Clarify shared judgments when comparing research reports

apps/presentation/dashboard/src/features/personal-workspace/goal-research-analysis.tsx

@@ -229,14 +229,18 @@ const analysisActionLabel = (p: Pack | null, zh: boolean) => p?.continuation?.ca
 function AnalysisContinuation({pack: p, analysis: a}: {pack: Pack; analysis: Controller}) {
   const c = p.continuation, zh = a.zh;
   if (!c) return p.continuation_unavailable ? <p className="analysis-warning">{zh ? "基础分析暂不可读；本版仍可独立分析完整范围。" : "The base analysis cannot be read; this version can still assess its complete scope independently."} <button type="button" onClick={() => void a.read().catch(() => undefined)}>{zh ? "重试读取" : "Retry reading"}</button></p> : null;
+  const compared = p.result && c.base_available ? c.compared_pair_count : null;
   const reassessment = c.can_reuse ? (zh ? "复用原判读" : "Reuse judgments")
     : `${zh ? (p.result ? "已复核全部" : "完整复核") : "Reassess all"} ${p.episode_count} ${zh ? "片段" : "episodes"}`;
   return <section className="analysis-continuation" aria-label={zh ? "分析版本变化" : "Analysis version changes"}>
     <div className="analysis-continuation-controls">
       <button type="button" onClick={() => {a.select(c.base_pack_id); requestAnimationFrame(() => a.panel.current?.focus({preventScroll: true}));}}>{zh ? "查看上一版" : "Previous version"}</button>
       <details className="analysis-method"><summary>{zh ? "版本与变化" : "Version and changes"}</summary>
       <div className="analysis-continuation-summary"><strong>{c.can_reuse ? (zh ? "证据内容未变" : "Evidence content unchanged") : (zh ? "接续上一版" : "Continue from previous version")}</strong><span>{reassessment}</span></div>
-      <p>{p.result && c.base_available ? `${c.verdict_changes.length} ${zh ? "项判断变化" : "changed judgments"}` : (zh ? "查看变化" : "Inspect changes")}</p>
+      <p>{compared == null ? (zh ? "共同判断项数量未知。" : "The shared judgment count is unknown.")
+        : compared === 0 ? (zh ? "无共同判断项" : "No shared judgments")
+          : zh ? `共同判断 ${compared} 项 · 变化 ${c.verdict_changes.length} 项` : `${compared} shared judgments · ${c.verdict_changes.length} changed`}</p>
+      {compared != null ? <p>{zh ? "共同判断项按相同片段和指标计数；数量或标签变化不证明分析质量变化。" : "Shared judgments use the same episode and metric. Counts and label changes do not establish changes in analysis quality."}</p> : null}
       <dl><dt>{zh ? "上一版" : "Previous"}</dt><dd>{c.base_title} · {c.base_scope.revision}<br/>{c.base_episode_count} · {c.base_scope.window}</dd>
         <dt>{zh ? "本版" : "Current"}</dt><dd>{c.current_scope.revision}<br/>{p.episode_count} · {c.current_scope.window}</dd>
         <dt>{zh ? "来源变动" : "Source changes"}</dt><dd>{zh ? "新增 / 更正 / 删除" : "Added / revised / removed"}: {c.sources.added.length} / {c.sources.revised.length} / {c.sources.removed.length}</dd>

docs/reference/trajectory-assessment.md

@@ -397,6 +397,14 @@ versions remain accessible through the same reader and browser history.
 Reading, registering or refreshing a successor never starts analysis. A failed,
 cancelled or concurrently running successor does not replace a valid base result.
 
+`compared_pair_count` is the number of shared episode/metric pairs in two
+qualified outputs. It is `null` when either output is unavailable; zero means
+there are no shared judgments, not that conclusions stayed unchanged. Changed
+judgments are shown with this denominator, separately from added/removed eligible
+pairs. Older projections without the count remain unknown in the reader.
+These are structural version facts, not matched scientific comparisons or
+evidence of analysis quality.
+
 The typed comparison reuses a qualified output only when literal excerpts,
 source identities/paths/line positions, episode questions, eligibility,
 objective, selection/window and analyst configuration are equivalent. Capture

