Integrity — Governed Decision Path

How we govern our own AI agents.

Our platform's own agents make tool calls — starting invocations, holding credit, dispatching work. Those calls are watched by an internal control plane that projects a live event ledger through the same abstract model our formal specs are checked against, and flags the runtime state that model says should be unreachable. This page is about that system, not about betting picks, odds, or grading — see /accountability for that.

Every claim below links to something you can independently check — a live receipt route, a published public key, or a document in this public repository. Nothing here is asserted without a way to verify it yourself.

What we govern

Detection

A live projection, watched

Every internal tool call folds the recent event ledger into an abstract state (claim phase, exposure phase, pending-attempt count) — the same shape our TLA+ specs reason over. Two forbidden states are watched for: two-or-more concurrently pending attempts on one invocation, and a rejected fingerprint with nothing binding it.

See the architecture pipeline

Posture

SHADOW is the only default

Detecting a forbidden state does not block anything by default. SHADOW mode always admits and only ever writes an evidence row. An enforcement mode exists, but it is reachable only from an explicitly lab-gated code path — never from a production route, cron, or worker.

See the enforcement posture

Evidence

Signed, publicly verifiable receipts

When a gated tool call is refused (lab-only today), it produces a signed receipt: what was asked, why it was refused, and an ed25519 signature over a canonical payload. Anyone can fetch our published public key and verify a receipt themselves — no shared secret required.

View the public keyring

Shadow evidence & active certificate

We do not publish a live shadow-would-refuse counter on this page — a number here with no way for an outside reader to recompute it would just be a claim, not evidence. Instead, the honest, independently-reproducible numbers live in one place:

Real TLC model-checking receipts, real state/depth counts — §3, "What is PROVED".
Whether any certificate version is currently active (today: no — activation is a human-only CLI action, never automated) — §10, "Active certificate".
Seven concrete steps to verify all of this yourself, not just trust it — §11, "Attack checklist for outsiders".

See it run

90-second walkthrough

Force a REFUSE, verify the signature

A scripted, end-to-end walkthrough: trigger a real violation, get a signed REFUSE receipt, open it over HTTP, fetch the public keyring, and verify the signature — all runnable from a clean checkout.

Read the demo script

Full record

The single honest status document

Architecture, what is actually proved (with real receipt numbers), enforcement posture, a full SHA-traceable receipts index, and every explicit non-claim — kept current against main, not a point-in-time snapshot.

Read SRQC_STATUS.md

Two different things called "integrity"

This page is about one of them. Keeping them separate matters — they share a value but not a subject, and merging them would let a strong claim about one imply an unearned claim about the other.

This page

Agent control-plane governance

How our own AI agents are governed when they call tools: admission bookkeeping, the SHADOW-by-default posture, and signed receipts anyone can verify against a published key. Subject: our software's behavior.

Elsewhere

Whether a published number is substantiated

Whether a rate, edge, or record we display is backed by coverage, a lower bound, CLV, and walk-forward provenance — and is withheld when it is not. Subject: the claims we make about picks.

How the gate decides Glass Ledger Accountability Proof of record

A signed receipt proves an agent decision happened as recorded. It says nothing about whether a betting claim is statistically supported — that is the Glass Ledger's job, and it stays sealed until the evidence clears its own bar.

What we do not claim

Not a parameterized (∀N) proof. Every model-checking result behind this page is a fixed-constant, finite-cutoff check, not a machine-checked universal statement.
Not a production enforce-by-default posture. SHADOW is the only default anywhere this admission check runs. Enforcement is reachable only from an explicitly lab-gated path, confirmed by direct code search, not assertion.
Not a claim about bet-settlement correctness or any user-facing betting logic. This entire surface is about internal AI/agent tool-call admission bookkeeping — nothing about odds, grading, or picks.
Not a SOC 2, ISO 27001, or EU AI Act certification. Related internal alignment work exists and is documented separately, but nothing on this page is a certification claim.
Not autonomous. Certificate-version activation is a human-only, manually-run action — never triggered by CI, a cron job, or any automated caller.