning,
+            reason: "no completion count in usage",
+        };
+    // An empty reply carries no evidence: nothing was generated, so nothing was
+    // skipped. Truncation and refusals land here, not in "no-thinking".
+    if (usage.completion === 0)
+        return {
+            state: "unmeasured",
+            completionTokens: 0,
+            reasoningTokens: usage.reasoning,
+            reason: "empty reply",
+        };
+    if (usage.completion < floor)
+        return {
+            state: "no-thinking",
+            completionTokens: usage.completion,
+            reasoningTokens: usage.reasoning,
+            reason: `completion_tokens ${usage.completion} below floor ${floor} with no reasoning tokens`,
+        };
+    return {
+        state: "thought",
+        completionTokens: usage.completion,
+        reasoningTokens: usage.reasoning,
+    };
+}
+const SKIPPED = {
+    state: "skipped",
+    completionTokens: null,
+    reasoningTokens: null,
+};
+/** Judge one reply to the short probe. Pure: pass any response body already collected. */
+export function judgeThinkingFloor(model, data, options) {
+    if (!mustAlwaysThink(model, options))
+        return SKIPPED;
+    return judgeReasoningUsage(readThinkingUsage(data), options?.minCompletionTokens ?? DEFAULT_MIN_COMPLETION_TOKENS);
+}
+async function floorResult(ctx) {
+    const res = await ctx.evidence("floorReply");
+    if (!res)
+        return SKIPPED;
+    if (res.status !== 200 || !res.data)
+        return {
+            state: "unmeasured",
+            completionTokens: null,
+            reasoningTokens: null,
+            reason: res.error ?? `http ${res.status ?? "none"}`,
+        };
+    return judgeReasoningUsage(ctx.wire.read.reasoningUsage?.(res.data) ?? null, ctx.checks.thinkingFloor?.minCompletionTokens ??
+        DEFAULT_MIN_COMPLETION_TOKENS);
+}
+export const thinkingFloorRule = defineRule({
+    id: "thinking-floor",
+    layer: "evidence",
+    needs: ["floorReply"],
+    check: "thinkingFloor",
+    applies: (ctx) => ctx.wire.read.reasoningUsage !== undefined,
+    judge: async (ctx) => {
+        const r = await floorResult(ctx);
+        return r.state === "no-thinking"
+            ? { severity: "fail", reason: `no-thinking: ${r.reason}`, data: r }
+            : null;
+    },
+    report: async (ctx) => ({ thinkingFloor: await floorResult(ctx) }),
+});
+//# sourceMappingURL=thinking-floor.js.map

package/dist/rules/thinking-floor.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"thinking-floor.js","sourceRoot":"","sources":["../../src/rules/thinking-floor.