RESEARCH00
Selected
research.
Papers and notes from the Transient model series. Receipts as training data, governance at the model level, long-horizon physical AI.
Receipts as training data
Governance receipts produced at runtime are not audit logs. They are labelled training examples, generated automatically as a side effect of governed agent action. This note explains what a receipt contains, why that makes it dense training material, and what the Rocky series demonstrated when models were trained exclusively on them.
Teaching models to read permission, not just action
Show a model the same action twice, with one permission variable changed. It has to learn why the label flipped. That training-side intervention produced a 110x lift in the model's ability to detect when permission had changed. Safety generalised first. Deployability is still being made reliable.
A purpose-built 6M model outperformed frontier AI at the decisions that matter.FLAGSHIP
A 6M-parameter model trained for one task: read the evidence, decide proceed or stop. It outperformed frontier AI 1,500 times its size. Paired with a deterministic rule layer, it reached 84% accuracy with zero unsafe decisions, without changing a single weight.
Rocky-α
A 6M-parameter model trained from scratch on structured Trace execution logs learned sharp allow, deny, and freeze boundaries in software-action contexts. Evaluated against scenarios it had never encountered, it almost never permitted actions it should have blocked. When it made mistakes, it stopped actions it should have allowed, not the reverse. Proof that governance receipts are dense training material for small, specialised decision engines.
Rocky-DAM-αFLAGSHIP
Rocky-DAM-α is a 6M-parameter model that reads physical sensor data alongside authority and governance evidence to decide whether infrastructure actions should proceed, be blocked, or halt. Tested on physical situations it had never seen during training, it showed novel decision making.