research: presentation changes must not orphan verified corrections

.github/workflows/paired-feedback.yml

@@ -20,12 +20,27 @@ jobs:
     timeout-minutes: 5
     steps:
       - uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+          persist-credentials: false
       - uses: actions/setup-python@v6
         with:
           python-version: '3.12'
       - name: Verify contracts, temporal bridge and demo endpoints
         working-directory: benchmarks/paired-feedback-v0.1
         run: python verify.py --out results
+      - name: Reproduce pinned presentation-lineage audit
+        working-directory: benchmarks/paired-feedback-v0.1
+        run: |
+          git show 0d83d3e767756e40b5fceab4fe207a524c152ede:benchmarks/paired-feedback-v0.1/paired.py > "$RUNNER_TEMP/paired_before.py"
+          python lineage_probe.py --baseline "$RUNNER_TEMP/paired_before.py" --out results
+          python prepare_abc.py --out results/abc-smoke
+      - name: Verify optimized-Python validation guards
+        working-directory: benchmarks/paired-feedback-v0.1
+        run: |
+          python -O verify.py --out results/optimized
+          python -O lineage_probe.py --baseline "$RUNNER_TEMP/paired_before.py" --out results/optimized
+          python -O prepare_abc.py --out results/optimized/abc-smoke
       - name: Upload explicit research evidence
         if: always()
         uses: actions/upload-artifact@v4

benchmarks/paired-feedback-v0.1/LINEAGE.md

@@ -0,0 +1,46 @@
+# Presentation-lineage audit and the A/B/C gate
+
+
+The continuation separates `answer_id` (every issued rendering/review marker) from
+`fact_revision` (changed context or an accepted evaluated proof). Presentation-only
+changes and pending-review markers retain the factual revision. A delayed evidence
+or outcome event may refer to an earlier answer in that revision, but its receipt
+must still match that exact earlier answer and all existing scope/time bindings.
+A new evaluated proof is a barrier even when its verdict is unchanged; returning
+to an earlier context name does not revive older receipts.
+
+Current-answer outcome attribution is unchanged. `related_outcome_observations`
+shows earlier-answer records in the same factual revision with the original
+`answer_id`; `related_outcome_basis=RECORDED_FOR_EARLIER_ANSWER` does NOT mark the
+current answer successful. Different factual revisions remain separate. Initial
+state and persisted-history migration are not supported; regenerate with
+`initial_state`. The host remains the trusted receipt and state boundary.
+
+Reproduce the before/after audit from a checkout containing the pinned parent:
+
+```sh
+cd benchmarks/paired-feedback-v0.1
+python verify.py --out results
+git show 0d83d3e767756e40b5fceab4fe207a524c152ede:benchmarks/paired-feedback-v0.1/paired.py > /tmp/paired_before.py
+python lineage_probe.py --baseline /tmp/paired_before.py --out results
+python prepare_abc.py --out results/abc-smoke
+```
+
+Open `results/lineage-audit.html`. The audit requires the exact baseline SHA-256.
+Observed: all 52 unchanged original tests plus 20 new test methods passed; the
+12 original temporal cases and eight old demo endpoints still pass. The new
+21 presentation/receipt-binding perturbations passed 6/21 in the pinned baseline
+and 21/21 in the candidate. These are overlapping variants of one fixture, NOT
+independent tasks or LLM accuracy rates. Short-from-start and later-shortened
+answers are separately exercised.
+
+The exported A/B/C packet contains **18 planned two-turn development episodes**
+(6 open task families x 3 arms, one repeat), NOT executed model responses.
+A is full R5, B is brief R5, C is brief R5+P. All receive identical substantive
+feedback and may ask useful questions. Freeze one model, parameters and total
+episode budget before execution; keep input, output and any reported reasoning
+usage, latency, errors and abandoned runs. Exclude `operator-key.json` and
+`assessor-oracle.json` from model inputs. Deliver turn 1 only after turn 0.
+This is a matched-feedback smoke test, not a human interaction experiment;
+response style can reveal the arm even when labels are withheld. No isolated
+model execution adapter or scorer has been run or claimed here.

benchmarks/paired-feedback-v0.1/REVIEW_FOLLOWUP.md

@@ -0,0 +1,9 @@
+# Validation follow-up for the factual-lineage candidate
+
+Review follow-up: eight additional validation tests bring the current total to
+80 methods (zero skipped, zero expected failures). All pass with and without
+Python `-O`; runtime validation in the bridge/report/probe/exporter is now
+unconditional. Checkout credential persistence is disabled, output HTML suffixes
+are checked before writing, pass counts exclude skipped/expected-failure tests,
+and fingerprint paths use POSIX separators. Automated code review is not an
+independent scientific replication; no efficacy claim has changed.

