Hand them the manifest, not a slide deck

The instinct in a model-risk review is to narrate: to walk the examiner through what the model does, how it works, why the team trusts it. The instinct is wrong. Narration invites argument — every sentence is a claim the examiner can contest, every explanation is a story the examiner can disbelieve. A record answers. The per-output manifest (BR-842) is the record: a hash-chained, append-only artifact where each field is a question the examiner would have asked, answered before they asked it. The move that wins the review is not to narrate the decision; it is to hand them the manifest and point.

We are pre-revenue and building the POC. This post is about why pointing beats narrating, and how the manifest is designed to be pointed at, not read aloud.

Every narration is a claim; every field is an answer

A narration is a sequence of claims: "the model used relevant data," "the team validated the output," "the decision was reviewed." Each claim is contestable; each claim invites the examiner to ask "how do you know?" A field is not a claim; it is an answer. "Which model? This field, hash-pinned. Which inputs? These NodeVersions, hash-linked. Can I replay it? Here is the replay anchor." The field does not invite "how do you know?" because the field is the how. The manifest is the submission that answers before it is asked, which is what an examiner reads as competence.

The fields that replace the narration

• Model identity — "which model?" becomes model_version, hash-pinned, not a name.
• Input snapshot — "which data?" becomes input_nodeversions, L1 to L4 refs, hash-linked.
• Agent trace — "which agents?" becomes agent_trace, CRDT-merged, which of the 23 participated and in what order.
• Output and chain link — "what came back?" becomes output_bytes, ts, and chain_link, sealed at the moment of production.
• Replay anchor — "can I check?" becomes replay_hash, the field that lets the examiner verify, not trust.

Why the slide deck fails

A slide deck is a narration in bullet points. It tells a story the examiner can disbelieve without contradicting any evidence, because there is no evidence in the slide — there is only a claim. A manifest is evidence; the slide is a story about evidence. The examiner who reads the slide and then asks for the underlying record is doing the work the team should have done; the examiner who reads the manifest is doing the review. The difference is who is doing the work. The manifest is the submission that lets the examiner review instead of reconstruct — and the six weeks of archaeology that used to be the audit is the reconstruction the manifest replaces.

Narration invites argument; a record answers. The slide is a story about evidence; the manifest is the evidence. The difference is who is doing the work.

What the manifest does not narrate

The manifest does not narrate accuracy. It records provenance and reproducibility — never truth. The claim "the decision was correct" is a claim the manifest refuses to make, and that refusal is what makes everything else it records credible. Calibration is a separate, gated loop (Gate 1b): shadow runs against ground truth, scored with Brier and log-loss via deino-eval. The manifest is the audit story; the calibration loop is the accuracy story. A firm that hands the examiner the manifest is handing over the record of how the decision was made; a firm that hands the examiner a slide is handing over a story about how the decision was made. The examiner can tell the difference.

Hand them the manifest, not a slide deck. Pointing is the move; narrating is the mistake. The record answers before the question is asked, and that is what an examiner reads as competence.

— review submission note

The honest scope

We are building the POC, pre-revenue. The manifest fields exist; calibration is gated. What the manifest gives a firm is not a guarantee of a passed review — it is a guarantee that the review is a review of a record, not a reconstruction of a memory. Six weeks becomes six hours because the substrate answers the question the regulator asks, not the one the firm wished it asked. The manifest is the question and the answer in one artifact; the slide is neither.