Surveillance & methodology
A partnership discussion with the agencies building this criteria class: attested telemetry that gives an AI-risk framework a live instrument, company-level, covenant-grade measurement under sector-level judgment.
The methodology gap
Every published view on an AI-dependent issuer rests on issuer-supplied information, management narrative, model inventories, attestation letters. The judgment is expert; the evidence basis underneath it is the issuer's own word.
Framework to instrument
Existing AI-risk frameworks run on financial statements. Attested telemetry gives the framework a live instrument: company-level, covenant-grade measurement that a sector-level view can rest on, rather than displace.
What the feed contributes to surveillance
| Machinery in the system | Surveillance analogue |
|---|---|
| Guarantee-invalidation event | A notifiable change in the evidence basis. |
| Re-certification (model change / shift / calendar) | A dated reset of the measurement. |
| Drift threshold breach | An accumulation signal at portfolio level. |
The systemic event
One bad model update propagated across every AI-dependent issuer at once is a single-event downgrade cluster no sector limit catches. The mechanism: model identity is on every token; unattested updates fall outside the attested stream, blocked at the gate or flagged outside covenant, before the correlation is on anyone's book.
"Portfolio and methodology teams should test that answer rather than take it."
The measurement
Distribution-free, finite-sample, assuming only exchangeability. The procedures are published: Angelopoulos, Bates et al. (RCPS, JACM 2021; Learn-then-Test; Conformal Risk Control, ICLR 2024). We apply them; we didn't invent them. A sample certificate is available for review.
How the bound is built →Honest boundary
Statistically certified action classes today are the Rules and Drift tiers. The Meaning tier joins only when its pre-registered experiment passes, and any criteria built on this architecture should say exactly the same. Coverage analysis →
The ask
1. Consideration of the four-conditions framework and the attestation evidence model as contributions to your methodology programme.
2. Where you judge it warranted, referral introductions. A counterparty can check the proof; the consumer of the rating gets the proof behind it.
There is no product ask.
Status, without varnish
The complete integrated system is not yet built and tested; research risk is concentrated in semantic governance. The USPTO application was filed 6 January 2026.
Triodian is not a lender, credit rating agency, valuation agent or financial adviser, does not hold an Australian Financial Services Licence, and does not issue, arrange or advise on credit or rating products.