First-party evidence
The credible thing a pre-launch company can offer is not a wall of logos. It is a complete, honestly-graded account of what it has measured, what it hasn't, and the experiment that would move a claim from our word to an outsider's.
The honest boundary
Everything on this page is first-party evidence produced by Triodian. Independent assurance has not yet been produced. This page shows exactly what we have tested ourselves, how, and the independent validation we have committed to, and when.
The claim-status register
Grouped by what the claim is about, measurement, governance, enforcement, never blurred into one list. Each status badge links to the Status legend; each claim links to its methods definition on the Source Dossier.
Measurement
Set-level statistics detect a decision stream drifting out of a declared appetite even when every individual decision passes. Method →
internal-testedAdversary
A model optimising locally while portfolio exposure compounds.
Independent proof
Established statistics; a design partner reproduces the drift signal on their own stream at a pre-agreed sensitivity.
Distance in embedding space as a proxy for whether an action means what policy meant. Method →
design-intentAdversary
An output that satisfies the form of a rule while violating its intent.
Independent proof
Moves validating → proven when an independent lab shows measured divergence separates compliant from violating actions at a pre-registered floor, on held-out data no one tuned.
Governance
Semantic (Mode A) and structural (Mode B) constraints resolve to allow / redact / deflect / block. Method →
internal-testedAdversary
A prompt-injected or hallucinated tool call in the software path.
Independent proof
A red-team exercise reproduces the four dispositions on a held-out prompt set under agreed pass/fail cases.
Every meaningful action path passes through the governed boundary; alternates are removed or separately controlled. Method →
design-intentAdversary
An out-of-band route around the boundary, a second network path, tool, or display.
Independent proof
An architecture review by a third party confirms mediation for a named decision stream in a specific deployment.
Enforcement
A fixed-order hardware datapath no software at any privilege level can reorder or suppress. Method →
roadmapAdversary
A privileged insider or compromised host that owns the enforcement software.
Independent proof
Reduction to practice on an FPGA build, then a third-party lab attempts bypass under a published threat model.
A hardware root of trust signs that the loaded policy is unmodified before emission is enabled. Method →
design-intentAdversary
A substituted or tampered policy loaded into the enforcement path.
Independent proof
A security audit verifies the attestation chain end-to-end and publishes the findings table.
The harness
The harness runs the enforcement claims against declared cases and publishes the raw result set alongside the criteria that decide a pass. First-party. Not yet independently reproduced.
See the harness →Pre-registered pilot protocol
We are recruiting 3–7 institutions for time-boxed 60–90 day evaluations. Success metrics are agreed at kickoff; results are co-owned and publishable, positive, negative, or null.
Committed in advance
The hypothesis, the exact metric definitions, and the success/failure thresholds, a cheap-kill before the expensive work.
Who holds the evidence
The design partner or a neutral third party, not Triodian alone.
Publish either way
A negative or null result is published as readily as a positive one, and time-stamped.
Independent-validation roadmap
We commit to publishing the findings table, not a pass/fail badge.
An independent lab evaluates the enforcement and semantic-divergence claims against the pre-registered protocol.
ISO/IEC 42001, NIST AI RMF, ISO/IEC 27001, mapped now; conformance is a later, evidenced step.
A disclosure channel, then a staged bug-bounty programme as the appliance matures.
What we are not claiming