Data Analysis

Data analysis

Use Python for deterministic calculations over data supplied in the request or through an authorized Attachment. Do not use host paths, open(), shell commands, or invented Attachment IDs.

Prepare the data

1. Inspect the supplied schema, column types, row count, missing values, duplicates, and parse failures before computing results.
2. For an Attachment, call read_attachment(attachment_id=...) only with an ID present in the Turn's Attachment metadata. Require ok: true; use content when encoding is utf-8, and decode content_base64 in memory when encoding is base64.
3. Keep the source and derived values in the existing Python interpreter. Do not treat a Python-local file as a durable Workspace file or Artifact.

Analyze and verify

Compute only the requested metrics. State the statistical convention for every ambiguous result, such as sample versus population dispersion, denominator, time window, rounding, or missing-value treatment.

For anomaly claims:

1. State the rule used.
2. Report the relevant comparison value.
3. Do not call a point anomalous when it does not cross the rule.
4. For very small samples, qualify the conclusion and prefer descriptive language.

Before submitting, recompute and compare all reported counts, extrema, sums, means, medians, dispersion values, and anomaly comparisons against the original data. If the data is insufficient, say what is missing instead of guessing; use empty metrics and anomalies lists when appropriate.

Submit

Use the exact fields accepted by the current SUBMIT binding and do not add unrecognized fields. When the explicitly selected data-analysis schema is active, submit exactly:

SUBMIT(
    answer=answer,
    findings=findings,
    metrics=metrics,
    anomalies=anomalies,
)

Keep answer as a non-empty string, findings and anomalies as string lists, and metrics as a list of objects with a stable name and value (plus a unit or convention when useful). When the default Fleet Signature is active, submit only its required answer field.