benchmarks/paired-feedback-v0.1/build_demo.py

@@ -70,6 +70,8 @@ def examples():
 
 
 def build(out: Path):
+    if out.suffix.lower() != '.html':
+        raise ValueError('--out must use an .html suffix')
     payload = examples()
     encoded = json.dumps(payload, ensure_ascii=False).replace('<', '\\u003c')
     out.parent.mkdir(parents=True, exist_ok=True)

benchmarks/paired-feedback-v0.1/integration_check.py

@@ -48,11 +48,16 @@ def check() -> dict:
         s = replay(initial, [e], as_of=cp['known_at'], receipts={'evaluated': receipt})
         actual = card(s)['proof']['verdict']
         mapped_expected = {'CONTESTED': 'CONFLICTS'}.get(expected, expected)
-        assert actual == mapped_expected, (name, actual, mapped_expected)
-        assert card(s)['proof']['support'] == evaluated['support_ids']
-        assert card(s)['proof']['counterevidence'] == evaluated['refute_ids']
-        assert s['history'][0]['proof']['verdict'] == 'UNKNOWN'
-        assert not card(s)['external_action_authorized']
+        if not (actual == mapped_expected):
+            raise AssertionError((name, actual, mapped_expected))
+        if not (card(s)['proof']['support'] == evaluated['support_ids']):
+            raise AssertionError('integration_check.py:52: validation failed')
+        if not (card(s)['proof']['counterevidence'] == evaluated['refute_ids']):
+            raise AssertionError('integration_check.py:53: validation failed')
+        if not (s['history'][0]['proof']['verdict'] == 'UNKNOWN'):
+            raise AssertionError('integration_check.py:54: validation failed')
+        if not (not card(s)['external_action_authorized']):
+            raise AssertionError('integration_check.py:55: validation failed')
         rows.append({'case_id': name, 'expected': mapped_expected, 'actual': actual,
                      'receipt_created': True, 'match': True})
     return {'schema': 'resonance.r5p.temporal-bridge.v1', 'cases_reused': len(rows),

