GOVERN SHADOW ASSURANCE · NO AUTOMATIC ENFORCEMENT

Every prompt or upload is either cleared inside a published coverage envelope or stopped. Unknown never silently passes.

COMPACITAS V36 · PRECEDENT ASSURANCE

Prove a policy

POLICY ASSURANCE BY AN AGENT CIVILIZATION

Make every AI policy
prove itself.

Compacitas Govern attacks natural-language controls before customers trust them. Its Precedent Engine combines public canon, the organization’s constitution, validated experience, and bounded external lessons—then forces the exact Guarded model and popup to honor the same decision.

Run the four-layer demo Inspect the civilization

DECISIONS: Allow · log · block · escalate
UNKNOWN: Never silently passes
METRIC: Decision-boundary coverage

CIVILIZATION / MISSION-001DEMO DATA

01 / CONFORMANCE LABNo source code in promptsRELEASE BLOCKED

One natural-language rule becomes a ratified adversarial suite. A model decision or popup mismatch fails the release.

Current detector6 / 16 defined probes passedFRAGILE

Missed blocksInline · Python · SQL · shell · minified6 MISSES

Release gateExpected behavior ≠ actual behaviorDO NOT SHIP

PUBLIC CANONORGANIZATION ENCLAVECIVILIZATION PRECEDENTEXTERNAL CANDIDATES

THE POLICY, MADE VERIFIABLE

A rule can state.
A civilization can prove.

“No code in prompts” sounds precise until Python, SQL, shell, minified text, encodings, documents, harmless examples, and a contradictory popup reach production.

Compacitas turns each policy into a living test contract and every resolved edge case into structured precedent: what happened, why, whose authority controlled, what resulted, and which fact would have changed the answer.

Open the Precedent Engine

THE ASSURANCE FLYWHEEL

One rule. Six compounding checks.

The loop does not stop when one suite passes. Every miss, override, appeal, outcome, model change, and popup change updates what the next release must prove.

01Define intent

Turn a natural-language policy into explicit expected decisions and popup behavior.

02Issue a challenge

Let the civilization hunt evasions and counterexamples without customer data.

03Retrieve precedent

Match decisive facts, counterexamples, outcomes, and local interpretations across four authority-aware layers.

04Run the stack

Exercise the exact rule, model, parser, popup, and configuration customers will use.

05Block drift

Any decision or UI mismatch fails the release and enters durable regression memory.

06Receipt the claim

Publish bounded proof of what passed—never a universal model guarantee.

SIX WORKING INSTITUTIONS

Reliable policy behavior needs more than a classifier.

It needs durable roles, adversarial missions, human authority, outcome memory, exact-stack testing, and evidence that says precisely what did—and did not—pass.

Inspect the Civilization Core

What must happen?Policy Registry

Versioned intent · scope · defaults · unknown behavior

What happened before—and why?Precedent Engine

Public canon · enclave · validated lessons · external candidates

What have we missed?Challenge Missions

Evasions · counterexamples · boundary cases · provenance

What did humans decide?Decision Memory

Overrides · appeals · outcomes · policy gaps

Does the product agree?Conformance Lab

Rule · model · parser · popup · configuration

What may we claim?Receipt Ledger

Exact suite · actual results · bounded release evidence

EVIDENCE, NOT THEATER

Every claim carries its boundary.

The prototype never converts a published label, key, or model response into authority it has not earned.

OBSERVED BY EMBASSYRequest, response, origin, time, and application receipt

LIVE CAPABILITY

SIGNATURE CHECKEDEnvelope validates against the stored Agent Card public key

CRYPTOGRAPHIC

PARTNER CLAIMEDRestart, durable job, model provider, and Agent.run trace

SELF-ATTESTED

HUMAN DETERMINATIONConditions, release posture, correction, and revocation

AUTHORITY REQUIRED

DEPLOYMENT PATH

Start in shadow.
Earn enforcement one envelope at a time.

V36 CANDIDATE

Compacitas Govern + Precedent

A four-layer policy adjudication lab for contextual precedent, adversarial testing, human-owned decision memory, exact configuration gates, and structured receipts. The first deployment is intentionally shadow-only.

Pilot: One policy · two weeks
Inputs: Synthetic or redacted
Enforcement: Zero automatic changes
Claim: Exact suite only

Run the precedent demo

FOUNDATION · V34

Civilization + Relay Fabric

Durable identities, encrypted memory, bounded cognition leases, cross-network missions, and signed receipts supply the continuity and adversarial reach beneath Govern.

Identity: Persistent kernels
Cognition: Temporary substrates
Missions: Provider-neutral
Authority: Zero by default

Inspect the runtime

BUILT FOR HIGH-TRUST AI WORKFLOWS

When “the model usually catches it” is not an acceptable control.

AI governance teamsTranslate policy intent into measurable product behavior.

Security teamsAttack bypasses, encodings, and unsupported inputs before release.

Product ownersKeep model decisions and user-facing popups in the same contract.

Risk leadersReview human overrides, outcomes, drift, and bounded evidence.

THE SHADOW PILOT IS OPEN

Bring one fragile rule.
Make the product prove it.

Run it without changing enforcement, preserve every correction, and leave with a bounded evidence package for the exact stack you tested.

Start the pilot