Skip to content

v2.9.0 — adoption release

Choose a tag to compare

@pedrohcgs pedrohcgs released this 07 Nov 03:30
· 10 commits to main since this release
4c6280e

An adoption release carrying what the 2026-11-06 audit found worth building or linking. Everything in it is additive: a new skill, new scripts and a template, guidance, and links to outside work, each listed below.

Added

  • The ledger-coverage gate understands an inline hook launcher. scripts/check-ledger-coverage.py reads sh -c '<code>' <name> <script> as running <script>, so a settings file that starts its hooks through the launcher drafted for #179 and #180 is still checked hook by hook. A typo in the launcher's script path is still caught, and a launcher with no script fails loudly. Qualified in the ledger.
  • /verify (62nd skill). Since 2.1.286, Claude Code tells Claude to run a project skill named verify right before committing; docs-only and tests-only commits are exempt. This one runs the backtest, and the quality score on staged .qmd/.tex/.R files, and stops the commit on a failure. It skips itself when the git pre-commit hook will run the same gates. It is a reminder, not enforcement: the hook and /commit remain the gates. It is bridged for Codex and Antigravity.
  • templates/triangulation-ledger.md. For each robustness check, write down and commit what result would support the claim, weaken it or leave it undecided, before the check runs; the result is filled in afterwards in a new commit. It is method-neutral and joins spec-ledger.md on date and Script:line.
  • Template, not fork, for unpublished work (README and guide, step 1, placed before the clone commands). The repository is now a GitHub template repository, so a copy can be made private; a fork of a public repository cannot.
  • scripts/data-inventory.py: a dataset manifest that never carries a value. For each file it lists the path, size and SHA-256 and, for tables, the row count, the column names, and the columns whose names suggest personal data (name, email, address, birth date, ...).
    • CSV and TSV use the standard library. A first row is listed as column names only with --header, your check that every file given has a header, because no rule can tell Alice,London from a header. Even with --header, a numeric first row is withheld.
    • .dta, .sav and .sas7bdat use pyreadstat's metadata-only read, and Parquet uses its footer, when those packages are installed.
    • --no-counts drops the row counts.
    • On restricted data, the author runs it, and the whole manifest goes through the agreement's disclosure review before it leaves the machine or reaches a model.
    • /replication-package starts its dataset manifest from it.
    • Six portability cases pin it, under cp1252 and with non-ASCII file names. Three mutants each fail their own case: one leaks a value, one drops the numeric-row check, and one lists names without --header. Codex's review on PR #184 found the all-text-first-row leak and the clearance gap.
    • The idea follows the World Bank's reproducibility skill; the code is our own.
  • scripts/beamer-qa.py: what a compiled Beamer deck looks like, page by page. It saves a screenshot of every page, prints the frame and overlay step each page belongs to (from Beamer's .nav), and lists the errors and warnings that matter from the .log beside the PDF. With --band-mm FROM TO it flags every page with ink in a strip just above the footline: a frame whose content runs into the footline often compiles without a warning. The strip depends on the theme, so it is off unless given. It needs poppler and no Python packages. /visual-audit runs it for Beamer decks. A full run deletes the page screenshots an earlier run left in its folder, and nothing else, so a deck that lost pages shows no stale ones (Codex's review on PR #185). Errors are caught in -file-line-error form too (also from that review). Six portability cases pin it, and eleven mutants each fail their own case. Adapted, with its MIT notice, from shoot.py in claesbackman/AI-research-feedback.
  • /replication-package: AI-generated analytical outputs. When a model produced data, the README records the model ID and date, the prompts, the parameters, the raw outputs, the post-processing and who checked them. AI coding help stays a disclosure matter. The idea follows the World Bank's reproducibility guidance; the wording is our own.
  • Links: Brodeur et al. 2026 (PNAS: AI-assisted teams reproduce about as often as human-only teams but catch fewer major coding errors; AI-led teams reproduce far less) in verification-ladder.md; paulgp/econ-ai-detector in writing-with-ai.md and the ecosystem section; posit-dev/skills in README and the guide.
  • TROUBLESHOOTING: "Every Bash, Edit and Write is refused, with a hook error", covering a missing hook script (recovery from a terminal) and a missing python3 on Windows.

Inventory at release: 62 skills, 18 agents, 37 rules, 11 hooks, 11 gates
(standing hook battery: 322 cases, seconds to run; portability suite: 245 cases).