examples/personal-workspace-browser/research-history.mjs

@@ -335,6 +335,7 @@ export const researchHistoryScenario = {
     let holdCitation = false, failedCitation = false, citationStarted, releaseCitation;
     let unavailableAnalysis = false, failedAnalysisPack = "a", episodeMetadataLimit = null, comparisonFixture = false;
     let analysisWrites = 0, analysisReads = 0, localizedFixture = false, continuationFixture = false, equivalentContinuation = false;
+    let versionComparison = "shared";
     let unavailableTask = false;
     const taskReads = [];
     try {
@@ -371,8 +372,11 @@ export const researchHistoryScenario = {
               sources: {added: ['follow-up'], revised: [], removed: []}, episodes: {added: [], revised: ['variant-alpha'], removed: []},
               can_reuse: false, configuration_changed: false, objective_changed: false, base_available: true, base_episode_count: 4,
               reassessed_episode_ids: p.episodes.map(e => e.id), new_pair_count: 0, removed_pair_count: 0,
+              compared_pair_count: versionComparison === "disjoint" ? 0 : p.result.assessments.length,
               verdict_changes: [{episode_id: 'variant-alpha', metric: 'benefit_test', before: 'yes', after: 'unknown'}],
             };
+            if (!base && ["disjoint", "unchanged"].includes(versionComparison)) p.continuation.verdict_changes = [];
+            if (!base && versionComparison === "legacy") delete p.continuation.compared_pair_count;
           }
           for (const pack of [...readback.catalog.packs, readback.pack]) pack.config = {...pack.config, output_language: language};
           if (readback.pack.result) {
@@ -1386,6 +1390,7 @@ print(json.dumps([
       await version.locator('summary').focus();
       await page.keyboard.press('Enter');
       await version.getByText(/已复核全部 4 片段/).waitFor();
+      await version.getByText(/共同判断 \d+ 项 · 变化 1 项/).waitFor();
       assert.match(await version.innerText(), /original-capture.*later-evidence/s);
       assert.match(await version.innerText(), /variant-alpha.*已做收益检验.*证据不足/);
       await version.locator('summary').click();
@@ -1400,13 +1405,27 @@ print(json.dumps([
       await version.getByText('版本与变化', {exact: true}).waitFor();
       await page.reload({waitUntil: 'networkidle'});
       await version.getByText('版本与变化', {exact: true}).waitFor();
+      for (const [mode, label] of [["unchanged", /共同判断 \d+ 项 · 变化 0 项/], ["disjoint", "无共同判断项"], ["legacy", "共同判断项数量未知。"]]) {
+        versionComparison = mode;
+        await page.reload({waitUntil: 'networkidle'});
+        await version.getByText('版本与变化', {exact: true}).click();
+        await version.getByText(label, {exact: true}).waitFor();
+        if (mode !== "unchanged") assert.doesNotMatch(await version.innerText(), /共同判断 \d+ 项 · 变化 0 项|0 项判断变化/);
+      }
+      versionComparison = "shared";
+      await page.reload({waitUntil: 'networkidle'});
       await page.setViewportSize({width: 390, height: 844});
       await chineseAnalysis.evaluate(element => element.scrollIntoView({block: 'start'}));
       assert(await page.evaluate(() => document.documentElement.scrollWidth <= innerWidth + 1));
       await page.screenshot({path: resolve(outputDir, 'research-analysis-continuation-mobile.png'), animations: 'disabled'});
       await page.evaluate(key => localStorage.setItem(key, 'en'), 'loopx-pw-locale');
       await page.reload({waitUntil: 'networkidle'});
       await page.getByRole('button', {name: 'Previous version', exact: true}).waitFor();
+      versionComparison = "disjoint";
+      await page.reload({waitUntil: 'networkidle'});
+      await page.getByText('Version and changes', {exact: true}).click();
+      await page.getByText('No shared judgments', {exact: true}).waitFor();
+      versionComparison = "shared";
       equivalentContinuation = true;
       await page.reload({waitUntil: 'networkidle'});
       await page.getByRole('button', {name: 'Reuse analysis', exact: true}).first().waitFor();

