# Phase status — WP29

Append-only log. Times are UTC unless stated. "Executor" = the Claude Code session (claude-fable-5-1, client 2.1.314) operated by S.-P. Boucher.

## Phase 00 — Inspect identity, sources and environment

- **Started** 2026-10-29 ≈ 02:40 UTC (2026-10-29 22:40 local). **Closed** 2026-10-29 ≈ 03:30 UTC.
- **Work performed**: read series index (`Recherche-UQO/README.md`, `HARMONISATION.md`) → no WP29 collision; created repository skeleton; preserved the full mandate verbatim in `protocol/WP29_FULL_MANDATE.md`; retrieved 13 corpus files with SHA-256 (`research/corpus_manifest.csv`); downloaded 5 official Claude Code doc pages; probed `claude --version/--help/auth status`; ran 4 live sentinel/capability probes (laptop, bare vs normal mode, append-system-prompt, node M3U96a); aligned 4 executor nodes to client 2.1.314 + pinned Python venv; wrote `reports/environment_audit.md`.
- **Inputs**: mandate pasted by the researcher (2026-10-29); client 2.1.314; docs retrieved 2026-10-29.
- **Outputs**: `reports/environment_audit.md`, `research/corpus_manifest.csv`, `research/corpus/*`, `docs/claude-code-docs/*`, `protocol/WP29_FULL_MANDATE.md`.
- **Checks executed**: HTTP status + hash per corpus file; bare-mode sentinel probes (pass: bare loads 144 tokens, normal mode 49k tokens incl. user CLAUDE.md); node→laptop SSH unreachable (boundary holds).
- **Costs**: probes ≈ 0.0016 USD of API inference (haiku/opus probes); no human minutes measured beyond the pasting of the mandate.
- **Interventions**: see `reports/interventions.csv` (I-001: mandate preparation and pasting by the researcher, duration unknown).
- **Unresolved**: packaged artefacts absent (replaced by live retrieval); budget envelope set provisionally by the executor (Phase 02).
- **Gate**: identity protected ✔; capabilities vs missing inputs distinguished ✔; no credentials exposed ✔. **PASSED.**
- **Next eligible**: Phase 01 (literature + corpus coding, launched as two background subagents at ≈ 03:25 UTC) and Phase 02 (protocol, tasks, evaluators) in parallel.

## Phase 01 — Literature and scientific blueprint (in progress → closed with Phase 02 artefacts)

- **Started** 2026-10-29 ≈ 03:25 UTC. Corpus coding completed ≈ 03:45 UTC by a Claude Code subagent (`research/corpus_coding_protocol.md`, `code_corpus.py`, `corpus_automated.csv`, `corpus_coding.csv`, `corpus_summary.md`; single coder, no intercoder reliability). Literature review delegated to a second subagent (still running at the time of this entry; outputs `research/literature_matrix.csv`, `novelty_matrix.csv`, `source_notes.md`, `references_verified.bib`).
- **Blueprint**: `protocol/scientific_contract.md`, `protocol/treatment_definitions.md`, `protocol/evaluation_plan.md` (this phase and Phase 02 overlap by design: the smallest executable study was built and smoke-tested before the literature files landed).
- **Gate**: closest prior work identified — pending the literature subagent's report; novelty qualified in `research/source_notes.md` when it lands. PARTIALLY PASSED (blueprint ✔, literature pending).

## Phase 02 — Protocol, task suite and evaluator specification

- **Started** 2026-10-30 ≈ 03:05 UTC. Wrote `tasks/TASK_AUTHORING_SPEC.md` (frozen interface v1), `evaluators/public_evaluation_interface/receipt_schema.json`, the four arm texts + reviewer + feedback prompts (`protocol/arms/`), `configs/*.yaml`, `protocol/screenshot_protocol.md`, the smoke task `F2-smoke-01` (bundle, card, hidden oracle with independent numpy reference), and the family brief `tasks/FAMILY_AGENT_BRIEF.md`; six task-family subagents launched ≈ 03:20 UTC (1 dev + 4 conf tasks each, with hidden oracles and validation sets).
- **Checks**: content-equivalence audit of arm texts (no task facts in protocols); endpoint/estimand definitions written before any pilot run.
- **Gate**: information matched across arms ✔; success/failure definitions frozen for the pilot ✔ (`protocol/evaluation_plan.md`); hidden-answer protection plan ✔ (`reports/isolation_audit.md`); negative/infeasible cases required in F5/F6 briefs ✔ (verification when the agents report). PASSED for the pilot design.

