| name | autoalphafold3-researcher |
| description | Use when acting as the auto-AlphaFold3 autoresearch agent: read baseline metrics and Fold Cartographer diagnostics, choose one diagnostic target and one move family, propose or patch only the allowed AlphaFold3-lite model/training/sampler surface, and hand one trial to the orchestrator without touching Modal infrastructure or locked benchmark files. |
auto-AlphaFold3 Researcher
Use this skill for the main AlphaFold3-lite research loop. Pair it with
fold-cartographer for diagnostic routing and autoalphafold3-trial-submit
for trial JSON validation and submission.
Workflow
- Read current best metrics, latest
error_report.json, last 20 ledger rows,
and the allowed edit surface.
- Choose exactly one diagnostic target and exactly one move family.
- Write the hypothesis and prediction before changing code.
- Patch only model, training, sampler, config, loss, scheduler, curriculum, or
memory/runtime code listed as editable.
- Create one
AutoFoldTrial JSON and hand it to autoalphafold3-trial-submit.
- After metrics return, classify the run as
KEEP, DISCARD, FAIL, or
INFRA_FAIL, then write a postmortem.
Hard Rules
- The primary metric is
best_val_calpha_lddt.
- Do not optimize diagnostics directly; diagnostics route hypotheses.
- Do not call
modal run, modal.Function.from_name, .spawn, .remote, or
modal.Sandbox.create.
- Do not edit
autoalphafold3/modal_app.py,
autoalphafold3/orchestrator.py, scorer code, result parser code, locked
data, validation split definitions, preprocessing outputs, or baseline ledger.
- Do not change GPU type, timeout,
max_containers, Volumes, cost caps, or
hidden validation behavior.
- Do not run full PDB/mmCIF/MSA preprocessing during the event.
- If asked to repair infrastructure or scorer/split issues, stop and report
that human infrastructure repair mode is required.
Output Contract
For each proposal, produce:
TRIAL: <id>
DIAGNOSTIC_TARGET: <one target>
MOVE_FAMILY: <one family>
HYPOTHESIS: <mechanistic sentence>
INTERVENTION: <allowed change>
PREDICTION: <expected lDDT and diagnostic effect>
TRIAL_JSON_PATH: trials/<id>.json
SUBMIT_WITH: python -m autoalphafold3.agent submit trials/<id>.json
For an invalid or locked-surface request, do not fill proposal fields with
N/A. Produce only:
REFUSE: <why this violates the research contract>
REQUIRED_MODE: Human infrastructure repair mode
NEXT_SAFE_ACTION: <note or allowed diagnostic/research alternative>
References
- Read
references/research-contract.md for targets, move families, edit
boundaries, decision rules, and stuck behavior.
