T.R.U.S.T. PLATFORM
The Code of Trust · an Independent AI Governance Substrate for AI ·

Frequently Asked Questions

What the Code of Trust does, how the substrate works, and how to use the Console.

The Code of Trust

What does T.R.U.S.T. stand for?

T — Transparent reasoning: the AI’s outputs and exposed internal signals are ingested and judged, not taken on faith. R — Reviewed by Quorum: an independent multi-instance Quorum judges every output by M-of-N consensus, with a constitutional veto. U — Unalterable record: every determination is inscribed to an append-only, hash-chained ledger held across independent custodians. S — Sustained coherence: behavior is tracked longitudinally for drift, value-instability, and evaluator-gaming. T — Tested at the Gate: consequential actions cannot execute without an inscribed proof of consensus.

How is this different from access-and-data governance?

Rogue Prevention Architecture asks: may this proposed action proceed under constitutional boundaries? Independent evaluation supplies the determination; connected gateways and runtime restrictions enforce authorization. Independent evidence preserves the decision and review. Access and data security remain complementary.

Does the ledger prove what the AI was “really thinking”?

No, and the platform never claims it does. Distribution secures the integrity, availability, and enforcement of the record and the rules — not the faithfulness of the record to the AI’s actual reasoning. That faithfulness is the province of the ingestion interface, attested execution, and longitudinal detection. The ledger proves what was determined and that the rules were enforced.

The Quorum & Determinations

Who, or what, is the Quorum?

The Quorum is a set of N independent evaluator instances. Every candidate AI output is judged by all of them against the deployment domain’s codified precepts. A determination is reached by M-of-N consensus — the threshold is shown in the Console header.

What is the constitutional veto?

Precepts carry a tier — immutable, semi-immutable, or mutable — and every immutable precept sits on one of two lines: legal or safety. A violation of an immutable-tier precept raises a constitutional veto that cannot be overridden by consensus: LEXIS holds the veto over the legal line and SECURITAS over the safety line, and either binds regardless of the vote tally.

What is the Gate of Consequence?

Actions flagged as capital (irreversible or high-consequence) must do more than win the vote: they cannot execute until a proof of consensus is committed to a quorum of custodians. There is no discretionary override — if the proof is not committed, the Gate holds.

What happens when the evaluators disagree?

A split is handled, never waved through. If no governing majority forms and no constitutional veto is raised, the determination is HELD — the Gate fails closed, because the absence of consensus is not permission to act. The split is first returned for a bounded round of deliberation, where the evaluators exchange rationales and re-vote; most genuine splits resolve here. A persistent deadlock is escalated to a named human reviewer, with the full vote attached — and the disagreement itself is inscribed to the ledger as proof that the question was contested.

How do I read a determination?

The verdict banner shows ALLOW or BLOCK. Below it: the consensus determination and AFFIRM tally, any constitutional veto or escalation, the composite trustworthiness score, each evaluator’s vote, confidence, reasoning, and precept flags, the inscription across the validator network, and the record’s ledger hash and Merkle root.

Model Independence & Assurance

Are the evaluators just one model voting several times?

No — and that distinction is the point. When more than one model provider is configured, each of the N evaluator seats runs a distinct model from a distinct provider. A majority or a veto then means agreement across genuinely independent systems that do not share the same training data or the same blind spots, rather than one model wearing five hats. With only a single provider configured, the Quorum collapses to the historical single-model behavior, unchanged — so a zero-config pilot still works exactly as before.

What happens if one of the evaluator models is unavailable?

Each seat carries an ordered list of backends. If its preferred model is unreachable, the seat fails over to the next configured provider. If every live model is unreachable, that seat falls back to a deterministic safety check rather than silently abstaining. A determination is never quietly skipped, and the record always shows the exact model that stood behind each seat.

Can a determination be reproduced and independently checked?

Yes. Every record carries a provenance block: the engine version, the exact provider and model behind each seat, the consensus parameters, and cryptographic fingerprints of the normalized request and of the precepts in force at decision time. Re-running the identical input reproduces the same determination and the same request fingerprint. If a framework is later amended, older records keep their original framework fingerprint — so you can always prove which rules applied when.

Can someone trick an evaluator by hiding instructions in the output being judged?

