Trust Layer for Agentic AI

What it is

AI agents accumulate knowledge about the person who runs them — and nearly every memory system ships with the agent deciding what gets remembered. This is the inverse: a governance plane where every write is a proposal, every acceptance is a human ruling, and the ledger records both sides of that exchange.

The invariants I'll name

• Append-only event ledger. State is a fold over events, never a stored column. Same spine as everything else I build.
• Human-ruled writes. Nothing enters the knowledge base on a model's word — the same rule my claim ledger applies to politicians, applied to my own agents.
• Crypto-shred redaction. Deleting something means destroying its key. The hash chain stays intact; the payload is gone everywhere at once, backups included.

There are more invariants than this. They're the interesting part, which is why they're not on this page yet.

Status

This page promised to become a post-mortem if the feasibility probe failed. It failed, so here it is.

The probe ran seven days against three pre-registered numeric criteria and met none of them. The machinery held up its end — nearly two hundred candidates captured and distilled autonomously across five working domains, still running cleanly past the window — but the human ruling habit collapsed after day one, and the probe's own pre-registered hypothesis was confirmed by the data: session-level taint granularity is too coarse, and it mis-triaged 78% of candidates into the high-risk lane. That is the design correction v1 needed, measured rather than argued.

The design still stands; the build stays gated. v1 waits until the habit problem has an answer, not just the software. A probe you can't fail isn't a probe.