← Trust LayerThe institution

Queldrex is built as a government

Not a scanner and not a vendor of alerts. An institution that governs the actions AI agents take, and is itself accountable through a record no one can rewrite. A government makes rules, enforces them, judges what happened, keeps the record, and answers for it. A system that only detects and blocks is policing, and policing without adjudication, record, or appeal is not legitimate authority.

One precise choice runs through all of it: we govern actions, not agents. An AI agent cannot consent and cannot be held responsible, so we bind each governed tool call to the accountable operator instead, and scale oversight to how much autonomy the agent was given. That is the form the emerging literature on separation of powers for AI is converging on, and it is why this institution is complete within its honest limits even though it can never govern an agent that never routed through it.

Separation of powers is a security property here, not a slogan. If the component that detects is also the one that judges and also the one that keeps the record, then compromising any one of them compromises all three, and nobody can prove what happened. That is why the organs below are separated in code, and why each one is marked live or building honestly rather than described as finished.
5
Separated organs
20
Published laws
2026.07.2
Ruleset in force

The organs

1

Legislature

Owns: Declares the rules.

Cannot: Makes no decision about any specific case.

A corpus of citable laws, each with a permanent id, plain-language text, a rationale you can argue with, the ruleset version it came into force under, and whether a finding under it can be contested. A rule you cannot read is one you cannot comply with, so the whole corpus is public.

2

Executive

Owns: Applies the rules to actions.

Cannot: Cannot write them.

Enforcement runs inside your own process as a zero-dependency package. Every allow, require-approval, or deny is reachable deterministically with no model call, and each decision now names the law it was made under, so a record reads "decided under QX-L-0004" instead of just "decided".

3

Judiciary

Owns: Determines what happened, assigns proportionate consequence, hears challenges.

Cannot: Cannot change the rules to fit the case.

Pending decisions route to a human, and each one now names the law it was made under so a challenge can contest something specific. A contestable finding can be appealed from the console; overturning it records a signed exoneration on the same ledger that carried the accusation, as visibly as the original finding, and re-pins the version. The original is answered, never erased. A tripped canary stays not-contestable at decision time, and the appeal path says so rather than opening a case that could not succeed.

4

Archive

Owns: Keeps the record. Answers to nobody, including us.

Cannot: Cannot quietly rewrite what it already recorded.

Each decision produces an Ed25519 receipt a third party can verify offline, with no Queldrex account and without trusting this page. Blocks are sealed to a hash-chained ledger. A Merkle commitment over that ledger lets anyone prove a specific record is included, seeing only sibling hashes, and a consistency proof lets anyone confirm the log only ever grew, that nothing already recorded was removed, reordered, or rewritten. Both verify offline with our published package. The ledger head is now timestamped daily at an independent authority, so even we cannot back-date or rewind a root a third party already witnessed. A public transparency-log anchor (Rekor) is the fuller version still to come.

5

Constitution

Owns: The shared standard that binds Queldrex too.

Cannot: Cannot exempt its own author.

The open Trust Receipt spec and the published npm verifier are how anyone holds us accountable without our permission. It was built first, which is the correct order: the way to check us existed before the thing being checked.

Rights of the governed

You can read every rule before you are judged by it, and a stricter rule shipped tomorrow does not make yesterday's receipt a lie.

A decision names the law it was made under. The full reasoning goes to the record and the account owner, never to the agent being governed, which would be a map for an attacker.

Contest a specific ruling from the console and have a reversal recorded as visibly as the accusation, on the same tamper-evident ledger, verifiable offline. The original finding is answered, not erased. A finding a human can overturn on the record is safer to make than one that cannot.

Consequences are graduated and recoverable before they are permanent. There is no automated action at the scope where a wrong decision would take a whole fleet offline at once.

We commit to the record, never to your payloads. A Merkle commitment over the ledger lets anyone prove a specific entry exists and was not altered, seeing only sibling hashes. Your operations stay yours; only the commitment is permanent.

We measure our own judge for drift and publish it. The deterministic layer scores the whole corpus repeatedly with zero verdict variance, and every change across ruleset versions is attributable to a named, dated rule rather than to silent drift. This is the piece almost nobody does.

Restraints on our own power

An institution is defined as much by what it refuses to do. These are limits on Queldrex, not on the governed.

Where this sits in the field

We did not invent these ideas, and pretending otherwise would fail the first thing this institution is built on: being checkable. Each governed agent is a non-human identity with a scoped set of permissions, and the structure below has real prior art. What we bring is that it is shipped, commercial, and verifiable, not a proposal.

Separation of powers for AI agents
Prior art: AgentCity (arXiv 2604.07007) and centuries of constitutional design
We ship it as separated code with signed evidence, rather than describe an agent economy.
Capability envelope
Prior art: AgentBound (arXiv 2606.30970); the envelope metaphor predates it
Our executive applies a bounded permission set per agent, and adds a pre-call scan and a receipt.
Contestability, the right to contest a decision
Prior art: GDPR Article 22(3); a large contestable-AI literature
Our appeals path is aligned with the principle. It is not a compliance claim; these regimes govern decisions about people, not agents.
Non-human identity, graded by how it was established
Prior art: NIST SP 800-63 IAL/AAL; OIDC AMR/ACR
Our trust-rung ladder stamps how firmly an identity was proven, so evidence is never overstated.
Transparency log with inclusion and consistency proofs
Prior art: Certificate Transparency, RFC 6962 / 9162; in-toto; Sigstore
We apply the same construction to agent actions, verifiable offline with our published package.
Legitimacy comes from being more auditable than the things we judge. Payment never changes a verdict. Verification does not depend on trusting us. The rules, the benchmark, and our own limits are public. A company that appoints itself to govern AI has to earn that, and the way it earns it is by being checkable. Read the honest limits on the threat-model page, the concepts we build on credited here.
Institution model versioned with the rulebook. Ruleset queldrex:baseline 2026.07.2 · 20 published laws · the pieces marked building are tracked openly and are not described as finished until they are.