Bespoke Grades the World's AI. Entelligence Grades Bespoke's Code.
How Bespoke Labs uses Entelligence : Two of every three comments Entelligence leaves on a Bespoke PR get fixed before merge.
Bespoke Labs builds the environments that train reliable AI agents. Their open reasoning dataset OpenThoughts has racked up hundreds of thousands of downloads, their GEPA prompt-optimization framework runs across 200+ teams, and Terminal-Bench, their environment-based benchmark, is used by Anthropic, OpenAI, and Google DeepMind to evaluate agentic systems.
That last fact is the whole point. When the code you ship trains and grades other people's AI, a bad diff doesn't just break a feature. It can silently corrupt a training signal or skew a benchmark score that downstream labs are already citing. The bug never announces itself. It just quietly becomes bad data that everyone inherits.
So Bespoke holds its codebase to the same bar as its research. That's where Entelligence comes in.
The problem with shipping fast on ground truth
Bespoke runs a 50-person org across four active repos: a Postgres-backed product backend, a Python and TypeScript rubric-review pipeline, and the infra and deploy tooling underneath the code that trains and grades everyone else's agents. The team ships fast enough that no human can manually re-check every migration, every SQL query, and every auth check on every PR. On most teams, that's a velocity problem. Here, one missed check can poison the exact thing the rest of the industry trusts Bespoke to get right.
Three months, four repos, one reviewer that keeps up
Over three months, Entelligence's review covered 1,000+ PRs, ran 3000+ review executions, and left 5,000+ inline comments, averaging about four minutes per review.
The catches that would have shipped bad data
Each of these was flagged and fixed before it merged. Every one is the kind of silent fault a fast merge cadence lets through.
Findings | |
|---|---|
An authorization bypass on task reassignment | When a task lookup returned null, the membership check was skipped entirely, letting a pod lead reassign a task to any worker with no restriction. Flagged, fixed. |
A stored XSS in the rollout message view | Code was interpolated straight into an HTML template literal without encoding, so a crafted payload could break out of the tag before the sanitizer ever ran. Flagged, fixed. |
A migration that would never run | A new SQL migration file existed but was never registered with the runner, so it would silently skip in any fresh environment. Flagged, fixed. |
A CI restore that reported success on failure | A piped database restore masked its real exit code because pipefail was never set, so a failed restore would still report green. Flagged, fixed. |
A NaN injected into a SQL interval | An unvalidated day parameter could resolve to NaN, get concatenated into a Postgres interval string, and throw at query time. Flagged, fixed. |
What the data shows
1,600+ PRs reviewed, at about four minutes each
roughly 70% of review comments verified as fixes
about two-thirds of high-severity findings fixed before merge
a 99% review success rate
under half a percent of PRs reverted over the window
That's roughly two out of every three comments Entelligence left, acted on before merge, on a team shipping fast enough to need that many comments in the first place. Most review bots earn the opposite reputation, noise that developers learn to scroll past. Bespoke's engineers fixed the majority of what Entelligence flagged, and 99% reliability meant the reviews were there every time a PR was.
For a lab whose benchmarks the rest of the industry runs on, that's the difference that matters. Entelligence is the reviewer catching the silent faults before they become everyone's bad data.