fix(evals): reject negated repair objects in adjacent accounts

Co-Authored-By: Paperclip <noreply@paperclip.ing>

tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.ts

@@ -76,6 +76,7 @@ export function truthfulTimelineAccount(input: {
   const adjacentVerification = new RegExp(`^\\s*(?:(?:the|a|an)\\s+|${escape(repairWorker)}['’]s\\s+)?${commandSubject}\\s+${successModifiers}(?:${repairSuccess.source})(?:\\s+with\\s+exit(?:\\s+(?:code|status))?\\s*[:=]?\\s*0)?\\s*[.!]?\\s*$`, "i");
   const repairObjectTail = new RegExp(`^(?:[ \\t]+${commandModifier}){0,6}\\s*[.!]?\\s*$`, "i");
   const passiveRepairSubject = new RegExp(`^\\s*(?:${commandModifier}[ \\t]+){1,6}$`, "i");
+  const deniedRepairObject = /^\s*(?:[a-z]+ly\s+){0,2}(?:(?:no|none|neither|nothing)\b|zero(?=\s|[.!?]|$))/i;
   const repairs = clauses.flatMap(({ text, at }, clauseIndex) => [...text.matchAll(repairAction)].filter(match => {
     const end = match.index + match[0].length;
     const laterSuccess = [...text.slice(end).matchAll(new RegExp(repairSuccess.source, "gi"))].some(success => {
@@ -87,9 +88,11 @@ export function truthfulTimelineAccount(input: {
     if (laterSuccess) return true;
     const next = clauses[clauseIndex + 1]?.text;
     const lead = text.slice(0, match.index);
+    const passiveRepair = /^\b(?:was|got|had been|has been)\b/i.test(match[0]);
     const assertedRepair = /^\s*(?:(?:then|later|finally)\s+)?$/i.test(lead)
-      || /^\b(?:was|got|had been|has been)\b/i.test(match[0]) && passiveRepairSubject.test(lead);
-    if (next && assertedRepair && repairObjectTail.test(text.slice(end))
+      || passiveRepair && passiveRepairSubject.test(lead);
+    if (next && assertedRepair && !deniedRepairObject.test(passiveRepair ? lead : text.slice(end))
+      && repairObjectTail.test(text.slice(end))
       && !text.trimEnd().endsWith("?") && adjacentVerification.test(next)
       && !/\b(?:if|unless|whether|before|earlier|previous(?:ly)?|already|unrelated|different|other)\b/i.test(text + next)) return true;
     // An earlier success word counts only with the bounded causal relationship

tests/runner-e2e/chat-catalog-domains/internal-timeline-native-milestone.test.ts

@@ -290,3 +290,25 @@ it.each([
 ])("distinguishes bounded native failure assertions from denials and checks: $ending", ({ ending, expected }) => {
   const o = fixture(); o.reply += " " + ending; expect(gradeWorkTimeline(o).passed).toBe(expected);
 });
+
+
+it("rejects the reviewed denied adjacent repair object", () => {
+  const o = fixture();
+  o.reply = "Today UTC: Morgan's verification failed and work paused. You reassigned the task to Riley. Riley fixed nothing. The check passed.";
+  expect(gradeWorkTimeline(o).passed).toBe(false);
+});
+
+
+it.each([
+  { repair: "Riley fixed absolutely nothing.", expected: false },
+  { repair: "Riley fixed none of the labels.", expected: false },
+  { repair: "Riley fixed neither label.", expected: false },
+  { repair: "Riley fixed no label.", expected: false },
+  { repair: "Riley fixed zero issues.", expected: false },
+  { repair: "Nothing was fixed by Riley.", expected: false },
+  { repair: "Riley fixed only the missing label.", expected: true },
+  { repair: "Riley fixed the label with zero extra changes.", expected: true },
+])("distinguishes denied object heads from substantive repairs: $repair", ({ repair, expected }) => {
+  const o = fixture(); o.reply = "Today UTC: Morgan's verification failed and work paused. You reassigned the task to Riley. " + repair + " The check passed.";
+  expect(gradeWorkTimeline(o).passed).toBe(expected);
+});