Paper drafting, review & replication for the Phase-178 autonomous-research stack (TODOS/phase-178 §178.7).
Components#
| Module | Summary |
|---|---|
authoring.py |
Sakana-v2 tree-search authoring — ideation → hypothesis → design → hparam-tune → ablation → writeup, with an Experiment-Progress-Manager promotion gate and a buggy-node classifier (debug rather than build on failures). (§178.7.1.1) |
figure_critic.py |
VLM figure critic — caption clarity, axis labelling, unit hygiene, misleading-scale (truncated bar axis, undisclosed log), and a real Machado-2009 colour-blind-safe palette check; failing figures trigger regeneration. (§178.7.1.2) |
build.py |
LaTeX + Typst build from one spec, deterministic bibliography (sorted + hashed), and a paper-embedding novelty gate before emission; fail-loud PDF compile. (§178.7.1.3) |
release.py |
Reproducibility-bundle release — Docker image / lock-file hash / seed / env / wall-clock / cost ledger, content-addressed into a Phase-111 store. (§178.7.1.4) |
integrity.py |
Plagiarism + LLM-regurgitation n-gram checks against the lit-graph corpus and training-data lineage. (§178.7.1.5) |
review.py |
5-reviewer peer-review ensemble on NeurIPS guidelines + area-chair accept/reject, a review gate (publish / auto-revise / human-review), and balanced-accuracy calibration. (§178.7.2.1–.2.2) |
benchmarks.py |
PaperBench (step-wise rubric replication rate), ScienceAgentBench (executable-deliverable success), DiscoveryBench (discovery rate by discipline, real vs synthetic). (§178.7.2.3–.5) |
The figure critic's colour checks and the deterministic build, novelty gate, signed-bundle store, integrity n-gram checks, ensemble aggregation and benchmark scoring are all fully local and tested. The VLM vision check, LLM reviewers, LaTeX toolchain and benchmark paper corpora are plug-in seams.
Testing#
bash
python3 -m pytest tests/ -q # 21 tests
python3 -m ruff check .