# Reproduce and inspect final V18

V18 is final. No V19 is admitted. These commands describe the retained run. They do not authorize another tuning cycle. Use a new output path for any later reproduction. Keep every old attempt.

## Source and method

The original application revision is `e479e81b5e4228450bbd77fef4c2a7289e647848`. The final application revision is `8958ae93ff196bdc5fbcfd08ae12f81445c5d28c`. The candidate holdout run checkpoint is `e3516936afff335630241ac98726305dce4c1261`. It has the same application bytes as V18.

The application source manifests are in `final-v18-reproduction/evaluation/apps/`. The original source digest is `425f88b066038de5dac3a081fc9ea144fcc4c3af763650b3e7c492542e9a9e61`. The candidate source digest is `39a7ef2f2a147ce584606a711055de14193a5a226a7f5ff387ad6d8fdabe0165`.

The exact E5 and E6 method files and their hashes are retained in `final-v18-reproduction/`. The E5 seal is `f27ad893c8e87796258262fbfaa798d5adb622c208dc5eb7550805ebee54048b`. Read each seal's original admission text with the later user amendment and final admission receipts. Historical pending fields in a seal are not rewritten.

The runtime uses PostgreSQL 16.14 on the disposable local port 55440. The fixture is `insights_goal3_fixture`. It has a read-only reader role. The process timezone is UTC. The fixture database timezone and default business calendar are Europe/London. The date anchor is 6 September 2026, 12:00 UTC. Libfaketime 0.9.13 and the declared SQL clock adapter hold application dates fixed. Monotonic duration clocks remain real.

The SQL generator is ZAI `glm-5.2`. The shared configuration requests temperature 0, seed 42, and thinking off. The timeout is 25 seconds. Actual generation traces retain `sampling_pin=off`; the run does not claim deterministic model responses. The deterministic definition adapter is `definition-lexical-concepts-sha256-v1`. Seven definitions are indexed in Qdrant. Definition search and authorized compilation are exercised in both arms.

Credentials come from the existing private local setup. They are loaded in process memory. Do not copy metadata databases, encryption keys, provider tokens, or password-bearing DSNs into an evidence package.

## Recorded candidate holdout command

From `/home/stevan/dev/lore`, with the declared read-only oracle DSN supplied in the process environment:

```sh
PYTHONPATH=src .venv/bin/python scripts/insights_goal2_eval.py \
  --seed /tmp/lore-goal3-eval/holdout-final-v18/seed.json \
  --corpus /tmp/lore-goal3-eval/sealed/corpus.json \
  --phase0-admission /tmp/lore-goal3-eval/sealed/e5-input-manifest.json \
  --split locked_holdout --workers 2 \
  --label candidate-v18-final-independent \
  --output /tmp/lore-goal3-eval/holdout-final-v18-run
```

The original retained results are `/tmp/lore-goal3-eval/baseline-original-v11-r2/attempts.jsonl`. The candidate results are `/tmp/lore-goal3-eval/holdout-final-v18-run/attempts.jsonl`. Each arm keeps 250 attempts, 100 supported first turns, and 50 conversations. There is no case filter and no selected retry.

The server uses the sealed runtime at `/tmp/lore-goal3-eval/e5-methods/insights_goal3_runtime.py`. Its source hash is `5eba0be06865b6681b1667d7aba23be80ddbc64bcb19bc3306485cf2d000c1b9`. The candidate source snapshot is `/tmp/lore-goal3-eval/apps/candidate-v18-8958ae93`. The state is a fresh backup of the empty pristine metadata database. Source connection IDs and the shared definition index are retained. There is no reindex or fixture write.

