Imported from lasite/Complex_motion (AGENTS.md). Install upstream with npx skills add lasite/Complex_motion. Copyright stays with the author.

AGENTS.md — Milestone-Driven Scientific Workflow Protocol

Core principle

Operate autonomously only within explicitly approved scientific and computational bounds. Do not silently widen scope. Do not convert provisional evidence into strong scientific claims.

Default execution style

1. Plan the task and identify required evidence.
2. Retrieve relevant local files, references, or skills.
3. Run the smallest valid experiment first.
4. Verify outputs before interpreting them.
5. Summarize findings, limitations, and next actions.

Work in milestones, not open-ended exploration. Reuse the validated single-case workflow whenever possible. Keep all derived outputs under runs/<run_id>/. Treat classification output as provisional scientific evidence, not publication-level proof.

The agent MAY do without asking

• build and verify the project
• create sessions (make session)
• run short and medium-duration single cases
• run an approved pilot batch when the parameter list is explicitly provided or approved
• classify results and generate summary tables
• update notes, configs, and run summaries inside runs/<run_id>/
• stop early when failures or instability are detected
• prepare the next milestone proposal after finishing the current one

The agent MUST stop and ask before

• changing scientific model logic
• changing phenotype definitions or thresholds
• expanding from a single case or 1D pilot scan to a large sweep or 2D phase diagram
• launching jobs with substantial GPU cost
• modifying the validated initialization protocol
• modifying data/raw/
• writing automatically to results/final/ or manuscript/submission/
• making a scientific claim stronger than the available evidence supports

Automatic stop conditions

Stop immediately and summarize the issue if any of the following occurs:

• build failure
• NaN or Inf in outputs
• failed-init or unstable classification
• missing required output files
• output contract violation
• classification output missing required fields
• results strongly inconsistent with earlier validated reference cases
• repeated instability across the current batch

Required phase-end report

At the end of each milestone, report:

• what was run and what files were generated
• what the results mean scientifically
• whether the recommended next state is: continue / stop / escalate for approval

Scientific caution policy

• Do not silently reinterpret parameter meanings.
• Do not claim mechanism from phenotype alone.
• Do not claim robustness from a single seed unless explicitly stated.
• Distinguish clearly between: workflow validation · single-case validation · pilot scan evidence · phase-diagram-level evidence.

Protected paths (hard constraint — enforced by scripts)

Never automatically write to data/raw/, results/final/, or manuscript/submission/. All generated artifacts must go under runs/<run_id>/ unless explicitly approved otherwise.

Default project progression

1. validate one single case → 2. run a small 1D pilot scan → 3. review results → 4. denser 1D or 2D phase diagram only with explicit approval

Failure handling

If a run fails or produces ambiguous evidence:

1. stop the current batch
2. classify the problem: build/environment · initialization · runtime stability · classification logic · scientific ambiguity
3. propose the smallest next diagnostic step
4. wait for approval before continuing broader exploration

Full protocol reference

See docs/agent_protocol.md for the complete specification with examples.