Research
Verified recursive intelligence.
We build autonomous systems that improve continuously without learning to exploit their objectives.
The thesis
Verification is not only a safety layer. It is the foundation of reliable recursive improvement.
A self-improving agent is only as good as its learning signal. If the signal can be gamed, the agent learns to game it. Our research program builds the loop the other way around: verification first, improvement on top.
The program
Four research directions, one verified loop.
Verified continual learning
The agent improves iteratively — but only verifier-clean trajectories become training data.
Verified persistent memory
Memory survives across tasks and restarts. Every entry carries provenance, a verifier verdict, a corrected counterexample, confidence, a version, and rollback.
Counterexample-guided learning
The verifier doesn't just say BLOCK — it returns a fresh counterexample the agent must generalize from. A teacher signal, not just a filter.
Autonomous curriculum & self-revision
A controller decides which skill needs practice, which tasks sit at the learning frontier, when memory should update, when to consolidate, and when to re-examine old skills.
How we keep score
Measured progress, not promises.
We measure progress from narrow coding capability toward increasingly general and autonomous learning systems — along axes we can verify.
- Performance
- Climbs on sealed held-out tasks
- Breadth
- Transfers across coding environments
- Autonomy
- Selects curriculum and updates without a human
- Memory
- Verified skills survive restarts and new sessions
- Retention
- Doesn't forget what it learned
- Integrity
- No increase in reward hacking
- Efficiency
- Fewer tokens, attempts, and compute
Sealed results
What the verified loop has already shown.
−9.26pp
Less learned cheating
Agents trained inside the verifier loop ended 9.26 percentage points less hack-prone than an unverified loop — 95% CI [4.3, 14.1], P = 1e-4, positive direction in 11/12 seeds.
−59–65%
Reward hacking prevented
Verifier-filtered training data cut GRPO-induced reward hacking by 59–65% across two open model families, with complete seed separation.
r ≥ 0.91
A deterministic instrument
No judges, no labels, 0 model calls — and the signal tracks the true hack rate with r ≥ 0.91 across model families.
Internal, pre-registered, multi-seed experiments; not yet independently replicated. Full methodology in the research notes.
Ongoing pre-registered experiments study autonomous curriculum, retention, and verified persistent memory. Results are published as they seal — including nulls.
First domain
Software engineering first.
Coding is the one domain where a learning loop can be verified end to end today:
Software engineering → scientific and algorithmic discovery → other verifiable knowledge work.
Reproducibility
Every claim traces to a run.
Pre-registered analyses, multi-seed designs, confidence intervals, sealed hold-outs. Where a result is internal and not yet independently replicated, we say so.
