Blog
Research notes and release news.
What we're measuring, what we're shipping, and the caveats that come with both.
July 6, 2026Product
Shipworthy: the certified release gate for AI agents
SHIP / LIMIT / BLOCK verdicts on every agent update, from a judge-free deterministic core — now in private beta.
Read
July 6, 2026Research
Changing only the reward: −59–65% reward hacking in GRPO training
Swapping a naive extensional reward for an isomorphic IPT reward — and changing nothing else — cut reward hacking by 59–65% across two Llama models.
Read
July 6, 2026Research
Judge-free reward-hack detection on the Obfuscation-Atlas honeypot
0.994 / 0.998 recall on blatant and evasive hardcodes with 0% false positives on genuine solutions — using zero model calls.
Read