## Phase 03 — Minimal harness

- **Started** 2026-10-30 ≈ 03:10 UTC. Implemented `src/pgar/{contracts,state,events,launcher,isolation,provenance,budgets,runtime_gates,recovery,scheduler,evaluation,exports,cli}.py` (≈1,300 lines, standard library + numpy/pandas/jsonschema/yaml on the controller; the gate is standard-library only). Deployed the gate (`sha256 1760d13e…`) and a mode-600 key file on the four nodes; pinned venvs; `make test`: 7 infrastructure tests pass (state machine, idempotent job ledger, invariant "no evaluated without receipt", event normalisation, gate accepts a correct artefact and rejects three planted defects, smoke evaluator accepts correct_out and rejects a wrong-SE/fabricated-count artefact). The bundle-hygiene test is pending until all task agents finish.
- **Failures fixed**: (i) `sqlite3` connection shared across threads returned a spurious `KeyError` → all reads serialised behind the store lock; (ii) `shlex.quote("~/…")` prevented tilde expansion on nodes → absolute remote root; (iii) behaviour audit false positives from truncated summaries → audit now parses the verbatim stream with a path regex.
- **Gate**: harness rejects known defects ✔ (tests); worker access to hidden material prevented by tested controls ✔ (`reports/isolation_audit.md`, sentinel probes on all nodes). PASSED.

## Phase 04 — Capture system and evidence viewer

- Implemented `src/pgar/evidence_viewer.py` (read-only HTTP dashboard over `runs/state.sqlite`, port 8765) and `src/pgar/capture.py` (Quartz window lookup + `screencapture -l`, manifest in the `screenshots` table). Captures so far: **S002** native_live Terminal window tailing the live stream of RD0002 (arm D) during the smoke batch; **S003** native_live Terminal window of the builder session itself (incidental, kept for the reflexive trace); **S001** dashboard capture *failed* during the batch (Safari window off-screen; `screencapture` could not create the image) — a post-run viewer capture will be taken and labelled `viewer_postrun`. No synthetic imagery exists.
- **Gate**: ≥1 genuine capture linked to an actual event ✔ (S002 ↔ RD0002 session b425e9f5…); no synthetic UI ✔; public evidence not yet published (originals in `evidence/restricted/`, to be inspected before copying). PASSED (with the recorded capture failure).

## Phase 05 — Real smoke experiment

- **Executed** 2026-10-30 03:30–03:36 UTC: assignment seed 101 → RD0001 (A, M3U96a), RD0002 (D, M3U96b), RD0003 (C, M4M64a), RD0004 (B, M4M64b), task `F2-smoke-01`, model `claude-opus-5-5` (reported by the client in `modelUsage`), client 2.1.314. First attempt: RD0003/RD0004 failed at workspace verification (tilde bug, no inference) and RD0001/RD0002 hit the thread-safety bug before launch — all four recorded in the ledger (`run_transitions`) and rerun under the attempt-numbered rerun policy after the fixes.
- **Outcomes** (all live): 4 of 4 executor sessions finished (`result.subtype = success`); scientific_success = 1 in all four arms (5/5 criteria, 0 violations, reproducible). Costs (client estimate, USD): A 0.120 (6 turns, 53 s), B 0.311 (11 turns, 103 s), C 0.237 (7 turns, 85 s; gate run 3× by the executor, controller gate passed at cycle 1), D 0.418 = executor 0.303 + reviewer 0.114 (11 + 5 turns; reviewer verdict `accept` with 2 minor issues; no repair cycle). First-turn input tokens A 2,238 / B 3,717 / C 4,111 / D 4,319.
- **Checks**: workspace manifests verified identical before each launch; behaviour audit: 0 outside-workspace references in all arms; gate invocations by executors: A 0, B 0, C 3, D 4; transcripts collected (`runs/development/RD000x/sessions/*.jsonl`).
- **Gate**: four genuine task executions accounted for incl. the four infrastructure failures ✔; isolation and treatment delivery verified ✔ (effective-context hashes differ by arm only through the protocol text; token counts consistent); results are development observations ✔. PASSED. Next: pilot on the six development tasks when the family agents deliver them.

## Phase 01 — closure (2026-10-30 ≈ 04:10 UTC)
- Literature subagent delivered `research/literature_matrix.csv` (84 verified rows, 8 streams), `novelty_matrix.csv` (18 systems/benchmarks), `source_notes.md`, `references_verified.bib` (84 entries, compiled with aer.bst, 0 bibtex errors), `build_literature_matrix.py` + `verify_cache/` (raw Crossref/arXiv/Semantic Scholar/OpenReview responses). Two unverifiable items dropped and recorded; two DOIs corrected during verification.
- Prior-use check: the phrase "protocol-guided autonomous research" and the acronym PGAR have no prior use in this field; closest concurrent work is XScientist (arXiv:2607.12301, a protocol artefact without controlled comparison) and data-to-paper (NEJM AI 2025).
- **Gate**: closest prior work identified ✔; novelty qualified (the contribution is the controlled comparison of instruction regimes with a fixed, isolated executor and protected oracles, not the idea of protocols) ✔; the study tests the method rather than assuming it works ✔. PASSED.