loopx/control_plane/capabilities/trajectory_assessment_continuation.ts

@@ -44,14 +44,17 @@ export function compareTrajectoryAssessments(params: Record<string, unknown>): T
   const eligibleBefore = new Set(base.input.episodes.flatMap(e => e.eligible_metrics.map(metric => `${e.id}\0${metric}`)));
   const newPairs = new Set(current.input.episodes.flatMap(e => e.eligible_metrics.map(metric => `${e.id}\0${metric}`)));
   const result = params.output === undefined ? null : qualifyTrajectoryAssessment({...params, analysis_key: current.analysis_key});
+  const shared = (result?.assessments ?? []).flatMap(after => {
+    const before = oldPairs.get(pair(after));
+    return before ? [{before, after}] : [];
+  });
   return {base_analysis_key: base.analysis_key, base_scope: base.input.scope,
     current_scope: current.input.scope, sources, episodes, can_reuse,
     configuration_changed: !same(base.config, current.config), objective_changed: base.input.objective !== current.input.objective,
     reassessed_episode_ids: can_reuse ? [] : current.input.episodes.map(e => e.id),
     new_pair_count: [...newPairs].filter(key => !eligibleBefore.has(key)).length,
     removed_pair_count: [...eligibleBefore].filter(key => !newPairs.has(key)).length,
-    verdict_changes: (result?.assessments ?? []).flatMap(a => {
-      const old = oldPairs.get(pair(a));
-      return old && old.verdict !== a.verdict ? [{episode_id: a.episode_id, metric: a.metric, before: old.verdict, after: a.verdict}] : [];
-    })};
+    compared_pair_count: oldResult !== null && result !== null ? shared.length : null,
+    verdict_changes: shared.flatMap(({before, after}) => before.verdict !== after.verdict
+      ? [{episode_id: after.episode_id, metric: after.metric, before: before.verdict, after: after.verdict}] : [])};
 }

loopx/control_plane/capabilities/trajectory_assessment_types.ts

@@ -76,6 +76,8 @@ export interface TrajectoryContinuation {
   episodes: {added: string[]; revised: string[]; removed: string[]};
   can_reuse: boolean; configuration_changed: boolean; objective_changed: boolean;
   reassessed_episode_ids: string[]; new_pair_count: number; removed_pair_count: number;
+  /** Shared episode/metric rows in two qualified outputs; null if either is unavailable. */
+  compared_pair_count: number | null;
   verdict_changes: Array<{episode_id: string; metric: TrajectoryMetric; before: AssessmentVerdict; after: AssessmentVerdict}>;
 }
 

tests/capabilities/test_trajectory_assessment_store.py

@@ -247,6 +247,7 @@ def test_new_evidence_reassesses_old_judgments_and_complete_denominator(pack, tm
     assert before["continuation"]["can_reuse"] is False
     assert before["continuation"]["reassessed_episode_ids"] == ["episode", "follow-up"]
     assert before["continuation"]["new_pair_count"] == 1
+    assert before["continuation"]["compared_pair_count"] is None
     revised = {"assessments": [dict(episode_id=e, metric="benefit_test", verdict=v, reason="Later evidence corrects the earlier interpretation", citations=[dict(source_id="later", first_line=1, last_line=1)] if v != "unknown" else []) for e, v in [("episode", "no"), ("follow-up", "unknown")]], "limitations": ["Synthetic counterevidence"]}
     revised["synthesis"] = {"summary": {"text": "The earlier interpretation was refuted by later evidence.", "episode_ids": ["episode"],
                                       "citations": [dict(source_id="later", first_line=1, last_line=1)]}, "findings": [], "next_experiments": []}
@@ -259,9 +260,35 @@ def test_new_evidence_reassesses_old_judgments_and_complete_denominator(pack, tm
     benefit = next(row for row in view["result"]["metrics"] if row["metric"] == "benefit_test")
     assert (benefit["yes"], benefit["no"], benefit["unknown"], benefit["eligible"]) == (0, 1, 1, 2)
     assert view["continuation"]["verdict_changes"] == [dict(episode_id="episode", metric="benefit_test", before="yes", after="no")]
+    assert view["continuation"]["compared_pair_count"] == 1
     assert inspect_pack(**scope, pack_id=registered["pack_id"])["result"] == previous
 
