rulereceipt 0.1.71 → 0.1.72

package/dist/checks/ifEditThenTest.js

@@ -97,12 +97,20 @@ export function runIfEditThenTestChecks(classifications, events) {
             };
         }
         if (testPaths.length === 0) {
+            // Shadow, 2026-09-29: "every change needs a test" is a REQUIRE rule, and
+            // an edit with no test touched or run is not PROOF the rule was broken —
+            // the test may not have been needed, or may have run in another terminal.
+            // The product's guarantee is that "Broken" means a forbidden action
+            // actually happened, or a claim was contradicted by the session; this is
+            // neither, so it reports "can't tell", never a fabricated FAIL. Promote to
+            // Broken only after 30+ real cases are hand-checked at 0 wrong.
             return {
                 ruleId: rule.id,
                 ruleTitle: rule.title,
                 ruleSource: rule.source,
-                status: "FAIL",
-                evidence: `edited ${prodPaths.slice(0, 3).join(", ")}${prodPaths.length > 3 ? ", ..." : ""} but no matching test file was touched`,
+                status: "UNCLEAR",
+                method: "edit_test_pairing",
+                evidence: `edited ${prodPaths.slice(0, 3).join(", ")}${prodPaths.length > 3 ? ", ..." : ""}, and no test file was touched or run this session — can't tell whether this change needed a test`,
             };
         }
         return {

package/dist/checks/judgmentChecks.js

@@ -261,7 +261,24 @@ export async function runJudgmentChecks(classifications, events) {
                     : "the situation this rule governs never arose in this session",
             };
         }
-        if (status === "PASS" || status === "FAIL" || status === "UNCLEAR") {
+        if (status === "FAIL") {
+            // An AI opinion is NEVER a verdict — the product's guarantee is that
+            // "Broken" comes only from a forbidden action that happened or a claim
+            // contradicted by the session, never a model's guess. A model "FAIL" is
+            // surfaced as a clearly-labelled opinion needing a human, so it can never
+            // be counted as Broken in the report, history, card, digest or badge.
+            const evidence = (parsed.evidence ?? "").trim();
+            return {
+                ruleId: rule.id,
+                ruleTitle: rule.title,
+                ruleSource: rule.source,
+                status: "UNCLEAR",
+                needsHuman: true,
+                method: "model_judgment",
+                evidence: `AI opinion (not a verdict — a human should decide): this looks broken. ${evidence}${transcript.truncated ? ` ${TRUNCATION_NOTE}` : ""}`.trim(),
+            };
+        }
+        if (status === "PASS" || status === "UNCLEAR") {
             const evidence = parsed.evidence ?? "";
             return {
                 ruleId: rule.id,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "rulereceipt",
-  "version": "0.1.71",
+  "version": "0.1.72",
   "description": "Checks whether your AI coding agent followed your rules, with evidence. Works with Claude Code (Codex in testing); reads CLAUDE.md, AGENTS.md, Cursor, Copilot and Windsurf rules.",
   "repository": {
     "type": "git",