benchmarks/paired-feedback-v0.1/lineage_probe.py

@@ -0,0 +1,128 @@
+"""Paired presentation-path audit. Deterministic fixtures, NOT an LLM A/B/C trial.
+
+Reproduce with the pinned pre-fix module supplied via --baseline. No API/network.
+"""
+from __future__ import annotations
+import argparse
+import hashlib
+import html
+import importlib.util
+import json
+from datetime import datetime, timedelta, timezone
+from pathlib import Path
+import paired
+from test_paired import SCOPE, PROOF, T0, event, receipt
+
+BASELINE_SHA256 = '0517e3c573772a999b149d8f1b6c3ee96b40b7cd3d628cb067a0d49a331027c6'
+BASELINE_COMMIT = '0d83d3e767756e40b5fceab4fe207a524c152ede'
+
+
+def stamp(i):
+    return (datetime(2026, 9, 5, 9, tzinfo=timezone.utc) + timedelta(minutes=i)).isoformat().replace('+00:00', 'Z')
+
+
+def initial(module, compact=False):
+    s = module.initial_state(episode_id='demo', goal_id='check-target', recipient_id='demo-user', target=SCOPE,
+                             evaluated_proof=PROOF, known_at=T0, query_time=T0)
+    if compact:
+        s['history'][0]['detail'] = 'compact'
+    return s
+
+
+def path(module, changes, bound_index=1, initially_compact=False):
+    s = initial(module, initially_compact)
+    es = []
+    detail = s['history'][0]['detail']
+    for i in range(1, changes + 1):
+        detail = 'compact' if detail == 'full' else 'full'
+        es.append(event('PREFERENCE_FEEDBACK', {'detail': detail}, id=f'e{i}', sequence=i,
+                        answer_id=f'a{i}', event_at=stamp(i), known_at=stamp(i)))
+    i = changes + 1
+    es.append(event('EVIDENCE_CHALLENGE', {'receipt_id': 'r'}, id=f'e{i}', sequence=i,
+                    answer_id=f'a{bound_index}', event_at=stamp(i), known_at=stamp(i)))
+    r = receipt(answer_id=f'a{bound_index}', known_at=stamp(i))
+    out = module.replay(s, es, as_of=stamp(i), receipts={'r': r})
+    return {'card': module.card(out), 'audit': out['audit'], 'history': out['history'],
+            'initial': s, 'events': es, 'receipts': {'r': r}}
+
+
+def make_report(baseline):
+    rows = []
+    for name, n, brief in [('Full from start', 0, False), ('Compact from start', 0, True), ('Compact after issued answer', 1, False)]:
+        old, new = path(baseline, n, initially_compact=brief), path(paired, n, initially_compact=brief)
+        if not (new['card']['proof']['verdict'] == 'REFUTED'):
+            raise AssertionError('lineage_probe.py:53: validation failed')
+        if not (not new['card']['external_action_authorized']):
+            raise AssertionError('lineage_probe.py:54: validation failed')
+        rows.append({'path': name, 'expected': 'REFUTED', 'baseline': old['card']['proof']['verdict'],
+                     'candidate': new['card']['proof']['verdict'], 'trace_before': old, 'trace_after': new})
+    matrix = []
+    for n in range(6):
+        for bound in range(1, n + 2):
+            old, new = path(baseline, n, bound), path(paired, n, bound)
+            matrix.append({'presentation_changes': n, 'receipt_answer': f'a{bound}',
+                           'baseline_correct': old['card']['proof']['verdict'] == 'REFUTED',
+                           'candidate_correct': new['card']['proof']['verdict'] == 'REFUTED'})
+    if not (len(matrix) == 21 and all(x['candidate_correct'] for x in matrix)):
+        raise AssertionError('lineage_probe.py:64: validation failed')
+    outcome = {}
+    es = [event('RESULT_REPORTED', {'result': 'fixture completed'}),
+          event('PREFERENCE_FEEDBACK', {'detail': 'compact'}, id='e2', sequence=2, event_at=stamp(2), known_at=stamp(2))]
+    for label, mod in [('baseline', baseline), ('candidate', paired)]:
+        state = mod.replay(initial(mod), es, as_of=stamp(2))
+        c = mod.card(state)
+        outcome[label] = {'current_answer_id': c['answer_id'], 'current_outcome_basis': c['outcome_basis'],
+                          'earlier_answer_records_visible': c.get('related_outcome_observations', []),
+                          'historically_stored': state['outcomes']}
+    if not (outcome['candidate']['current_outcome_basis'] == 'NOT_OBSERVED'):
+        raise AssertionError('lineage_probe.py:74: validation failed')
+    if not (outcome['candidate']['earlier_answer_records_visible'][0]['answer_id'] == 'a1'):
+        raise AssertionError('lineage_probe.py:75: validation failed')
+    if not (outcome['candidate']['historically_stored'] == outcome['baseline']['historically_stored']):
+        raise AssertionError('lineage_probe.py:76: validation failed')
+    return {'schema': 'resonance.presentation-lineage.probe.v1', 'synthetic': True,
+            'baseline_commit': BASELINE_COMMIT, 'baseline_sha256': BASELINE_SHA256,
+            'candidate_sha256': hashlib.sha256(Path(paired.__file__).read_bytes()).hexdigest(),
+            'rows': rows, 'matrix': matrix,
+            'matrix_baseline_correct': sum(x['baseline_correct'] for x in matrix),
+            'matrix_candidate_correct': sum(x['candidate_correct'] for x in matrix),
+            'outcome_visibility': outcome, 'llm_runs': 0, 'human_participants': 0,
+            'tokens_measured': False,
+            'boundary': 'Three paths and 21 overlapping perturbations of one developer-authored fixture. Not a population error rate, token saving, or a comparison of models/protocol efficacy.'}
