Gaia Research

Bake-off Workflow (reproducibility artifact)

The four-strategy compaction bake-off is a Claude Code dynamic workflow. The exact script that produced Context Diet — Lab 001 is preserved here so the experiment can be replayed.

What it does

1. Inventory — extract an exhaustive, atomic rule inventory from the original file (the scientific control). Each rule tagged loadBearing if CI-enforced/incident-codified.
2. Compact — run 4 candidate strategies in parallel (externalize, condense, telegraphic, hybrid), each producing a proposed rewrite + any linked files.
3. Verify — adversarially audit each candidate against the inventory: every rule present / weakened / missing (a rule moved to a linked file counts as present).
4. Score — disqualify candidates over the hard limit or that drop a load-bearing rule; winner = max faithfulness, tie-break on larger reduction.

How to replay on a different context file

1. Edit the constants at the top of bakeoff.workflow.js:
   • CLAUDE → absolute path to the target file.
   • LIMIT / TARGET → the harness char budget and your headroom target.
   • The CONSTRAINTS string’s “must stay fully inline” list → re-derive the do-not-touch set for your file (which sections are CI-enforced or incident-codified?).
2. In a Claude Code session in the target repo, invoke the workflow with that script.
3. Feed the returned winner.claudeMd + winner.newFiles to Phase C (apply), then run context_diet.py <file> --json for the after-measurement and make_charts.py for the charts.

See METHODOLOGY.md §4 for the full reproduction protocol and §5 for threats to validity. The workflow is stochastic in candidate wording but deterministic in which rules must survive — cache the winning corpus and re-score it for a strict replay.

The canonical script is committed alongside this note as bakeoff.workflow.js.