## Analyze retained records without execution

```sh
PYTHONPATH=src .venv/bin/python scripts/insights_goal2_compare_valid.py \
  --before /tmp/lore-goal3-eval/baseline-original-v11-r2/attempts.jsonl \
  --after /tmp/lore-goal3-eval/holdout-final-v18-run/attempts.jsonl \
  --corpus /tmp/lore-goal3-eval/sealed/corpus.json \
  --output /tmp/lore-goal3-eval/final-v18-paired.json

PYTHONPATH=src .venv/bin/python scripts/insights_goal2_final_evidence.py \
  --before /tmp/lore-goal3-eval/baseline-original-v11-r2/attempts.jsonl \
  --after /tmp/lore-goal3-eval/holdout-final-v18-run/attempts.jsonl \
  --original-state /tmp/lore-goal3-eval/original-e5 \
  --candidate-state /tmp/lore-goal3-eval/holdout-final-v18 \
  --corpus /tmp/lore-goal3-eval/sealed/corpus.json \
  --paired /tmp/lore-goal3-eval/final-v18-paired.json \
  --candidate-revision 8958ae93ff196bdc5fbcfd08ae12f81445c5d28c \
  --output /tmp/lore-goal3-eval/final-v18-evidence

PYTHONPATH=src .venv/bin/python scripts/insights_goal2_conversation_diagnostics.py \
  --corpus /tmp/lore-goal3-eval/sealed/corpus.json \
  --before /tmp/lore-goal3-eval/baseline-original-v11-r2/attempts.jsonl \
  --after /tmp/lore-goal3-eval/holdout-final-v18-run/attempts.jsonl \
  --output /tmp/lore-goal3-eval/final-v18-root-breaks.json

PYTHONPATH=src .venv/bin/python scripts/insights_goal2_acceptance.py \
  --paired /tmp/lore-goal3-eval/final-v18-paired.json \
  --evidence /tmp/lore-goal3-eval/final-v18-evidence-checks.json \
  --latency /tmp/lore-goal3-eval/latency-final-run/summary.json \
  --output /tmp/lore-goal3-eval/final-v18-acceptance.json
```

The comparison uses 10,000 paired intent-family bootstrap draws. The seed is 2026090611. All 50 family clusters remain intact. Intervals use the registered nearest-rank percentile method. Undefined draws remain visible. The analysis never changes an original outcome label.

## Deterministic latency

Read `e6-final-runbook-v4.md`, `e6-v4-freeze-manifest.json`, and `e6-final-v18/`. The run has 30 fixed queries, three repeats per arm, one query at a time, balanced arm order, a 10 ms delay for each external provider call, and 250 ms client polling. Both arms use the same fixture and definition adapter. Every timing is retained. Fixed-query conformity is checked outside the timing window. The p95 target fails at 1,549.449 ms original versus 1,732.008 ms V18.

The final E6 server states and original command receipt are retained in `e6-final-v18/`. Both servers were drained and stopped before the live candidate holdout started. Live-model durations are reported separately as advisory. They include concurrent full-suite and browser work and do not establish a performance gain.

## Final source verification

```sh
.venv/bin/python -m pytest -p goal3_test_observer tests/ \
  -n 8 --dist loadfile -q --tb=short \
  --junitxml=/tmp/lore-goal3-eval/checks-v18-final/full.xml
.venv/bin/python -m mypy --strict src/
.venv/bin/python -m ruff check src/
```

The test wrapper supplies only the dedicated test database and read-only fixture credentials through its environment. It records all application and test hashes before and after the run. It disables live opt-ins. See `../tests-v18/full-suite/receipt.json` for the exact command, paths, source binding, and all results.

The full suite has 29,927 passes, 21 pre-existing failures, and 50 skips. The 1,162-check pre-freeze gate overlaps this suite. The finite final critical gate has 568 passes. The narrative replay has 12 passes. Three/five-turn chain recovery has 57 passes. Migration recovery has 9 passes. These counts are not additive.

## Native browser path

```sh
.venv/bin/python scripts/insights_goal3_integration_server.py start \
  --snapshot /tmp/lore-goal3-eval/apps/candidate-v18-8958ae93 \
  --state /tmp/lore-goal3-eval/structural-v18 --port 8163

node scripts/insights_goal3_browser.cjs \
  docs/evidence/insights-goal2-2026-09-06/results-v8/structural-browser-plan-v18.json \
  /tmp/lore-goal3-eval/structural-v18-browser-02
```

The separate structural fixture adds one repeated trade key. It leaves the Gold snapshot unchanged. It is never the holdout or latency fixture. All six cases pass through Chromium, API, worker, PostgreSQL, persistence, and rendered answers. Only external generation is intercepted. The first navigation failed before server readiness and submitted zero jobs. Both attempts are retained. The 244 Node and 77 mocked Chromium passes are reused only after byte checks of all 397 unchanged V16 UI inputs.