Blog
Writing from Rekursor.
Results, methods and working notes on reliable AI systems that can improve without losing what already works.
All posts
- Jul 21, 2026
A New Bar for Legal AI on Harvey LAB
Rekursor's rubric-blind system achieved a 34% all-pass rate and passed 94.9% of individual criteria on a held-out, 50-task sample of public Harvey LAB tasks.
- Jun 16, 2026
Continual Learning Needs a Bouncer
We demonstrate continual learning on the sequential Atari problem John Carmack describes: an agent discovers a skill from its own failures, proves it helps, and keeps prior games intact. Remove the verification gate — the bouncer — and the forgetting comes right back.
- Jun 2, 2026
Beyond Frontier Model Performance for Legal AI Agents
Five results on Harvey LAB: held-out library transfer, autonomous all-pass revision, autonomous library generation from a firm's own graded work, reliability of revision, and a scaling-law curve where Rekursor's routing holds as RAG collapses.
- May 20, 2026
Continual Learning on Harvey's Legal Agent Benchmark: First Results
A first result on Harvey LAB's open legal benchmark: with our learning layer attached, the same agent and judge go from 45/48 to 48/48 on a corporate-governance task, with zero regressions.
- May 13, 2026
AI Agents Are Not Magic. Here's What Actually Matters.
A working filter for the AI agent moment: what an agent actually is, where the basic version fails, and four questions to ask any vendor.
- May 1, 2026
Why AI Agent Learning Plateaus
Why AI agents plateau on autoresearch loops, and what breaks through. A controlled comparison on Karpathy's autoresearch fork.
- Apr 16, 2026
Rekursor: AI Agents That Keep Getting Smarter
A learning layer that makes any frozen model smarter with every task — no retraining, no weight changes, fewer tokens. Results across Terminal-Bench 2.0, SWE-bench, EDGAR, and drug discovery.