Splats as the Unit of AI Process
Abstract
If you want to audit what an AI actually did, the conversation is the wrong record. This paper argues for recording the smallest complete unit of work instead — what changed, in what order, under whose approval — and shows what becomes possible once that unit is hash-chained: replay, attribution, and a log a third party can check.
Terms used
splat — one recorded unit of agent work, hash-chained so it can be replayed.
meridian — the ordered spine a session’s splats are filed along.
The record problem
Transcripts are a poor audit surface. They mix intent with output, they are lossy about ordering, and they say nothing about which actions were actually taken against a real system. An audit that cannot be replayed is a story, not a record.
What a splat is
A splat is the smallest complete unit of agent work: an input, the action it resolved to, the diff it produced, and the approval that let it through. Each one is hashed and chained to the one before it, so the sequence cannot be silently edited after the fact.
What it buys you
- Replay — the sequence can be re-run and compared.
- Attribution — every change has a human decision attached to it.
- Verification — a third party can check the chain without trusting us.
The full argument, the schema, and the worked examples are in the PDF. The schema itself is public in splat-schema.
PENDING: this page currently carries the abstract and summary. Drop the full paper body in here, or keep the PDF as the canonical read — founder’s call.