I build checks that can fail, and I try to keep what I claim under what I can show.

Receipts first — the externally-checkable parts, and their status:

• mutmut#546 · merged — a branch nothing in that mutation tester mutated; the fix was accepted upstream.
• nox#1153 · merged — an environment-dependent test failure; the fix was accepted upstream.
• tree#143 · confluent-kafka-python#2319 · open, awaiting maintainer — reproduced free-threading use-after-frees.
• private work · not publicly claimable — judge me on the four lines above, not on what I say is behind the wall.

The rest of this page is the longer version of those lines.

That sounds like a slogan, so here is what it costs. The robot below won its contest and I published the source unmodified, with the bug still in it and a note on why it stays. The audit pipeline in my own public repo turned out to emit hardcoded scores — composite: 8.2, recommendation: ACCEPT, computed from nothing — so I deleted it instead of repairing it, because repairing it would have turned a red X into a green badge certifying invented numbers.

firing-checks · Small checks that prove they can fail.

A guardrail nobody has seen fire is theater. The usual failure isn't a missing test — it's a check that reports clean because it silently couldn't run: a parser that returned an empty list, a subprocess whose exit code got dropped, a predicate never reached. Each tool ships negative controls that construct the violation and assert rejection; --selftest runs them. Three tools, and the third enforces that contract over the other two — and over itself, registered in its own registry with no exemption for the enforcer. It counts the control lines a run printed rather than trusting the summary a script writes about itself. No dependencies, Python 3.10+, each file standalone.

scanner-false-negatives · A pre-registered study that refuted its own hypothesis, twice.

I built a scanner, hypothesized where its false negatives cluster, and designed the test so it could kill the claim rather than confirm it. It did — the naive version on Bandit, and the successor on semgrep, using a discipline rating of each layer sealed before any defect was classified, so the concordance could not be circular. All 61 labels are published so a stranger can re-grade them. Exploratory: one grader, no inter-rater agreement yet, and the topic is not new — the limits and prior art are on the page.

WarmLogic-OSS · Evidence-centric governance tooling — public subset.

An evaluation harness where a run produces a manifest, a decision log, and a verify report, and verification refuses a run whose stored verdict its own event log does not support. Toy workloads; the refusal is real and one shipped run triggers it.

xinrui-robot-2023 · 1st place, 7th Xinrui Innovation Robotics Contest — HIT, May 2023.

Team 21, out of 20 teams and ~90 participants. Freshman year, four of us from four countries. Two Arduino boards split by concern, joined over serial. The source is published exactly as it ran, including a bug, with the reasoning for leaving it there. Match report (university, third-party record).

Elsewhere. Three patches to projects I do not own: two merged, one closed — and the closed one's maintainer had a better reason than my patch.

mutmut#546 — merged. A ternary is a branch nothing in that mutation tester mutated, and branch coverage cannot see an uncovered arm either, so an untested arm read as a pass twice over. Two mutants per ternary now force it down each side. The review on an earlier attempt had asked to drop the precedence wrapping; that is right for or True and wrong for and False, because and binds tighter than a top-level or — over all 16 assignments of a if b or c else d, the unwrapped form differs from the original on 2 of 16 rather than 6, and only when b is falsy, i.e. a mutant tests mostly cannot kill. I brought the table rather than the opinion. Closes their #196.

nox#1153 — merged. Their uv download tests failed on a machine without uv installed. +15/-0.

PyO3#5774 — closed, and rightly. A Vec<u8> fast path for u8-compatible buffer exporters. The maintainer's reason was better than my patch: there is no guarantee the exporting buffer is properly synchronized, so that choice belongs to user code rather than to the extraction impl. CodSpeed also measured an 11.6% regression. Worth the round trip, and it stays on this list — a record that only shows the accepted ones is not a record.

Free-threading (no-GIL) memory safety. Two use-after-free reports in C extensions, each with a minimal reproducer and three controls — no-mutator, decoy, and GIL-on — so the crash is shown to be a free-threading fault, not a generic race.

google-deepmind/tree#143 — a borrowed dict key from PyDict_Next is dereferenced across __hash__/__eq__ in assert_same_structure; 10/10 SIGSEGV, and live by default because pybind11 declares the module no-GIL-safe when it is not.

confluentinc/confluent-kafka-python#2319 — a borrowed NewTopic list item is cast to a C struct and its fields read against a stale count; 10/10 SIGSEGV with the GIL forced off, and latent until the module opts into free-threading — the GIL-on control stays clean, and the report leads with that rather than overstating. The scanner that surfaced these is a static tool I wrote; a general release is deliberately withheld.

Most of my work is in private repositories, and it stays private — so I will not dress it up as an accomplishment here. A claim you cannot check is exactly the thing this profile is about not making. Judge me on the public evidence: the reproducers, the negative controls, and the labels published for re-grading.

Reach me at espressolee1@gmail.com. It is in the sidebar too, but GitHub shows a profile email only to signed-in visitors, so it is repeated here where everyone can read it.