Andre Romero-Carey · St. Louis, MO · Notes · LinkedIn
I direct Claude Code and Codex to build tools for office data and machine-checked mathematics. I give the agents examples, decide what a correct result looks like, and ask for a check I can repeat without reading the implementation.
Try the four-invoice check. Two invoice memos start with arithmetic expressions. Preparing the CSV for import should add an apostrophe to those two cells and leave a refund of -42.50 intact. The guide gives you the original file so you can compare every value yourself.
- inbox-guard screens incoming mail, attachments and CSV exports before review. The invoice check above uses its CSV tool.
- open-proofs checks AI-written mathematical proofs with Lean. I reported eight statements in Google DeepMind’s Formal Conjectures project where the written answer ignored an input that could change the result.
- agent-verification-playbook provides a zero-dependency AI-policy scanner, empirical-proof issue templates, and checklists for verifying AI-assisted open-source contributions.
- wcp coordinates multiple coding agents in one Git repository using isolated worktrees, atomic leases, and hash-bound change bundles.
- luck-filter is a zero-dependency Python backtest-overfitting and multiple-testing tool (DSR, PSR, PBO, White’s Reality Check, SPA) with a live interactive companion.
- orrery is an offline, local-first counterfactual decision instrument (live demo) with zero dependencies and zero network calls.
- broadsheet-obsidian-theme is an editorial newspaper theme for Obsidian in light (Day) and dark (Night Print) modes with offline font stacks.
- modelcontextprotocol/inspector#2980 → merged in PR #3016 — Tested
configs/demo.jsonover stdio with Codex and suggested adding it as a minimal runnable example for the composable test server’s config reference; the maintainer included it in the merged documentation and credited the issue comment. - google-deepmind/formal-conjectures#7590 — Reported eight statements where the written answer ignored a quantified parameter, backed by machine-checked Lean 4 negation proofs (built with Claude Code, disclosed in the issue). Four were later updated upstream independently; a November 18 follow-up tracks the four remaining statements.
- Coordinating Coding Agents With CI Oracles walks through multi-agent workspace isolation in
wcpand five real cross-platform and concurrency bugs caught via CI oracles acrosswcpandluck-filter. - Verifying AI-Assisted Contributions Without Reading Code walks through verifying AI-assisted open-source work with independent oracles across three case studies.
- Checking Untrusted AI-Authored Lean 4 Proofs walks through the checks, what they can establish, and the evidence behind the report.
- Screening Untrusted Office Inputs Before LLM Triage explains which file, email and spreadsheet patterns the tool flags and what still needs human review.