Notes from the workbench.
Real technical writing from the people building it: provenance, token efficiency, and verifiable memory.
- Provenance
The receipt, not the answer
Most AI tools ask you to trust the output. Provenance means handing back the receipt instead, so a claim can be checked rather than believed.
- Token efficiency
Atomic facts and the token budget
Stuffing whole documents into context is expensive and lossy. Querying one atomic fact at a time cuts tokens by a large margin and returns a hash you can verify.
- Engineering
Building verifiable memory
A walk through the design constraints behind a memory layer where every mutation is reversible, inspectable, and signed, and the daemon stays on your machine.