Layers
The kernel
- Every event is written to an append-only log framed with CRC32C checksums, with rotation and crash recovery.
- Large tool output is stored by hash in a blob store and kept out of the model’s context.
- A deny-first policy engine checks each
bashcommand against named rules; the first match wins.
The admission gate
The gate admits a candidate genome only when eight ordered predicates hold. In outline, from the Agent Kalki paper:- Candidate and incumbent are compared with k live runs per case, from matched starting states.
- Any loss on a protected set of cases that pass today, beyond a Bonferroni-corrected flake margin, rejects the candidate.
- Gains must exceed a measured noise ceiling τ.
- Token cost must stay within a bound κ.
- Held-out tasks guard against overfitting to the cases behind the change.
Repository layout
Related research
- Viveka: why “keep it if it scores higher” admits no-op changes half the time.
- The Harness Is the Lab: held-out self-improvement curves for AI scientists.
- Evaluation principles: the rules the gate encodes.
Contributing
Runcargo xtask ci, sign off commits with git commit -s, and open an ADR before changing the protocol, the log
format, the policy engine or the gate.
