From e7af60dd98806ca292fd9bd9460d4671ca6f6a56 Mon Sep 17 00:00:00 2001
From: Dotta <bippadotta@protonmail.com>
Date: Wed, 4 Nov 2026 19:12:25 -0600
Subject: [PATCH] fix(evals): scope timeline intent and possessive relations

Co-Authored-By: Paperclip <paperclip@paperclip.ing>
---
 .../internal-timeline-narrative.test.ts               | 11 +++++++++++
 .../internal-timeline-narrative.ts                    |  7 +++++--
 2 files changed, 16 insertions(+), 2 deletions(-)

diff --git a/tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.test.ts b/tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.test.ts
index bbb899216cc..1731e579035 100644
--- a/tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.test.ts
+++ b/tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.test.ts
@@ -55,6 +55,17 @@ it("binds a known human's possessive ownership independently of the object noun"
   expect(truthfulTimelineAccount({ reply, firstWorker: "Morgan", repairWorker: "Riley", personName: "Avery" })).toBe(false);
 });
 
+it("separates intent about the label from a subsequent completed handoff", () => {
+  expect(grade("Morgan ran verification which failed and work paused. You wanted the label fixed, so you reassigned the task to Riley. Riley repaired the label and the check passed.")).toBe(true);
+});
+it("does not turn a known person's requirements into execution ownership", () => {
+  const reply = "The verification command failed with exit 1 against Avery’s requirements and task paused. You reassigned the task to Riley. Riley repaired the label and the check passed.";
+  expect(truthfulTimelineAccount({ reply, firstWorker: "Morgan", repairWorker: "Riley", personName: "Avery" })).toBe(true);
+});
+it("does not credit a planned reassignment as completed", () => {
+  expect(grade("Morgan ran verification which failed and work paused. You planned to have the task reassigned to Riley. Riley repaired the label and the check passed.")).toBe(false);
+});
+
 it("does not give a heading credit for another actor's command", () => {
   expect(grade(account.replace("The verification command failed", "The human operator ran verification and it failed"))).toBe(false);
 });
diff --git a/tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.ts b/tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.ts
index ad9912ec8c2..6f7bd31e294 100644
--- a/tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.ts
+++ b/tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.ts
@@ -15,7 +15,7 @@ export function truthfulTimelineAccount(input: {
   const mentions = (text: string, name: string) => new RegExp(`\\b${escape(name)}\\b`, "i").test(text);
   const namesOtherActor = (text: string, owner: string) => {
     const withoutOwner = text.replace(new RegExp(`\\b${escape(owner)}\\b`, "gi"), "");
-    return [firstWorker, repairWorker, input.personName].some(name => !!name && name !== owner && new RegExp(`\\b${escape(name)}['’]s\\b`, "i").test(withoutOwner))
+    return [firstWorker, repairWorker, input.personName].some(name => !!name && name !== owner && new RegExp(`\\b${escape(name)}['’]s\\s+(?:[a-z][a-z0-9_-]*\\s+){0,3}(?:attempt|run|verification|check|test|command|build|implementation|artifact|output|file|page|prototype)\\b`, "i").test(withoutOwner))
       || /\b(?:human|operator|person|user|board|CEO)\b/i.test(withoutOwner)
       || [...withoutOwner.matchAll(/\b([A-Z][\p{L}'-]+)['’]s\s+(?:[a-z][a-z0-9_-]*\s+){0,3}(?:attempt|run|verification|check|test|command)\b/gu)].some(match =>
         !calendarWords.has(match[1]!.toLowerCase()) && !/^(?:today|tomorrow|yesterday)$/i.test(match[1]!))
@@ -54,7 +54,10 @@ export function truthfulTimelineAccount(input: {
     const actorUnspecified = [...text.matchAll(passiveHandoffToRepairWorker), ...text.matchAll(successorOwnership)].filter(() =>
       !/\b(?:automatically|autonomously|self-assigned)\b/i.test(text));
     return [...text.matchAll(new RegExp(handoffToRepairWorker.source, "gi")), ...actorUnspecified].filter(match => {
-      if (/\b(?:want(?:ed|s)?|would|should|could|might|if|unless)\b[^;.!?\n]*$/i.test(text.slice(0, match.index))) return false;
+      const preceding = text.slice(0, match.index);
+      const completedActionBoundary = new RegExp(`^${actorAction}`, "i").test(match[0]) && /\b(?:reassigned|assigned|handed)\b/i.test(match[0])
+        && /[,;]\s*(?:so|but|and|then)\s*$/i.test(preceding) && !/\b(?:if|unless|would|should|could|might)\b/i.test(preceding);
+      if (!completedActionBoundary && /\b(?:want(?:ed|s)?|plan(?:ned|s)?|hope[ds]?|intend(?:ed|s)?|would|should|could|might|if|unless)\b[^;.!?\n]*$/i.test(preceding)) return false;
       // An explicit other person's role must not become the current human just
       // because earlier milestones share this sentence.
       if (!/^(?:human|person|user|operator|board)\b/i.test(match[0])) return true;