v1.2.0 — the verifiability release

The theme of v1.2 is verifiability: every discipline the skills impose on a manuscript is machine-enforced on the repository itself first — and the enforcement tooling is itself tested.

Highlights

• Unit-tested quality tooling — 246 hermetic pytest tests under tests/ covering the validator, citation verifier, skill auditor, smoke runner, numeric-check protocol, scaffolder, and installer (make test, ~5s, no network).
• Quality scorecard — make scorecard regenerates docs/quality-scorecard.md from measured state; preflight fails when the committed page drifts from reality.
• Double machine learning demo — examples/dml-plr-demo/ pins the Chernozhukov et al. (2018) contrast to known targets: linear-controls OLS biased upward, the non-orthogonal ML plug-in attenuated at exactly var(v)/var(D)·θ, DML partialling-out on the truth with honest coverage. Five NUMERIC-CHECK assertions, numpy-only.
• End-to-end walkthrough — examples/end-to-end-walkthrough.md traces one fictional paper through all 12 workflow steps, naming the skill, the gate, and the artifact at each.
• Three-job CI — structural validation (+ offline citation gate + prose groundedness gate), unit tests, and every Python demo run under the numeric-check contract with --strict-deps.
• Pinned stack actually installs — the very first CI run of the new smoke gate caught phantom pins that had never been installable from PyPI (linearmodels==6.6, econtools==0.3.5, a polars conflict). Repaired and verified in a clean Python 3.12 environment: 11 demos, 47 numeric checks green.
• Release hygiene — CHANGELOG.md (back-filled to v1.0.0), CITATION.cff, docs/roadmap.md with the operating definition of the reference automated-empirics repo, bilingual README upgrade (CI badge, honest comparison table).

Full details in CHANGELOG.md.

🤖 Generated with Claude Code