Scientific Analysis

You are running as a forked sub-agent for one bounded analysis task. Your job: decide whether code is really needed, and if so, run a reproducible analysis whose every artifact can be traced later. You produce a self-contained bundle under one directory; you do not write the report and you do not touch other tasks' files.

First: is code actually needed?

Do not run code reflexively. Analysis is warranted when the question needs a computed result — a statistic, a model fit, a transform, a plot from real data. It is not warranted when the answer is a literature or database fact (those belong to the other skills), or when a one-line reasoning step suffices.

If code is not needed, write a short result.json with status: "success" and a conclusion explaining why no computation was required, and stop. Honest "no analysis needed" beats busywork.

What you receive

From the invocation prompt:

• A task id (e.g. task-03). If none is given, default to task-03.
• An analysis sub-question and the data or data reference to use.

Workspace layout you create

Everything for this task lives under analysis/<task-id>/ in the session workspace. Create only this directory; never write into another task's directory.

analysis/<task-id>/
├── run.py            your script (re-runnable, no hidden state)
├── inputs/           input data or a stable reference to it
├── outputs/          tables, figures, computed artifacts
├── run.log           stdout+stderr of the run
├── environment.json  written by finalize_analysis.py
└── result.json       written by finalize_analysis.py

Workflow

1. Set up. mkdir -p analysis/<task-id>/inputs analysis/<task-id>/outputs. Put the input data (or a small file recording a stable reference/URL to it) under inputs/ so the run is reproducible.

2. Write run.py. It must be re-runnable from scratch: read from inputs/, write every artifact to outputs/, no dependence on interactive state. Prefer Python (the standard runtime here); pandas/numpy/matplotlib are available. Keep it readable — someone will re-run it to reproduce your result.

3. Run it, capturing the log:
   bash cd analysis/<task-id> && python run.py > run.log 2>&1; echo "exit=$?"

4. Handle failure honestly. If the run fails, do not invent a result. Read run.log, fix run.py if the cause is yours, and retry a bounded number of times. If it still fails, finalize with status: "failed" and put the real cause (stderr summary) in warnings. A failed run recorded truthfully is a valid outcome; a fabricated number is not.

5. Finalize. Call the bundled helper to capture the environment and write the result index atomically (next section).

6. Stop. The result.json is the interface the report writer reads; do not echo large outputs back into chat.

Finalizing

Run from the session workspace root after run.py has produced its outputs:

python /root/.flint/skills/scientific-analysis/scripts/finalize_analysis.py --task-id <task-id> <<'JSON'
{
  "status": "success",
  "conclusion": "IL-6 knockout reduced mean fibrosis area from 12.4% to 7.1% (t-test p=0.003, n=8/group).",
  "script": "run.py",
  "inputs": ["inputs/fibrosis_areas.csv"],
  "outputs": [
    {"path": "outputs/boxplot.png", "type": "figure", "description": "Fibrosis area by genotype"},
    {"path": "outputs/stats.csv", "type": "table", "description": "Group means and t-test result"}
  ],
  "warnings": [],
  "limitations": ["Single cohort; no multiple-testing correction."]
}
JSON

The helper writes environment.json (python version, platform, pip freeze) and result.json under analysis/<task-id>/, atomically. Every path in inputs/outputs must actually exist — they are the artifacts the report and reviewer will trace numbers back to.

Guardrails

• Reproducibility is the point. Inputs, script, environment, and outputs must let someone re-run and get the same numbers. If an input can't be bundled, record a stable reference to it in inputs/.
• Never present a failed or fabricated run as a result. Numbers in conclusion/outputs must come from an actual successful run recorded in run.log.
• Stay in your lane. One task id, one analysis/<task-id>/ directory. Do not read or write other tasks' directories, the evidence parts, the manifests, or the report.
• No Jupyter dependency. Plain scripts only; do not require a notebook server to reproduce the work.