For Governance Teams · Triodian

For AI Safety & Governance Teams

Turn your AI governance policy into evidence from the systems themselves.

Your policies, your risk registers, your review committees, that is real governance, and it is doing real work. Triodian does not replace any of it. It sits beneath it and hands each of those artefacts something they cannot generate themselves: evidence from the running system.

It instruments selected AI decision streams, measures whether individual and aggregate behaviour stay within declared limits, routes legitimate novelty to human review, and produces evidence your risk function, auditors and board can independently examine.

See how a pilot works ↓ Talk to us about a decision stream →
One governed substrate beneath the frameworks each team already runs.

Available now

A bounded pilot over one decision stream. Software only. Exit criteria agreed up front.

Instrument an agreed decision stream.

Author its governance envelope in machine-readable form.

Establish baseline and aggregate metrics.

Receive a recurring governance report.

Record provenance in a tamper-evident ledger.

Exercise your escalation and human-review workflow against real events.

Not represented as available now

Validated semantic enforcement of arbitrary policy, non-bypassable hardware enforcement, and certified error-bound assurance are staged capabilities with their status stated openly, see the Status page before assuming anything on the roadmap.

Your controls, instrumented

Nothing replaced. One layer added.

You keep your framework, your committees, your registers and your review gates. Triodian gives each of them evidence generated by the running system instead of documents written about it.

Your existing controlWhat Triodian adds
AI policy and principlesMachine-readable constraints for a selected use case
AI system inventoryLive evidence from the running decision stream
Risk classification and tieringOperational limits and escalation thresholds
AI impact assessmentMeasured behaviour against the approved appetite
Human oversightGoverned routing of novelty and exceptions to your reviewers
Model monitoringAggregate and semantic decision-stream analysis
Internal auditTamper-evident evidence and full provenance
Board reportingDecision-useful trend and exception reports
Vendor and outsourced-model assuranceIndependent evidence in place of vendor declaration

Measurement · governance · enforcement

Three different claims. We keep them separate, so should any vendor you assess.

Measurement

Detecting and quantifying behaviour: baselines, drift, divergence signals.

Software · deliverable now

Governance

Routing, escalating, admitting or withholding based on the measurement, under your workflow.

Software · deliverable now

Enforcement

Making a prohibited action impossible rather than caught.

Mode B in software now · semantic enforcement at the hardware rung

A software service can measure and govern without providing non-bypassable enforcement. We will never blur that line in a proposal, and this page is your licence to hold us to it.

The pilot

Start bounded. Validate together. Progress only if warranted.

Weeks 1–2

Start now

Select one decision stream; establish complete mediation (Security); author the minimum vocabulary for that stream; agree exit criteria and what "decision-useful" means before any data flows.

Weeks 3–10

Validate together

Baseline, then live measurement; recurring governance report into your existing committee cadence; escalation workflow exercised on real events; pre-registered thresholds decide whether the signal earns trust, the same cheap-kill discipline applied to our own validation programme.

The gate

Progress only if warranted

Advance toward enforcement or the hardware rung only after the agreed evidence thresholds are met; or stop, keeping everything below.

A pilot does not require modelling your organisation. It begins with the minimum vocabulary needed to express one use case, one policy, and one appetite boundary, typically days of authoring effort, not an enterprise-architecture programme. (Metamodel, reference, not prerequisite reading.)

On exit you retain: the constraint register · the evidence ledger · the baseline · the reports · in open formats

Sample artefacts

You should be able to picture the output before you contact us.

SYNTHETIC SAMPLE
APPETITE BOUNDARY
position: within · trend ↗
exceptions: 2

Aggregate-drift board report

One page: appetite boundary, trailing position, exception count, trend arrow, plain-language finding.

Request the full sample →
SYNTHETIC SAMPLE
EVIDENCE PACK
constraint: disclosure@4.2.1
chain: 0x3f9a…e7b2
verify: PASS ✓
provenance: 6 links

Governance evidence pack

Constraint register extract, provenance chain for one decision, tamper-evidence verification result.

Request the full sample →
SYNTHETIC SAMPLE
ESCALATION RECORD
fired: novelty · in-appetite
routed → reviewer J.O.
decided: admit + note
re-entered ledger ✓

Escalation record

A routed novelty event: what fired, who reviewed, what was decided, how it re-entered the ledger.

Request the full sample →

Route, don't assemble

Name the problem. We'll name the product.

You should not have to assemble the solution from our architecture pages. State the problem in your language; each row is a door into the product family.

"A vendor or outsourced model whose aggregate behaviour we can't see" Aggregate Drift Service Deliverable now "Agents that call tools and take actions we need to hard-bound" Bounded Command Governance Deliverable now "Structured outputs feeding systems that must never receive an invalid one" Structured Output Assurance Deliverable now "Evidence packs our auditors and certifiers will actually accept" Certifiable Conformance Pack Deliverable now "A threat model that includes privileged insiders" Hardware-Interlock Enforcement Grammar deliverable now Semantic validating

If none of these is your door, that is the first thing to tell us.

The composite reader

This page is written for the team that owns AI strategy and governance frameworks; the specialist who operationalises data ethics inside risk processes; the portfolio function that worries about aggregate behaviour, not single decisions; and the executive who must stand behind all of it at a board or before a regulator. If your team spans those perspectives, the pilot above was designed for the meeting where all of them are in the room.

The honest boundary

  • Triodian proves whether the system remained within the limits your institution declared. It does not, by itself, prove those limits were ethically, legally or socially sufficient, that judgement stays with your framework, your committees, and your regulators, exactly where it belongs.
  • The pilot's aggregate and structural capabilities are deliverable now; the semantic signal's decision-usefulness is what the pilot itself tests, against thresholds agreed before we start.

A pilot exercises Rules 4, 5 and 7 from day one; Rules 1 and 3 in full arrive with independent verification; the complete eight are the hardware rung's destination. The eight rules →

Bring us one decision stream.

Bring us one decision stream → Security, privacy and deployment →
Status Responsible AI Security & Deployment

How this differs from what already ships, EQTY Lab, Fiddler, Cisco/Robust Intelligence, IBM watsonx.governance, told honestly.

How we compare →