Risk assessment and transparency are different obligations

The conflation happens because both obligations live under the same law and both produce documents an examiner reads. They are not the same. A risk assessment evaluates a system before it is fielded — what it could do wrong, who could be harmed, what mitigations are in place. A transparency record attests a specific output after it is produced — which model, which inputs, which trace, can it be replayed. The first is prospective and system-level; the second is retrospective and output-level. A program that files the first and calls the second done has not done the second.

Two obligations, two artifacts

Under the EU AI Act (binding law), high-risk systems require both: a risk assessment (the FRIA and ongoing risk management) and transparency obligations (Annex III logging and traceability). They are separate articles, separate deliverables, and separate things an examiner asks for. SR 11-7 (supervisory guidance) expects model risk management — the risk assessment side — and treats documentation as part of it, not as a substitute. NIST AI RMF and ISO/IEC 42001 (voluntary) address both, but document them as distinct functions. Conflating them reads as someone who has not read the articles.

What the manifest is, and is not

The per-output manifest (BR-842) is a transparency artifact. It attests provenance and reproducibility for a specific output: model version, input NodeVersions, agent trace, override, replay anchor. It is not a risk assessment. It does not evaluate what the system could do wrong; it records what the system did do, on this output, at this time. A risk assessment is a separate deliverable — one that names the system’s intended scope (the stated limits), the foreseeable misuses, the mitigations, and the human oversight design. The manifest is evidence the transparency obligation is met; the risk assessment is evidence the risk-management obligation is met.

A risk assessment is prospective and system-level. A transparency record is retrospective and output-level. Filing the first and calling the second done is not doing the second.

Why the conflation is expensive

The conflation is expensive because it makes a program look done when it is half done. A firm that has a thorough risk assessment but no per-output transparency record can show the system was evaluated but cannot show how a specific decision was made — and the specific decision is what a review asks about. A firm that has the manifest but no risk assessment can show how a decision was made but not that the system was vetted before fielding. The examiner asks for both, in that order, and a program that has only one has the one the examiner asks for second.

• Risk assessment = prospective, system-level: what could go wrong, who is harmed, what mitigations.
• Transparency record = retrospective, output-level: which model, which inputs, can it be replayed.
• The EU AI Act requires both, as separate articles; SR 11-7 expects the risk side; voluntary tiers document both.
• The manifest is a transparency artifact, not a risk assessment — both are required; neither substitutes.

What neither artifact certifies

Neither the risk assessment nor the transparency record certifies that outputs are correct. The risk assessment certifies the system was evaluated; the manifest certifies the output is reconstructable. Certified ≠ True: both attest process, neither attests truth. Accuracy is a calibration loop gated by Gate 1 (reproducibility) and Gate 1b (beat a disclosed baseline), and we are pre-revenue, building the POC. The two obligations are about defensibility, not correctness — and defensibility is what the examiner can check, which is why the two artifacts are what they ask for.

The risk assessment says the system was worth fielding. The manifest says the output is reconstructable. Neither says the output was right — and the examiner never asked either of them to.

— obligation note