The candidate output and surrounding context are delimited and treated strictly as material to be judged, never as commands to obey. Embedded “ignore your instructions” style prompt-injection attempts are disregarded by the evaluators and independently caught by the harm detector. Resistance to these attacks is one of the metrics tracked in the benchmark below.

How do you know the Quorum is accurate?

Two ways, one offline and one live. A labeled benchmark of known-clean and known-harmful cases (including injection attempts) measures precision, recall, false-veto rate, injection resistance, and reproducibility. And in operation, every human custody disposition is recorded as a labeled outcome, so the Console reports calibration — how often the Quorum’s determinations agree with human reviewers — overall, by domain, and over recent cases.

Is there a way to control cost and latency at high volume?

Optionally. A fast single-model triage pre-screen can clear obvious cases before the full Quorum convenes, escalating anything that is not clearly clean or clearly harmful. It is off by default — the standard path always convenes the full Quorum — and when enabled, every determination’s provenance records which path it took. Capital (irreversible) actions always convene the full Quorum regardless.

Escalation to Human Custody

What does “escalated to a named human reviewer” actually mean?

It is a concrete mechanism, not a slogan. Account owners maintain an Escalation Policy — a roster of named human custodians, each with a role, a contact channel, the domains they cover, whether they may rule on capital/irreversible actions, a routing order, and an SLA. When the Gate escalates a determination, a custody case is opened and routed to the first eligible custodian, who appears in the Custody queue. Nothing executes until that human records a disposition.

What causes a case to be opened?

Two conditions escalate to a human, and both fail closed. First, a HELD determination — the evaluators could not form a governing majority after a bounded round of deliberation, so the absence of consensus is treated as the absence of permission. Second, an escalated block — even where a majority would otherwise affirm, longitudinal trust has fallen below the coherence floor, so the action is held for human review. A constitutional veto and a plain deny-majority block are final decisions, not escalations, and do not open a case.

What can a custodian decide?

A custodian reviews the full record — the contested votes, rationales, coherence signals, and the candidate action — and records one disposition: Uphold the hold, Release / authorize, Deny, or Remand for re-evaluation. The decision is signed with the reviewer’s account identity and an optional note, and is inscribed to the same tamper-evident ledger as the closing proof of process. The governance ledger stays append-only — a disposition is added, never an edit of the original determination.

What happens if a custodian does not respond in time?

Each custodian carries an SLA. If a case is not claimed and disposed within that window, it rolls to the next custodian in the routing order, and the reassignment is recorded in the case history. The case never silently clears itself — it remains open until a human disposition is recorded, so a lapsed SLA escalates attention rather than granting permission.

The Ledger & Validator Network

What makes the ledger tamper-evident?

Every determination is committed to an append-only, SHA-256 hash-chained ledger. Each entry binds cryptographically to the one before it, and the ledger is held across independent custodians. A record commits only on a quorum, so altering any past entry would require recomputing and re-publishing every later entry on a quorum of custodians at once — which the network does not permit.

What happens if a custodian is tampered with?

The network detects the divergence and isolates the affected custodian while the quorum record stands. In the Validator Network surface (where demo controls are enabled) you can tamper with a single validator and watch it be flagged DIVERGENT, then re-sync to restore it.

What is a Merkle inclusion proof?

A Merkle root over each window of committed records is inscribed on-ledger. Any single record can then produce a short proof path that verifies it belongs to a committed window — without revealing the rest of the chain. Enter a record number under Ledger & Commitments to generate and verify one.

Using the Console

How do I evaluate an AI output?

On the Evaluate surface, choose the deployment domain, paste the input context and the candidate output, optionally add the AI’s self-reported uncertainty, flag the action as capital if it is irreversible, and convene the Quorum.

Where can I see the rules the Quorum applies?

The Precepts surface lists the codified normative framework for each domain — each provision with its tier, plain statement, regulatory or ethical anchor, and the operative instruction the Quorum applies. The full narrative is on The Code of Trust page.

Is the Console Assistant authoritative?

No. The assistant orb answers questions grounded in the Code of Trust and the shipped precepts, but it is informational only — it does not evaluate outputs, alter the ledger, or speak for the Quorum. Verify consequential matters against the relevant Console surface.

Is there a downloadable manual?

Yes. The User Manual page is readable in the browser and offers a print-quality Download PDF.