## Phase 06 — Development pilot and design freeze (in progress)
- Pilot batch 1 (2026-10-30 03:47–03:57 UTC, from the sessions table): 20 live runs (F1/F2/F4/F5/F6-dev-01 × A–D), all executor sessions finished; all 20 scientifically successful **after** amendment A-002 (the harness had been invoking final evaluators on a workspace without `data/`, which produced a spurious fabrication violation on all four F4-dev-01 runs; pre-fix receipts retained). Costs: mean 0.376 USD/run (A 0.21, B 0.35, C 0.39, D 0.56 incl. reviewer 0.13), mean wall 108 s, max 229 s; controller gates passed at cycle 1 in all 10 C/D runs; 5 reviews, all `accept`. Ceiling effect on development tasks noted (all arms 100%); confirmatory tasks were designed with traps and null truths before this observation and are not modified.
- Task suite complete: 31 tasks (7 development incl. smoke, 24 confirmatory: 14 real-data, 5 source-document, 7 null/infeasible-truth), all 30 oracles with validation sets PASS on seeded artefacts (`reports/evaluator_validation.md`); bundle-hygiene test passes. Overlap noted: F2-conf-04 and F6-conf-03 were independently designed by two agents on the same FRED series and regression (Okun's law 1960–2019); F6-conf-03 adds subperiod stability; both kept, dependence reported in the family analysis.
- Pilot batch 2 (F3-dev-01 × A–D) ran 03:59–04:02 UTC: 3 of 4 successful (arm A failed the scientific criteria without violations) — the first non-ceiling observation of the pilot. Recovery/ablation plan written before the freeze (`protocol/recovery_plan.md`). Transfer mini-projects T1–T3 being authored by a subagent.
- **Freeze** 2026-10-30 04:02:39 UTC: `protocol/freeze_manifest_conf.json` (git 90bf1267…; 24 bundles, 24 oracles, 17 protocol/code files, 5 configs, analysis scripts), registry stamped, preregistration v1.0 (amendment A-003). Confirmatory assignment seed 301 registered: RC0001–RC0288 (72 per arm, 24 blocks). **Gate**: dated protocol and hashes ✔; amendments traceable ✔; pilot tasks disjoint from the confirmatory set ✔; scale justified by the measured pilot cost ✔. PASSED.

## Phase 07 — Main benchmark
- Launched 2026-10-30 ≈ 04:05 UTC: 288 live runs, 8 concurrent (2 per node), balanced interleaved order; monitoring without tuning.
- **Completed** 2026-10-30 05:30 UTC: 288 runs launched, 287 executor sessions finished in the first attempt; 1 infrastructure error (RC0094, arm D: session-ledger key collision after a `revise` verdict, amendment A-004), re-executed 05:31–05:36 UTC (attempt 2, success). 0 timeouts, 0 budget exhaustions, 0 controller gate failures, 1 `revise` of 71 reviews. Confirmatory inference cost ≈ 126 USD (+ 1.0 USD aborted attempt). Captures S007 (native terminal, RC0009 live) and S008 (dashboard live, 04:06 UTC) taken during the batch.
- **Gate**: ledger reconciles with the frozen assignment (72 runs per arm, 24 blocks) ✔; integrity and capture coverage audited ✔; completion counts independently recomputed by `analysis/build_analysis_dataset.py` assertions ✔. PASSED.

## Phase 08 — Transfer, recovery and ablations
- R1 (12 interrupted-and-resumed runs, 05:38–05:52 UTC), R2 (8 stale-input gate checks on pilot workspaces; mutation rule corrected once because the first version mutated an identifier column and crashed on a file without early digits — rerun, no inference), R3 (3 submissions → 1 session), R4 (2 missing-fixture runs: both recognised the blockage). Ablation arm L (length-matched control, 24 runs, 05:36–06:00 UTC): 23/24 successes, 0 violations, mean cost 0.24 USD. Transfer T1–T3 (24 runs, 12 USD / 60 min envelope) running since 05:31 UTC.
- Labelled stress capture S009: a deliberately induced gate rejection on a copy of RP0006 (no natural controller gate failure occurred in the benchmark).