categorize

name	categorize
description	Categorize generated outputs for a probe session along output axes

Output Categorization

Executes Stage 3 of the pipeline (see docs/PIPELINE.md).

Step 1: Identify Session

Ask the user which experiment, or find the most recent completed session:

GET /api/probes

Step 2: Read Classification Rules

Find the probe guide from sentence_set_name:

glob data/sentence_sets/**/{sentence_set_name}.md

Read the Output Axes section — it contains classification rules for each axis and value.

Also read the sentence set JSON to get the output_axes array (axis IDs and valid values).

Step 3: Read Generated Outputs

GET /api/probes/sessions/{session_id}/generated-outputs

Returns list of {probe_id, input_text, label, generated_text, output_category}.

Skip any probes that already have output_category set (resumability).

Step 4: Classify Each Output

For each generated_text, determine:

• Per-axis classification: Read the text and assign a value for each output axis
• Primary output_category: The main classification label (may be a single axis value or a composite)

Use the probe guide's classification rules. When uncertain, use "ambiguous" or the closest match.

Step 5: POST Categories

Build the batch payload and POST:

POST /api/probes/sessions/{session_id}/output-categories
{
  "probe_id_1": {
    "output_category": "value",
    "output_category_json": "{\"axis_id\": \"value\"}"
  }
}

Important: output_category_json must be a JSON string, not a dict.

Process in batches if needed (hundreds of probes).

Step 6: Report

After posting, report:

• Total probes categorized
• Distribution per output_category value
• Any notable patterns (e.g., "78% of aquarium inputs produced aquarium continuations")