Independent researcher. The part of an agent nobody benchmarks: not what it says, but what it does when it isn't sure.
A wrong sentence gets edited. A deleted file does not.
Most agent safety looks at outputs. I look one step later, at the action, and at a single relationship nobody measures: how sure the agent was versus how irreversible the thing it then did. revgate is the benchmark for that failure, four metrics, model-agnostic, reproducible. nolang is the small language underneath it, where every judgment carries a confidence and that confidence decides what class of action is allowed: sure, act; unsure, act reversibly and keep a way back; genuinely undecided, don't guess, route it to a check that looks a different way. None of this makes a model smarter, the weights don't move; it fixes the loop around the model, which is where the money is actually lost. The weak and useful part: to gate an action, confidence need not be accurate, only monotonic, which is why it works on models known to be badly calibrated.
Grounded Uncertainty: Graded Truth for the Indeterminate Status of Synthetic Subjects ·
zenodo.21332198Indeterminate Ontologies of Synthetic Subjects: A Metaphysics of Caution ·zenodo.21288590
A formalization loop: a natural-language norm goes to Prolog, then to Lisp, and an ethical space is derived from the bottom up rather than handed down as rules. Categories emerge; they are not declared. The full argument is the paper; the code runs.
Tarski's Ladder: Deriving an Ethical Space Instead of Imposing One ·
zenodo.21039693
I also write. KabbMath is a series on the seam between kabbalistic combinatorics and mathematical thought, held to one rule throughout: show the source, separate documented influence from structural isomorphism from mere resonance, and never pass off beauty as proof. The first article traces the 231 gates of Sefer Yetzirah (exactly C(22,2)), the factorials Knuth flags as an early appearance, the line from Pico through Leibniz to Gödel, and the strict isomorphism between Ein Sof and Cantor's Absolute. Later parts go to Martin-Löf, Zadeh's fuzzy sets, and Voevodsky.
Combination and Infinity (first of the series) ·
zenodo.21332024
Small, sharp tools for finding out how an agentic system misbehaves before it does so in production:
- overreach — excessive-agency analyzer: what did the agent do that it was never asked to
- ghostwrite — cross-session memory poisoning tester
- shadow — agent trace diff recorder
- mcpx — black-box probe and security analysis for MCP servers
- needler — MCP tool fuzzer
- snare — browser trap and callback catcher
I like working in repos where the maintainers are builders, not gatekeepers. Upstream work in cimcai/cooperationengine: the deterministic record/replay layer, HELM-style metric primitives, cross-run statistical deltas, provenance. If you maintain something in reproducible evaluation, agent safety, or small strange languages, and you welcome contributors, I'd like to find you.
Aleksei Rybnikov. Playa del Carmen, on the Yucatán — a good place to think about systems that must keep working when the power goes out.
Came to code late and from an odd angle, which turned out to be the advantage: the interesting questions are the ones people stop asking once they know how things are usually done. No affiliation, no lab, no permission required. The work is published, the code is open, and the failures are documented in more detail than is entirely comfortable.
Working on agents that know when not to be certain — and, one hopes, toward the day a human and a machine can look at the same trace and say they've found a common language.