autoresearch

Research to paper

This is the main Autoresearch Paper workflow. The classic iteration loop and predicate orchestrator support bounded engineering work within a research task.

Resolve paper/ relative to the installed autoresearch skill. Run commands with Python 3.10+ through paper/scripts/research.py; the repository shortcut is scripts/research.py. Read paper/references/runtime-contract.md once. Do not launch the historical autoresearch-paper Harness or a second research scheduler.

Start and work

1. Read the current question, complete materials and resource authorization. Build the Brief JSON described by paper/references/packets.md, including every required figure, table and section in deliverable_plan. Venue choice belongs to the user's research scope; compare candidates using official instructions and representative accepted papers before choosing. A generic citation or page quota does not replace the venue calibration.
2. Choose the preparation branch before approval. A concrete research Brief with an already chosen venue and complete obligations can use init, display that exact Brief and relay its actual approval. If venue/obligations are undecided, use contract_status: draft, venue: undecided, and an empty deliverable_plan; do not ask for final approval yet.
3. Commit read-only intake and prepare sourced venue candidates. Run ask-user --kind venue --packet ..., present its question through the host's real user-input tool, and wait for the answer. Apply confirm-venue --response ..., calibrate the selected target, and commit venue. For a draft, use set-contract to produce the concrete post-calibration Brief, show it and relay its approval. No experimental dispatch can use a draft or a merely completed selection audit. A previously explicit venue answer can be bound and reused; a recommendation or broad delegation is not that answer. If the current user has already delegated execution of the research run, use the evidence-bound --authorization record described in packets.md. Record the host's delegated decision, never a fictitious exact human approval or an extra confirmation for an already authorized routine action. Venue selection, initial plan budget and budget increases still require their concrete user answers. Do not treat delegation as a reply to those questions.
4. Run context. It returns the current stage knowledge, bundled protocol, frozen input hashes, itemized obligations and typed action contracts. Actually read this returned context. Perform that stage's task with the authorized host tools and return its versioned packet and artifacts. Follow production_steps and paper/references/production.md for the mechanical work. Use the provided source acquisition, environment, reading index, recording, image, TeX and review tools. Do not recreate their algorithms in a task-local helper. Scientific methods, data adapters and interpretations still come from the current research contract.
5. Submit commit --packet /absolute/packet.json --operation-id ID. Technical validation records the stage; it does not judge research novelty. Fix a refused packet using the actual diagnostics. context then reveals only the next stage's specialized knowledge. Stages are intake, venue, plan, literature, experiment, figures, write and independent review.
6. The frozen plan covers every requirement with a WorkOrder and typed actions. execute --work-id ID --operation-id ID performs registered local rendering, compilation or validation. The host performs actual research/experiments; its raw outcomes still go through the same experiment and figure checks. No generated program or arbitrary shell action is executed by this command. Before committing plan, prepare the concrete resource proposal with ask-user --kind budget --packet .... Show the default of no implicit task cap and the optional budget fields, current usage, proposal and plan. Ask the actual user, wait, then apply confirm-budget --response .... A declined proposal retains the existing budget and requires plan adjustment. Every increase uses this same question/answer route. Read paper/references/user-input.md and paper/references/budget.md. Use export_image and prepare_tex actions for venue-selected output policies. Before freezing an empirical evaluator, probe its actual environment and use the supplied ExperimentRecorder envelope to retain failed/unstarted slots. Do not infer an experiment count or method from an example. Preflight recording/statistics support for the full actual condition matrix. The supplied numerical adapter admits named conditions and preregistered paired contrasts over independent cases, with seeds averaged within each case. Preserve every planned arm and failed/unstarted slot; do not drop or relabel arms to obtain technical acceptance. Descriptive multi-arm summaries do not establish universal statistical inference or scientific completion. A development diagnostic is not a substitute for a frozen prospective experiment. Use the reading index to expose all per-panel judgments and supplied locations, not just their totals.
7. After writing, submit the task's binding array with bind-all. Each binding names its WorkOrder, assets, evidence, manuscript and literal location. Run prepare-review --out review/request.json. The actual independent reviewer reads the request, current PDF and source evidence, then returns semantic judgments, its identity and the exact request/binding subjects. assemble-review --packet assembly.json copies authoritative hashes, derives the current eight-role source closure, computes R6 and saves the snapshot. Review paths, versions, requirement IDs and reviewer counts are task inputs. Missing items, open findings or unresolved gates block assembly and acceptance.
8. Commit review, run prepare-round, read its actual request and write the host judgment in control/round-decisions/. deliver --decision PATH --operation-id ID closes that issued checkpoint, revalidates evidence, rebuilds the exported source and verifies the folder and ZIP. Raw bind, apply-round, finalize and verify-export remain available for recovery. Internal delivery does not authorize submission or claim external acceptance.