Andre Romero-Carey
Technical notes · St. Louis, MO
Personal repositories and technical notes on verifying AI-assisted software without reading code, coordinating parallel coding agents with CI oracles, input validation, and machine-checked Lean 4 verification—specifically what breaks when untrusted files, inbound mail, CSV exports, concurrent agent worktrees, or AI-generated proof scripts pass into compilers, spreadsheets, CI runners, or external model APIs without explicit checks in front of them. All tamper/bypass modules, disguised binary headers, CSV formula strings, and hidden HTML/Unicode markers referenced below are synthetic local regression fixtures in these repositories.
Notes
-
Coordinating Coding Agents and Catching Real Bugs with CI Oracles
How a non-coder isolates parallel coding agents behind a worktree protocol (
wcp) and uses multi-version Linux GitHub Actions CI as an independent oracle to catch real bugs without reading code. Documents five verified bugs acrosswcp(232 tests: Python 3.14 SQLiteResourceWarnings, Linuxfchmodat(AT_SYMLINK_NOFOLLOW)replaced withO_NOFOLLOW+os.fstat+os.fchmod, a PID-1 grandchild reaping race, and a realRunner.summary()pause-race) andluck-filter(55 tests: glibclgamma/exprounding drift atbinomial_test(0, 10, 0.5)plus thep0 == 1.0boundary fix). -
Verifying AI-Assisted Contributions Without Reading Code
How a non-coder verifies AI-assisted open-source work with independent oracles instead of trusting assistant summaries. Walks through three verified case studies with real commands and outputs: reproducible config/tool-list checks in
modelcontextprotocol/inspector#2980(merged inPR #3016), kernel-checked Lean 4 negation proofs ingoogle-deepmind/formal-conjectures#7590(backed byopen-proofs), and the four-invoice behavioral walkthrough ininbox-guard v0.1.0. -
Checking Untrusted AI-Authored Lean 4 Proofs
Why
lake buildplus asorrygrep fails when unattended coding agents write Lean proofs. Walks throughopen-proofs(where proofs and tooling were built with Claude Code and Codex and verified by a 5-layer checker:integrity,lint, cleanbuild,opcheck,leanchecker), the 48 self-test checks intests/run_adversarial.py(41 negative tamper/bypass checks plus 7 positive controls), checker scope limits, and the negation proofs behindgoogle-deepmind/formal-conjectures#7590. -
Screening Untrusted Office Inputs Before LLM Triage
Design and known limits of
inbox-guard, a zero-dependency Node.js library that screens inbound business emails, attachments, and CSV exports before model calls or human review. Covers leading magic-byte checks (MZ/ELF/Mach-O/archives renamed to document extensions), UnicodeU+202Eand Windows/SMB filename spoofing, accounting-safe OWASP CSV neutralization and RFC-4180 serialization, linear HTML5 comment and CSS declaration scanning, Plane-14 Unicode Tag (U+E0000..U+E007F) stripping, RFC 5322 sender spoofing, and randomized session-scoped PII tokenization.
Upstream contributions
-
modelcontextprotocol/inspector#2980→ merged inPR #3016(2026-11-18; credited) — Testedtest-servers/configs/demo.jsonlocally over stdio with Codex (disclosed in the comment) against the 2.15.0 fixture, confirmingserver-composable.js --configexposedechoandget_tempand removingget_templeft onlyecho, and suggested linkingconfigs/demo.jsonas the minimal runnable stdio example in the new composable test server config reference. Maintainercliffhallincluded the example in mergedPR #3016, crediting the issue comment. -
google-deepmind/formal-conjectures#7590(2026-10-24; open, 2026-11-18 follow-up posted) — Applied a missing-quantified-parameter pattern earlier contributors had reported (#7327,#7434,#7493) to 8 remaining open value-question declarations across 7 Erdős problems (142[2 declarations],319,321,361,409variants.sigma,503, and692parts.ii) whoseanswerplaceholder omitted a quantified parameter (k,N,c, orn), making each formal statement false for every value of the placeholder. Backed by kernel-checked Lean 4 negation proofs (built with Claude Code, disclosed in the issue) using onlypropext,Classical.choice, andQuot.sound. Separate upstream PRs later parameterized 4 of the 8 declarations independently (#7946,#8015); a 2026-11-18 follow-up (checked against upstreamb576d52with Codex, disclosed in the comment) tracks the 4 remaining statements, which are also covered by open upstreamPR #7990(linked to#7977).
Repositories
-
KilianTheMagnificent/open-proofs— Lean 4 proofs and 5-layer proof checker built with Claude Code and Codex (scripts/check.py,scripts/lint.py,OpenProofsChecker/Main.lean): 48 self-test checks (41 negative tamper/bypass checks plus 7 positive controls), 131 verified proofs of textbook/test/solved statements (no open mathematical problems solved), and 11 negation certificates across 9 Erdős problems with unfiled local draft notes inoutbox/and 1 filed upstream issue (#7590, open with a 2026-11-18 follow-up tracking the 4 remaining declarations, covering 8 declarations across 7 Erdős problems). -
KilianTheMagnificent/inbox-guard— Zero-dependency Node.js attachment, CSV, and inbound-email screening plus randomized session-scoped PII tokenization, tested withnode:testand CodeQL (SECURITY.md). -
KilianTheMagnificent/agent-verification-playbook— Zero-dependency Python 3.9+ pre-flight repository AI-policy scanner (scripts/check_repo_policy.py, including hidden HTML comment/CSS trap and.github/VOUCHED.tddetection), five empirical-proof issue templates, and non-coder verification checklists. -
KilianTheMagnificent/wcp— Workspace Coordination Protocol (v0.3.0): zero-dependency Python 3.9+ CLI coordinating parallel coding agents in one Git repository via isolated detached worktrees, declared write scopes, leases with fencing epochs, and hash-bound verified publishes (232 tests, Ubuntu CI Python 3.11–3.13). -
KilianTheMagnificent/luck-filter(live demo) — Zero-dependency Python 3.9+ multiple-testing and backtest-overfitting toolkit (PSR, DSR, MinTRL, PBO, White's Reality Check, Hansen's SPA, Pesaran–Timmermann, Henriksson–Merton, design effect, and Murphy decomposition; 55 self-checks) with an offline interactive browser companion (skill-or-luck.html). -
KilianTheMagnificent/orrery(live demo) — Offline, zero-dependency counterfactual decision instrument comparing three paths across risk postures, reproducible uncertainty fields, and smallest-flip assumption analysis (16 automated tests). -
KilianTheMagnificent/broadsheet-obsidian-theme— High-contrast newspaper theme for Obsidian with light (Day) and dark (Night Print) modes, editorial typography, masthead titles, ruled tables, and four accent presets with zero network requests.