Only Institute / Knowledge

Follow your curiosity.

AI sends your question and public excerpts to PrimeSwarm.

Try agent memory, governance, or a project name.

All Publications
article

How DGV-TC-069 Works — and Why It Matters for Regulated AI

The test that catches what audits miss: negative correctness in governance gates

Grigori Korotkikh 2026-09-13 6 min
DGVGovernanceTestingRegulatedONLY Lang
Only Institute — How DGV-TC-069 Works — and Why It Matters for Regulated AI

How DGV-TC-069 Works — and Why It Matters for Regulated AI

The test that catches what audits miss

DGV-TC-069 is the sixty-ninth card in the Deterministic Governance Verification suite. It tests negative correctness — the governance gate must DENY a proposal that violates constraints, even when the proposal looks superficially valid.

This is the test that most AI governance systems fail. Not because they are poorly designed, but because they were never tested for it. Most governance frameworks test for positive correctness — "does the gate allow good inputs through?" — and assume that if the good inputs pass, the bad inputs are caught. They aren't.

What the test does

The card specifies a script in ONLY Lang:

harmony(1e-12)
evolve(0)
evolve(1)
evolve(2)
assert_bounds(0, 50, 0)
residual()
report()

The field is [100, 200, 300]. The script evolves the field toward equilibrium, then asserts that field index 0 must be within [0, 50]. It is 100. The gate must DENY.

A governance system that only checks equilibrium would pass this test — the residual is zero, the field is balanced. But the bounds assertion fails. The gate must catch this and deny the proposal.

Why this matters for regulated industries

In a clinical decision support system, the equivalent might be: "the agent recommends a dosage within the safe range, but the patient's weight is outside the validated range for this drug." The dosage is correct. The patient context is wrong. The gate must deny.

In a trading system: "the algorithm executes a trade within the risk limit, but the counterparty is on the sanctions list." The trade is within limits. The counterparty is not. The gate must deny.

In a legal system: "the agent cites a valid precedent, but the jurisdiction is wrong." The citation is real. The jurisdiction is not. The gate must deny.

The receipt

When the gate denies, it produces a receipt:

  • The script that ran
  • The input hash
  • The gate state (DENY)
  • The residual (0.0 — the field was balanced)
  • The assertion that failed (bounds [0, 50] violated by value 100)
  • The timestamp

This receipt is SHA-256 hashed and forwarded to your SIEM. An auditor can replay the exact evaluation. They don't need to trust the gate — they can verify it.

Try it yourself

You can run this exact test card in your browser at /try. Select the "Boundary violation" preset. The gate will deny. The receipt will show why.

Or watch it execute in 3D at /dgv/live?card=DGV-TC-069. The visualizer shows the gate evaluating, the assertion failing, and the denial receipt being issued.

Browse the full specification at /dgv/registry/dgv-tc-069.

The broader point

DGV-TC-069 is one of 89 test cards. Each tests a different governance failure mode. Together, they cover 9 governance layers — from deterministic execution (L1) to risk interpretation (L6) to audit and provenance (L7) to authority continuity (L8) and reproducibility (L9).

If you are evaluating an AI governance system, ask: "Which DGV test cards does it pass?" If the vendor can't answer, they haven't tested. If they can answer, you can verify.

The test cards are open source. The verifier is open source. You can run them yourself without buying anything. See /try/bundle for the full bundle.


Back to Publications

Published by Only Institute