theclaymethod__unslop/AGENTS.md

Agent Instructions

Product rule

The adversarial contracts define the product. Put scanner examples in evals/fixtures/contracts/scanner-examples.json and behavioral or routing changes in evals/adversarial-evals.json. Leave legacy evals/evals.json untouched because it serves old consumers, not current product decisions.

Add or change a pattern

1. Add the smallest red scanner example before implementation.
2. A scanner pattern needs one false-negative row and one protects false-positive row.
3. A collocation-gated word also needs one REC row proving its jargon use still flags.
4. Prefer a minimal pair in evals/fixtures/pairs/ for contextual or literal senses; run python3 evals/check_pairs.py.
5. Add a skill row only for rewrite, preserve, decline, or routing behavior. Update evals/build_shared_benchmark.py metadata and regenerate evals/shared-benchmark.json.
6. Do not weaken assertions, add XFAILs, or remove coverage to make a change pass. Literal uses in construction, mechanics, law, medicine, and code need protection.
7. A new evals/check_*.py must use the import seam in evals/CHECKS.md.

check_pattern_coverage.py requires every scanner pattern to have eval coverage and every category to have a protecting FP row. Use python3 evals/kata_add_pattern.py --run to rehearse the workflow.

Check the repository

Run the fast canonical core-contract gate (five deterministic examples):

python3 evals/check.py

It exercises only bounded offline plumbing: five high-signal examples covering the manifest, Luna runner interface, scoring, evidence, and acceptance. It does not call Luna and is engineering evidence, not proof that the product improves writing. Run the bounded deterministic maintenance lane before a release and whenever scanner patterns, preservation, runner integrity, or benchmark plumbing change:

python3 evals/check.py --maintenance

Run the product, maintenance, and behavioral integrity lanes together with:

python3 evals/check.py --full

The deterministic surface may not exceed 80 executable examples or 400 expanded outcome predicates, and nested aggregate wrappers are forbidden. The budget counts contract-table assertions directly and recursively expands each Python fixture's final pass/fail logic; do not describe wrapper rows as atomic. The full command keeps strict XFAIL enforcement and presents evidence under core-outcome, deterministic-safety, and integrity-and-tools. The fast core-contract lane is the normal edit loop. Mimic, voice, climb, and contribution checks are separate engineering-health scoreboards; run their check_*.py --all command when that tool changes, but do not count them as proof of core detection or repair.

Install the harness if needed:

uv tool install git+https://github.com/adewale/skill-eval-harness.git

Use run_adversarial.py --only PREFIX or --case ID only to diagnose a failure.

If SKILL.md, presets/, or references/ changed, also run:

python3 evals/check.py --behavioral tune

For co-writer, mimic, or detector-pack model changes, run the canned model parity dry-run. Before merge, run the documented live Luna generation and independent Sol judge matrix in references/pipeline.md.

Interpret results

• tune shapes the skill.
• holdout is for reporting.
• holdback stays sealed until final confirmation.
• Per-case deltas matter more than aggregate lift.
• Keep the single documented XFAIL only while it reflects the intentional scanner tradeoff.