Skip to content

Generate a Deep Dive

/deep-dive produces a comprehensive technical reference document (DEEP_DIVE.md) after an investigation or experiment phase is complete.

When to use

Run /deep-dive after:

  • A full /ml-lab investigation completes
  • A major experiment phase finishes
  • You want a standalone technical reference covering implementation details

Generate the document

/deep-dive [experiment_path]

If experiment_path is omitted, /deep-dive looks for experiment artifacts in the current directory.

What it covers

The generated DEEP_DIVE.md includes:

  • Data construction — how inputs were generated or collected
  • Model architecture — design choices and parameters
  • Scoring mechanics — rubrics, judge configurations, aggregation
  • Statistical methods — tests used, multiple comparison corrections
  • Per-test detail — individual test breakdowns with results
  • Quality gates — validation checks and their outcomes
  • Aggregation — how individual results compose into conclusions
  • Key design decisions — sourced from journal entries and plan documents

How it finds information

/deep-dive surveys:

  1. Experiment scripts (.py files)
  2. Results schemas (JSON/JSONL output files)
  3. Journal decisions (from .project-log/journal.jsonl)
  4. Existing documentation (REPORT.md, CONCLUSIONS.md, plan files)

It works with or without a project journal and degrades gracefully when artifacts are missing.

You have now...

Generated a comprehensive technical reference document that covers implementation details end-to-end, with sources cited from scripts, results, and journal entries.