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,OAAO,EAAE,KAAK,EAAE,GAAG,EAAE,MAAM,mBAAmB,CAAC;AAC/C,OAAO,EAAE,gBAAgB,EAAE,iBAAiB,EAAE,MAAM,iBAAiB,CAAC;AAEtE,OAAO,EAAE,UAAU,EAAgB,MAAM,SAAS,CAAC;AAEnD,MAAM,CAAC,MAAM,6BAA6B,GAAG,EAAE,CAAC;AA0BhD,MAAM,CAAC,MAAM,mBAAmB,GAAG,CAAC,MAAM,GAAsB,EAAE,EAAE,EAAE,CACpE,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,KAAK,EAAE,GAAG,gBAAgB,CAAC,CAAC,CAAC,GAAG,EAAE,YAAY,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC;AAEhF,MAAM,UAAU,eAAe,CAC7B,KAAa,EACb,OAA8B;IAE9B,OAAO,iBAAiB,CAAC,KAAK,EAAE,mBAAmB,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC;SAClE,YAAY,CAAC;AAClB,CAAC;AAED,0FAA0F;AAC1F,MAAM,UAAU,iBAAiB,CAAC,IAAa;IAC7C,MAAM,CAAC,GAAG,GAAG,CAAC,IAAI,CAAC,CAAC;IACpB,MAAM,CAAC,GAAG,GAAG,CAAC,CAAC,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC,eAAe,CAAC,CAAC,CAAC;IACpD,IAAI,CAAC,CAAC;QAAE,OAAO,IAAI,CAAC;IACpB,MAAM,OAAO,GAAG,GAAG,CAAC,CAAC,CAAC,2BAA2B,CAAC,CAAC,CAAC;IACpD,OAAO;QACL,SAAS,EACP,KAAK,CAAC,OAAO,EAAE,CAAC,kBAAkB,CAAC,CAAC,IAAI,KAAK,CAAC,CAAC,CAAC,oBAAoB,CAAC,CAAC;QACxE,UAAU,EACR,KAAK,CAAC,CAAC,CAAC,mBAAmB,CAAC,CAAC,IAAI,KAAK,CAAC,CAAC,CAAC,sBAAsB,CAAC,CAAC;KACpE,CAAC;AACJ,CAAC;AAED,MAAM,UAAU,mBAAmB,CACjC,KAA4B,EAC5B,KAAa;IAEb,IAAI,CAAC,KAAK;QACR,OAAO;YACL,KAAK,EAAE,YAAY;YACnB,gBAAgB,EAAE,IAAI;YACtB,eAAe,EAAE,IAAI;YACrB,MAAM,EAAE,mBAAmB;SAC5B,CAAC;IACJ,IAAI,KAAK,CAAC,SAAS,KAAK,IAAI,IAAI,KAAK,CAAC,SAAS,GAAG,CAAC;QACjD,OAAO;YACL,KAAK,EAAE,SAAS;YAChB,gBAAgB,EAAE,KAAK,CAAC,UAAU;YAClC,eAAe,EAAE,KAAK,CAAC,SAAS;SACjC,CAAC;IACJ,IAAI,KAAK,CAAC,UAAU,KAAK,IAAI;QAC3B,OAAO;YACL,KAAK,EAAE,YAAY;YACnB,gBAAgB,EAAE,IAAI;YACtB,eAAe,EAAE,KAAK,CAAC,SAAS;YAChC,MAAM,EAAE,8BAA8B;SACvC,CAAC;IACJ,4EAA4E;IAC5E,oEAAoE;IACpE,IAAI,KAAK,CAAC,UAAU,KAAK,CAAC;QACxB,OAAO;YACL,KAAK,EAAE,YAAY;YACnB,gBAAgB,EAAE,CAAC;YACnB,eAAe,EAAE,KAAK,CAAC,SAAS;YAChC,MAAM,EAAE,aAAa;SACtB,CAAC;IACJ,IAAI,KAAK,CAAC,UAAU,GAAG,KAAK;QAC1B,OAAO;YACL,KAAK,EAAE,aAAa;YACpB,gBAAgB,EAAE,KAAK,CAAC,UAAU;YAClC,eAAe,EAAE,KAAK,CAAC,SAAS;YAChC,MAAM,EAAE,qBAAqB,KAAK,CAAC,UAAU,gBAAgB,KAAK,2BAA2B;SAC9F,CAAC;IACJ,OAAO;QACL,KAAK,EAAE,SAAS;QAChB,gBAAgB,EAAE,KAAK,CAAC,UAAU;QAClC,eAAe,EAAE,KAAK,CAAC,SAAS;KACjC,CAAC;AACJ,CAAC;AAED,MAAM,OAAO,GAAwB;IACnC,KAAK,EAAE,SAAS;IAChB,gBAAgB,EAAE,IAAI;IACtB,eAAe,EAAE,IAAI;CACtB,CAAC;AAEF,0FAA0F;AAC1F,MAAM,UAAU,kBAAkB,CAChC,KAAa,EACb,IAAa,EACb,OAA8B;IAE9B,IAAI,CAAC,eAAe,CAAC,KAAK,EAAE,OAAO,CAAC;QAAE,OAAO,OAAO,CAAC;IACrD,OAAO,mBAAmB,CACxB,iBAAiB,CAAC,IAAI,CAAC,EACvB,OAAO,EAAE,mBAAmB,IAAI,6BAA6B,CAC9D,CAAC;AACJ,CAAC;AAED,KAAK,UAAU,WAAW,CACxB,GAA0B;IAE1B,MAAM,GAAG,GAAG,MAAM,GAAG,CAAC,QAAQ,CAAC,YAAY,CAAC,CAAC;IAC7C,IAAI,CAAC,GAAG;QAAE,OAAO,OAAO,CAAC;IACzB,IAAI,GAAG,CAAC,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI;QACjC,OAAO;YACL,KAAK,EAAE,YAAY;YACnB,gBAAgB,EAAE,IAAI;YACtB,eAAe,EAAE,IAAI;YACrB,MAAM,EAAE,GAAG,CAAC,KAAK,IAAI,QAAQ,GAAG,CAAC,MAAM,IAAI,MAAM,EAAE;SACpD,CAAC;IACJ,OAAO,mBAAmB,CACxB,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,cAAc,EAAE,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,IAAI,EAChD,GAAG,CAAC,MAAM,CAAC,aAAa,EAAE,mBAAmB;QAC3C,6BAA6B,CAChC,CAAC;AACJ,CAAC;AAED,MAAM,CAAC,MAAM,iBAAiB,GAAG,UAAU,CAAC;IAC1C,EAAE,EAAE,gBAAgB;IACpB,KAAK,EAAE,UAAU;IACjB,KAAK,EAAE,CAAC,YAAY,CAAC;IACrB,KAAK,EAAE,eAAe;IACtB,OAAO,EAAE,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,cAAc,KAAK,SAAS;IAC5D,KAAK,EAAE,KAAK,EAAE,GAAG,EAAE,EAAE;QACnB,MAAM,CAAC,GAAG,MAAM,WAAW,CAAC,GAAG,CAAC,CAAC;QACjC,OAAO,CAAC,CAAC,KAAK,KAAK,aAAa;YAC9B,CAAC,CAAC,EAAE,QAAQ,EAAE,MAAM,EAAE,MAAM,EAAE,gBAAgB,CAAC,CAAC,MAAM,EAAE,EAAE,IAAI,EAAE,CAAC,EAAE;YACnE,CAAC,CAAC,IAAI,CAAC;IACX,CAAC;IACD,MAAM,EAAE,KAAK,EAAE,GAAG,EAAE,EAAE,CAAC,CAAC,EAAE,aAAa,EAAE,MAAM,WAAW,CAAC,GAAG,CAAC,EAAE,CAAC;CACnE,CAAC,CAAC"}

