Profiler

pecca profile applies deterministic rules to each call's logged outputs. No LLM is involved.

1. Compute n_rows, n_unique_outputs, unique_ratio.
2. If ≥ 95% of outputs are JSON objects with the same keys: extraction (profiled only; not replaceable in v0.1).
3. Else if ≥ 98% parse as numbers: regression.
4. Else if n_unique <= max(50, 5% of rows) and unique_ratio <= 0.05: classification.
5. Else: free_text, not replaceable ("outputs are not a closed set").
6. Input type: text, multi_text or tabular.
7. replaceable also needs at least min_rows rows (default 500).
8. Languages (≥ 5% share, on a sample of 500 inputs), min_class_count and n_classes_below_20 are recorded.

With 60 classes you need at least 1,200 rows to be a "closed set" (5% rule).