.PHONY: help test gate bench verify validate retention fidelity suite holdout build pyz docker clean lint

help: ## Show this help

@grep -E '^[a-z]+:.*?## .*$$' $(MAKEFILE_LIST) | awk 'BEGIN{FS=":.*?## "}{printf " \033[36m%-10s\033[0m %s\n", $$1, $$2}'

test: ## Run the full test suite

uv run --with pytest --with pillow python -m pytest -q

gate: bench verify validate retention fidelity suite ## Run the full CI gate (non-inferiority + byte-fidelity + adversarial + recall + state probes + public benchmarks)

# Must stay identical to the gate steps in .github/workflows/ci.yml. `validate` was

# missing here while CI ran it, so a local green gate could still fail on push —

# which is the one thing this target exists to prevent.

fidelity: ## State probes: artifact state, overclaim, continuation, propagation

# Gated at the measured band, not 0: the reversible tier drops hedging on a

# small number of claims (docs/EVALUATION.md section 6.2). Gating at 0 would

# assert a property the code does not have.

uv run distil fidelity --max-silent 15

suite: ## Public benchmarks (third-party ground truth). First run fetches + caches.

# Zero API spend: grading is deterministic recall, not an LLM judge. Tier 1 is

# the payload an agent proxy actually risks breaking — tool schemas and retrieval.

# Gated at the MEASURED band on RICH rows only: bfcl, hotpotqa and squad all

# 1.000 support at n=25 (bfcl read 0.964 under the old prose matcher; scoring its

# golds as identifiers removed 11 accidental substring credits and 15 one-letter

# golds no text rule can adjudicate — see docs/RUNNING-EVALS.md). The band stays at

# 0.95 rather than 1.0: a gate pinned to a perfect score fails on sampling noise

# and teaches people to widen it, which is how a gate stops meaning anything.

# Controls are excluded by construction — gsm8k loses 17 bare-number answers here,

# which says nothing about compression.

uv run distil suite --tier 1 -n 25 --min-answer-recall 0.95 --min-support-recall 0.95 --allow-unavailable

bench: ## Corpus-wide non-inferiority gate

uv run distil bench

validate: ## Adversarial invariants on hostile input (reversibility, fail-open, recency)

uv run distil validate

verify: ## Byte-fidelity gate (reversibility + append-only)

uv run distil verify

retention: ## Fact-level recall gate (zero cost: no LLM, no API key)

uv run distil retention --max-lost 0

holdout: ## Holdout A/B savings with bootstrap CI

uv run distil holdout

build: ## Build wheel + sdist (PyPI distributables)

uv build

pyz: ## Build the single-file executable (dist/distil.pyz)

bash scripts/build_pyz.sh

docker: ## Build the container image

docker build -t distil:latest .

lint: ## Lint with ruff (pinned to the version CI gates on)

uvx ruff@0.15.10 check distil tests

uvx ruff@0.15.10 format --check .

clean: ## Remove build artifacts

rm -rf dist build *.egg-info .pytest_cache .ruff_cache