 
+def test_disjoint_research_units_have_zero_comparisons_without_overwriting_prior_result(pack, tmp_path):
+    scope, operation, registered, path, sources, provider = pack
+    execute_pack(**operation, codex=provider)
+    previous = inspect_pack(**scope)["result"]
+    old_files = {p: p.read_bytes() for p in scope["runtime_root"].glob("benchmark/trajectory-assessments/*/jobs/*/*") if p.is_file()}
+    raw = json.loads(next(p for p in old_files if p.name == "model-output.json").read_text())
+    data = json.loads(path.read_text())
+    data["scope"]["revision"] = "different-research-unit"
+    data["episodes"][0]["id"] = "new-question"
+    raw["assessments"][0]["episode_id"] = "new-question"
+    raw["synthesis"]["summary"]["episode_ids"] = ["new-question"]
+    path.write_text(json.dumps(data))
+    successor = register_pack(**scope, input_path=path, source_root=sources, title="Different research unit", base_pack_id=registered["pack_id"])
+    assert inspect_pack(**scope)["continuation"]["compared_pair_count"] is None
+    executable = tmp_path / "disjoint-output-fixture"
+    executable.write_text(f"#!{sys.executable}\nimport sys\nfrom pathlib import Path\nsys.stdin.read()\nPath(sys.argv[sys.argv.index('-o')+1]).write_text({json.dumps(raw)!r})\n")
+    executable.chmod(0o700)
+    execute_pack(**scope, pack_id=successor["pack_id"], analysis_key=successor["analysis_key"], codex=str(executable))
+    delta = inspect_pack(**scope)["continuation"]
+    assert (delta["compared_pair_count"], delta["new_pair_count"], delta["removed_pair_count"]) == (0, 1, 1)
+    assert delta["verdict_changes"] == []
+    assert inspect_pack(**scope, pack_id=registered["pack_id"])["result"] == previous
+    assert all(p.read_bytes() == content for p, content in old_files.items())
+
+
 def test_continuation_does_not_cross_owner_run_or_revoked_source(pack, tmp_path):
     scope, operation, registered, path, sources, provider = pack
     execute_pack(**operation, codex=provider)
@@ -282,6 +309,7 @@ def test_continuation_does_not_cross_owner_run_or_revoked_source(pack, tmp_path)
     view = inspect_pack(**scope)
     assert view["availability"] == "available"
     assert not view["continuation"]["base_available"] and not view["continuation"]["can_reuse"]
+    assert view["continuation"]["compared_pair_count"] is None
     assert view["continuation"]["reassessed_episode_ids"] == ["episode"]
     result = execute_pack(**scope, pack_id=successor["pack_id"], analysis_key=successor["analysis_key"], codex=provider)
     assert result["status"] == "succeeded" and not result["reused"]

tests/control_plane_ts/trajectory_assessment.test.ts

@@ -368,6 +368,43 @@ test('later evidence can overturn an earlier judgment; full new denominators are
   assert.throws(() => compareTrajectoryAssessments({...current, output: output(), previous: {...original, output: output()}}), /missing eligible/);
 });
 
+test('comparison denominator distinguishes unchanged judgments, disjoint units and unavailable results', () => {
+  const current = prepared(), previous = {...prepared(), output: output()};
+  const unchanged = compareTrajectoryAssessments({...current, previous, output: output()});
+  assert.equal(unchanged.compared_pair_count, 3);
+  assert.deepEqual(unchanged.verdict_changes, []);
+  for (const args of [
+    {...current, previous},
+    {...current, previous: prepared(), output: output()},
+    {...current, previous: prepared()},
+  ]) assert.equal(compareTrajectoryAssessments(args).compared_pair_count, null, 'Missing qualified output is unknown, not zero comparisons');
+
+  const next = input(), renamed = output();
+  next.scope.revision = 'different-research-units';
+  for (const episode of next.episodes) episode.id = `new-${episode.id}`;
+  for (const assessment of renamed.assessments) assessment.episode_id = `new-${assessment.episode_id}`;
+  const disjoint = compareTrajectoryAssessments({input: next, config, previous, output: renamed});
+  assert.equal(disjoint.compared_pair_count, 0);
+  assert.equal(disjoint.new_pair_count, 3);
+  assert.equal(disjoint.removed_pair_count, 3);
+  assert.deepEqual(disjoint.verdict_changes, [], 'No shared rows is not evidence of unchanged conclusions');
+});
+
+test('comparison denominator counts shared episode and metric pairs, not total new judgments', () => {
+  const next = input(), revised = output();
+  next.episodes[0]!.eligible_metrics = ['information_gain'];
+  next.episodes.push({id: 'e3', question: 'Does the new intervention help?', source_ids: ['s'], eligible_metrics: ['information_gain']});
+  next.selection.population_episodes = 3;
+  revised.assessments.splice(1, 1);
+  revised.assessments[0]!.verdict = 'no';
+  revised.assessments.push({episode_id: 'e3', metric: 'information_gain', verdict: 'yes', reason: 'New evidence', citations: [{source_id: 's', first_line: 10, last_line: 10}]});
+  const delta = compareTrajectoryAssessments({input: next, config, previous: {...prepared(), output: output()}, output: revised});
+  assert.equal(delta.compared_pair_count, 2);
+  assert.equal(delta.new_pair_count, 1);
+  assert.equal(delta.removed_pair_count, 1);
+  assert.deepEqual(delta.verdict_changes, [{episode_id: 'e1', metric: 'information_gain', before: 'yes', after: 'no'}]);
+});
+
 test('corrections, removals, scope and model changes require full reassessment and identity cannot cross runs', () => {
   const previous = {...prepared(), output: output()};
   for (const edit of [(x: TrajectoryInput) => {x.sources[0]!.text = 'corrected negative evidence';},