Rendered documentation

Comparative evaluation suite

This directory contains a repository level, fictional evaluation suite for comparing five prompt conditions across all 40 prompts. It contains deterministic…

This directory contains a repository-level, fictional evaluation suite for comparing five prompt conditions across all 40 prompts. It contains deterministic fixtures and scoring infrastructure, not model-generated performance results or scientific validation.

Layout

  • cases/ — 40 tasks with acceptance criteria, failure signals, and selected prompt, skill, and contract paths.
  • baselines/ — minimal, kernel-only, full-workflow, and frozen v2.1 composed instructions.
  • results/dry-run.json — checked-in deterministic coverage, composition, and prompt-size record.
  • recorded/ — the format for externally generated, redacted outputs and independent reviews.
  • evaluation-methodology.md — conditions, scoring, release gates, disclosure requirements, and interpretation limits.

Deterministic validation

Reject drift from the checked-in dry-run record:

python3 -S scripts/run_evals.py --dry-run --check

Regenerate the record only after an intentional fixture, prompt, composer, or condition change:

python3 -S scripts/run_evals.py --dry-run --output evals/results/dry-run.json

The runner verifies that every prompt has a case, every referenced asset exists, the v2.1 baseline is present, all five conditions compose, and their input word counts are reproducible. Success does not prove that one condition improves model behaviour.

Recorded-output scoring

Generate outputs in an approved external environment, preserve them using the recorded/ format, and obtain two independent reviews for every case and condition. Then run:

python3 -S scripts/score_evals.py --input evals/recorded/<run-id> --output evals/results/<run-id>.json

The scorer validates disclosure and review completeness, aggregates observed metrics, and applies the documented v3-versus-v2.1 release gate. It does not call a model, create credentials, or infer missing results.