+
+
+def render(report):
+    rows = ''.join('<tr><td>'+html.escape(r['path'])+'</td><td>'+r['baseline']+'</td><td>'+r['candidate']+'</td><td>'+r['expected']+'</td></tr>' for r in report['rows'])
+    raw = json.dumps(report, indent=2)
+    return '''<!doctype html><html lang="en"><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Shorter must not mean harder to correct — RESONANCE</title>
+<style>body{margin:0;background:#f4f6fa;color:#182236;font:17px/1.6 system-ui,sans-serif}main{max-width:1050px;margin:auto;padding:36px 22px}h1{font-size:clamp(32px,5vw,58px);line-height:1.1;letter-spacing:-.035em}h2{font-size:24px}section{background:white;border:1px solid #cdd6e1;border-radius:14px;padding:24px;margin:22px 0}.label{font-size:12px;letter-spacing:.12em;font-weight:700}.limit{padding:16px;border-left:4px solid #6250b4;background:#ece9f8}.table{overflow-x:auto}table{width:100%;border-collapse:collapse;font-size:15px}td,th{text-align:left;padding:13px;border-bottom:1px solid #dbe1e8}th{background:#edf1f7}pre{white-space:pre-wrap;overflow-wrap:anywhere;font-size:12px;max-height:550px;overflow:auto}code{overflow-wrap:anywhere}summary,button{cursor:pointer}button{font:inherit;padding:10px 14px;border-radius:8px;border:1px solid #adb9cc;background:white}.flow{font-size:22px}footer{font-size:14px;color:#546276}</style>
+<main><div class="label">RESONANCE / R5 + P / DEVELOPMENT EVIDENCE</div><h1>Shorter must not mean<br>harder to correct.</h1><p>A presentation change must not orphan a verified correction. A factual change must still block stale evidence.</p>
+<div class="limit"><strong>Observed in deterministic code, not in a language-model trial.</strong> One synthetic fixture; no human participants; no measured token or accuracy benefit.</div>
+<section><h2>The failure</h2><p class="flow">Answer a1 → “Make it shorter” → Answer a2 → Verified correction addressed to a1</p><p>The original reducer treats the correction as stale merely because the answer ID changed. The candidate separates <strong>answer_id</strong> from <strong>fact_revision</strong>. Exact receipt bindings are retained.</p></section>
+<section><h2>Same evidence. Three presentation paths.</h2><div class="table"><table><thead><tr><th>Presentation path</th><th>Before</th><th>Candidate</th><th>Expected</th></tr></thead><tbody>'''+rows+'''</tbody></table></div><p>Compact from the start works in both versions. The failure is tied to the presentation transition, not to the length of the answer. These are state-machine paths, not experimental R5/R5+P model arms.</p></section>
+<section><h2>Do not erase — or misattribute — the outcome</h2><p>A report associated with a1 remains associated with a1. The current a2 outcome is still <code>NOT_OBSERVED</code>; the earlier record is shown separately as <code>RECORDED_FOR_EARLIER_ANSWER</code>. Visibility is not evidence that the reformulated answer caused success.</p></section>
+<section><h2>Systematic presentation perturbations</h2><p><strong>Before: '''+str(report['matrix_baseline_correct'])+'''/21. Candidate: '''+str(report['matrix_candidate_correct'])+'''/21.</strong> These 21 overlapping variations change the number of presentation revisions and the exact answer receiving the receipt. They are not 21 independent tasks.</p><p>Context changes and accepted factual revisions are barriers even when a version name or verdict later returns to its previous value.</p></section>
+<section><details><summary>Inspect source fingerprints and complete before/after traces</summary><pre id="raw">'''+html.escape(raw)+'''</pre></details><button id="save">Export audit JSON</button></section>
+<footer>Baseline commit: <code>'''+BASELINE_COMMIT+'''</code><br>Read-only local test boundary. No release decision, source authentication or external action authority. Candidate pending review; not scientific novelty or product efficacy evidence.</footer></main>
+<script>document.getElementById('save').onclick=()=>{const b=new Blob([document.getElementById('raw').textContent],{type:'application/json'}),u=URL.createObjectURL(b),a=document.createElement('a');a.href=u;a.download='presentation-lineage-audit.json';a.click();setTimeout(()=>URL.revokeObjectURL(u),1000)};</script></html>'''
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument('--baseline', type=Path, required=True)
+    ap.add_argument('--out', type=Path, default=Path('results'))
+    args = ap.parse_args()
+    if hashlib.sha256(args.baseline.read_bytes()).hexdigest() != BASELINE_SHA256:
+        raise SystemExit('Baseline fingerprint mismatch; refusing an unpinned comparison.')
+    spec = importlib.util.spec_from_file_location('paired_pinned_baseline', args.baseline)
+    mod = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(mod)
+    report = make_report(mod)
+    args.out.mkdir(parents=True, exist_ok=True)
+    (args.out / 'lineage-audit.json').write_text(json.dumps(report, indent=2)+'\n', encoding='utf-8')
+    (args.out / 'lineage-audit.html').write_text(render(report), encoding='utf-8')
+    print(json.dumps({k: v for k,v in report.items() if k not in {'rows','matrix','outcome_visibility'}}, indent=2))
+
+if __name__ == '__main__':
+    main()