In March you decided "the server is in Seoul." In June it moved to Tokyo. Who went back and pulled the decisions that were made on the Seoul premise? Nobody. Once AI agents start making decisions, this gap widens on its own — an AI forgets yesterday's premise, and when you ask why, it makes something up.
This is a layer that records each decision together with its premise, and when a premise later changes, hands back exactly the decisions built on that premise with "please re-review" — leaving the rest alone. It begins where storage and search (RAG) end.
Those numbers come from a pre-registered public bench — reproducibility package: trajectory-governance-bench
Pick a scenario: a fact, a decision built on it, an unrelated decision, then the premise changes. Watch the layer hand back only the decision that now needs re-review. This runs the real kernel live — not a recording.
When AI starts to have a body (robots, autonomous equipment), a layer that automatically pulls back the plans standing on a changed safety premise becomes a safety mechanism.
See a live audit ledger (fictional company): https://showcase.cra.kr/work/?share=386434cde6d709bb&lang=en
No sign-up form, no approval queue. Verify your email, get a key on the spot.
curl -X POST https://crag.kr/v1/note \
-H "Authorization: Bearer <your-key>" \
-H "Content-Type: application/json" \
-d '{"text": "We decided X because Y"}'Prefer GitHub? You can also request a key via an issue at trajectory-governance-bench/issues — the same repo hosts our reproducible benchmark (governance behaviors +0.35 over RAG at zero recall loss, pre-registered). trajectory-governance-bench/issues
Honest labels: this is a learning-stage alpha. Free, with hard daily caps. No launch dates promised. Endpoints: POST /v1/note, POST /v1/recall, GET /v1/pending, GET /v1/ledger, GET /v1/version (all require a key).