The audit reads the substrate, not the slide

The slide is the story; the substrate is the evidence. An examiner listens to the slide and then asks for the substrate, because the slide is narration and the substrate is what the narration is about. A program that brings a polished slide and a thin substrate has brought the story and not the evidence — and a review is not a pitch. The manifest is the substrate; the slide points at it. Point, do not narrate; narration invites argument, a record answers.

The substrate is the record

The substrate an audit reads is the per-output, hash-chained manifest (BR-842): the model version (hash-pinned), the input NodeVersions (L1-L4 refs, hash-linked), the agent trace (the 23-agent A2A swarm, CRDT-merged), the override (AxisOverride, sealed), the output bytes, and a replay anchor. Every field is a question the examiner would have asked, answered before they asked it. The slide summarizes; the substrate is the thing the summary is a summary of — and the substrate is what verifies the summary, because the substrate is the computation, not a description of it.

Replay is the test the slide cannot pass

The thing the substrate can do that the slide cannot is be replayed. A third party with the replay anchor can regenerate the identical bytes — the full-chain-walk that Gate 1 requires. A slide cannot be replayed; it can only be reread, and rereading is not checking. The examiner who wants to know whether the decision was produced the way the firm says it was does not reread the slide; they replay the substrate. That is why a review that ends on the slide is a review that did not finish, and a review that ends on the substrate is one that did.

The slide summarizes; the substrate is what the summary is a summary of. Replay is the test the slide cannot pass — and the one the examiner actually runs.

The tiers all reach for the substrate

Every tier reaches for the substrate in its own force. The EU AI Act (binding law) requires logging and traceability that the manifest provides — Annex III points at the substrate, not the slide. SR 11-7 (supervisory guidance) expects model risk records that are retrievable and checkable — the substrate, again. NIST AI RMF and ISO/IEC 42001 (voluntary) document the governance that sits on top of the substrate. None of them accept the slide as the deliverable; all of them ask for the record the slide describes. Lumping them reads as someone who has not read them; naming the force reads as someone who has.

• The slide is the story; the substrate is the evidence. A review asks for the second.
• The substrate is replayable (Gate 1, the full-chain-walk); the slide is only rereadable.
• The manifest is the substrate — every field is a question answered before it is asked.
• Every tier — binding, supervisory, voluntary — reaches for the substrate, not the slide.

What the substrate does not make true

The substrate makes the output reconstructable, not correct. Certified ≠ True: the manifest attests provenance and reproducibility, and the one thing it refuses to attest is that the output was right. Accuracy is a calibration loop — Brier and log-loss via deino-eval, shadow runs against a realized timeline, Gate 1b before any forecast side is positioned as accurate. We are pre-revenue, building the POC; the substrate exists, and the calibration that would let us say more is gated. The slide is the story of the substrate; the substrate is the evidence the story is about; and neither is the truth of the output.

Point at the substrate, do not narrate the slide. The slide is the story; the substrate is what the story is about — and the audit reads what the story is about.

— review posture note