airmosaic-ai

Plan, review, generate and execute

English 中文

For a new study, begin with the planning method and skill record contract. Historical configurations are references, not default plans.

  1. Inspect authorized dictionary, data identity, split counts and development information; protect held-out outcomes.
  2. Save human-readable and machine plans before experiment-code generation. Explain every variable, bounded parameter choice, split and preprocessing decision.
  3. Default to waiting for human review. A real user delegation allows an agent review of the identified plan within scope; record both delegation and judgment.
  4. After approval, create a fresh experiment caller and task. Reuse skill calculations; identify template copies as copies rather than fresh LLM-generated algorithms.
  5. Check exact code, task/data bindings, syntax/scope, synthetic tests and dry-run. Save technical evidence before execution.
  6. Execute within approved scope. Verify results and interpretation, then conduct scientific review; preserve prior runs.

A study plan to inspect before generation

The AQ-Bench coordinate comparison maps literature and data-specific reasons to each variable, parameter, preprocessing step and proposed spatial protocol. It includes paired task drafts, a generation prompt and a code-review map. Its preparation and nested spatial evaluation remain unimplemented; no new models or experiment callers have been generated. Read the stated status before reusing it.

Executable handoff

The planned caller implements the ordering. Set the normal separate storage roots and explicit R library first. Replace all placeholders with personal invocation values:

python workflows/ml_regression/examples/planned_regression.py --repo "." --plan "<plan.json>" --input-csv "<prepared.csv>" --project-id study --out-dir "<fresh-project-result>"
python workflows/ml_regression/examples/planned_regression.py --repo "." --plan "<plan.json>" --review "<review.json>" --input-csv "<prepared.csv>" --project-id study --out-dir "<fresh-project-result>" --generate

The first command only checks the plan; without review it returns pending and writes no experiment code. For delegated judgment, add --delegation "<delegation.json>" to the second command. There is no self-approval flag.

Generation saves a plan-bound gen_script/run.py, task/review records, generation time and source hashes, plus runner preflight. Its code-check.pending.json intentionally leaves synthetic_tests=false. After actual tests and code review, create a separate code-check.json recording the evidence and time. Do not simply change a flag without checking.

python "<fresh-project-result>/gen_script/run.py" --repo "." --input-csv "<prepared.csv>" --project-id study --out-dir "<fresh-project-result>" --execute

Missing or failed checks prevent execution. Material or approved-plan changes require a fresh bundle/review as appropriate. Existing results are not overwritten. The registered runner keeps its normal path/dependency/receipt checks. Numerical outputs are indexed by artifacts.json; observed workflow boundaries are in events.jsonl.

This caller copies reviewed orchestration after approval; it does not provide an LLM API or arbitrary-code generation service. Other agent-authored callers must use the same skill checks. User evidence remains a trusted-host attestation, not authenticated identity proof.

The same caller supports an explicitly declared Python backend with execution.purpose=new_study. It records fixed implementation hashes and produces a pending-check preview bundle; this grants no execution. After actual checks, execution constructs review-bundle.json, reruns preflight and invokes the Python tool. R tasks without a backend keep the historical route. See the shared plan contract.

This site is open source. Improve this page.