jev-experiments: A Well-Known Developer's Public Jev Notebook

"jev-experiments" is not a unique name. Two unrelated repositories use it, published within days of each other, and both are exactly what the name suggests: a notebook of small, exploratory projects rather than a single polished tool. The difference between them is instructive. One, from a first-time contributor, uses Jev to play two simple games and is upfront that nothing in it has been independently verified. The other, published under developer-educator Nader Dabit's GitHub account (dabit3), has already collected 371 stars and includes a pre-commit security scanner, Commit Sentry, with a documented benchmark run behind its claims.

This article covers both, because the contrast says something about where in the maturity curve most Jev projects currently sit — and how much a known contributor's name can pull attention toward code that still needs the same scrutiny as anyone else's. For the general framing on typed decisions, see typed decisions vs free text agents; for the wider evaluation pattern, see real-time Jev: loops, games and trading.

Key Takeaways

• mittal-parth/jev-experiments (0 stars) uses Jev to play Chrome's offline dinosaur game and a local shooter arena, driven entirely by structured state — no screenshots ever reach the model.
• It ships a labeled heuristic fallback, so you can compare Jev's decisions against simple rule-based play, which is a genuinely useful design choice for a project this early.
• dabit3/jev-experiments (371 stars) is a different animal — developer-focused tooling, including Commit Sentry, a semantic pre-commit hook.
• Commit Sentry's own benchmark reports 338ms total hook execution and 58 hunks/second, against a regex baseline that caught only 6 of 17 findings the same run flagged.
• Star count tracks visibility, not verification. Neither repo's numbers have been independently retested here; the gap between 0 stars and 371 is attention, not proof.

mittal-parth/jev-experiments: games as a testbed for typed decisions

mittal-parth/jev-experiments (no license listed, 0 stars as of September 2026, /skills/jev-experiments-mittal-parth) contains two small demos. Dino-Jev drives Chrome's chrome://dino/ offline game: it snapshots pose, speed, and obstacle state without ever sending a screenshot, sends one request per tick, and executes the returned jump/duck/run decision through local Python. A web inspector at http://127.0.0.1:8766 shows the decisions live. Configuration knobs (--speed-cap, --lead-frames) tune jump timing, with defaults aimed at giving roughly 80ms of decision latency before a pterodactyl spawns.

Krunker-Jev is a local shooter arena, explicitly not live krunker.io — the author frames it as "the canvas is a view, it never goes to the model," observing structured state like field of view, cover, ammo, and health, and returning typed combat actions with what the README describes as speculative fan-out for parallel decisions.

The genuinely useful part of this repo, for a 0-star project, is that both demos run a labeled heuristic fallback when no API key is configured, driving the same action space with simple rules. That gives you an honest baseline to compare Jev's decisions against, rather than a demo that only works when the paid API is live. No win rates, scores, or latency numbers beyond the tick-timing design goal are published — this is a testbed, not a benchmarked result.

dabit3/jev-experiments: a known developer's toolkit, with one documented tool

dabit3/jev-experiments (no license listed, 371 stars as of September 2026, /skills/jev-experiments-dabit3) is published under the GitHub account of Nader Dabit, a well-known developer educator, which likely explains most of the gap in star count against its 0-star namesake. The repo's landing page carries a "Built by Devin" badge — Devin is Cognition's autonomous coding agent — though that attribution applies to the top-level page, not confirmed for every subproject inside the repo, including Commit Sentry specifically.

Commit Sentry is the most developed tool in the repository: a pre-commit hook that runs staged diff hunks through Jev across ten independent checks — leaked secrets and credentials, sensitive data written to logs, destructive database operations (DROP, TRUNCATE, unfiltered DELETE/UPDATE), skipped tests, changed API signatures, debug artifacts left in code, hardcoded internal IPs, and whether the commit message actually describes the diff. It also assigns an overall risk score from cosmetic to dangerous.

The project's documented benchmark run reports 338ms total hook execution on a Linux VM, p50 124ms and p95 240ms per request, and throughput of 58 hunks per second with 16 processed concurrently. Against that, a regex-based linter baseline ran in 0.55ms but caught only 6 of the 17 findings Commit Sentry flagged in the same run — a real speed-versus-coverage tradeoff, reported by the author, not independently reproduced here. Its verdict policy is threshold-based: leaked secrets or destructive changes report at 50% confidence and block at 70%; logged sensitive data blocks at 80%; skipped tests, API changes, and hardcoded values only warn, never block, regardless of confidence. A --strict flag tightens all of these. git commit --no-verify bypasses the hook entirely, which is worth knowing before you rely on it as a hard gate.

What to do next

1. Don't let star count stand in for verification. dabit3's repo has real benchmark data behind Commit Sentry specifically; that data is still the author's own, not independently retested.
2. Try mittal-parth's heuristic fallback comparison if you want a cheap, concrete way to see what a typed decision model adds over simple rules in a bounded, low-stakes setting.
3. If you adopt Commit Sentry, start without --strict and watch which findings land at the warn-only tier before deciding whether to promote any of them to blocking.
4. Remember --no-verify exists. A pre-commit hook, however well-calibrated, is not a substitute for CI-level enforcement if bypassing it is trivial.
5. See the broader field at the Jev benchmark landscape roundup and the Awesome Jev radar for more early-stage projects worth watching rather than trusting outright.