package/dist/rules/thinking-signature.d.ts

@@ -0,0 +1,43 @@
+/**
+ * Anthropic thinking signature.
+ *
+ * Claude attaches a server-generated `signature` to every thinking block. A
+ * relay serving some other model cannot produce one, so this is the only
+ * evidence here that does not rest on asking the model who it is: identity and
+ * behaviour probes can be coached with a system prompt, a signature cannot.
+ *
+ * What it does NOT prove: which tier answered. Sonnet returns a perfectly valid
+ * signature when sold as Opus, so a pass here still needs the tier checks.
+ *
+ * The `strict` replay (hand the block back so the vendor re-validates it) was
+ * built and measured: it does not work through a relay. A genuine block and the
+ * same block with 308 bytes of random base64 in place of the signature both
+ * came back 200 from two live upstreams, because a relay re-issues the turn to
+ * its own backend rather than passing our block through. It stays off by
+ * default and only tells you the endpoint accepts a thinking block.
+ */
+export type SignatureState = 
+/** Signed thinking block: a genuine Anthropic path answered. */
+"signed"
+/**
+ * Thought, but no signature. Expected when the lane crosses an
+ * OpenAI-shaped hop, which has no field to carry one, so this is
+ * unprovable rather than fake. Never a hard fail on its own.
+ */
+ | "unsigned"
+/** Thinking was requested and no block came back at all. */
+ | "no-thinking"
+/** Model or endpoint cannot be asked; carries no evidence either way. */
+ | "skipped";
+export type SignatureResult = {
+    state: SignatureState;
+    signatureLength: number;
+    /** First 32 chars, for the report; never the whole signature. */
+    signaturePrefix: string | null;
+    thinkingChars: number;
+    /** Only set when the replay check actually ran. */
+    replayVerified?: boolean;
+    reason?: string;
+};
+export declare const signatureRule: import("./types").Rule<"signature", "thinkingReply", "note">;
+//# sourceMappingURL=thinking-signature.d.ts.map

package/dist/rules/thinking-signature.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"thinking-signature.d.ts","sourceRoot":"","sources":["../../src/rules/thinking-signature.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAQH,MAAM,MAAM,cAAc;AACxB,gEAAgE;AAC9D,QAAQ;AACV;;;;GAIG;GACD,UAAU;AACZ,4DAA4D;GAC1D,aAAa;AACf,yEAAyE;GACvE,SAAS,CAAC;AAEd,MAAM,MAAM,eAAe,GAAG;IAC5B,KAAK,EAAE,cAAc,CAAC;IACtB,eAAe,EAAE,MAAM,CAAC;IACxB,iEAAiE;IACjE,eAAe,EAAE,MAAM,GAAG,IAAI,CAAC;IAC/B,aAAa,EAAE,MAAM,CAAC;IACtB,mDAAmD;IACnD,cAAc,CAAC,EAAE,OAAO,CAAC;IACzB,MAAM,CAAC,EAAE,MAAM,CAAC;CACjB,CAAC;AAwCF,eAAO,MAAM,aAAa,8DAaxB,CAAC"}