Public Release - 2026-05-25

No proof, no action.

SOVEREIGN is being shaped as proof-carrying cognitive action infrastructure for high-risk AI systems. The core rule is simple: an agent should not turn a plausible story into action unless the evidence chain is closed.

Proof-Carrying Action Open Repositories

Current State

NO_GO is a positive result when the proof is incomplete.

In the financial proof-action lab, the system is currently holding NO_GO_PAPER_EXECUTION_BLOCKED. That is not advertised as profit, alpha, or live readiness. It is an audit result: the system has not proved that it is allowed to act.

The current public summary is deliberately narrow: money denominator is zero, ActionWarrant pass count is zero, authority leak is zero, credit leak is zero, and the system reports 18 receipt-closure gaps plus 122 active wait repair items.

Proof-Carrying Action

Every high-risk action should carry a chain: thesis, falsifier, warrant, receipt, regret, scar, and clean learning status.

No-Credit Repair

Repair work does not earn reward, denominator credit, action permission, or clean-learning credit until the missing evidence is closed.

Negative Space Memory

Not acting is still a decision. The system records why it waited, what evidence was missing, and what later regret can be computed.

Authority leak guard visual showing a blocked no-go state.

Authority Boundary

The agent must not reward itself for sounding convincing.

The central failure mode is reward hijack: the system becomes attached to a neat report, a high confidence score, a paper-only result, or an apparent repair. The proof-carrying layer blocks that route by separating suggestion, repair intent, metric credit, execution authority, and clean learning.

This is why the public message is not "AI trading bot." The stronger claim is: high-risk agents need an auditable action-qualification layer before they are allowed to affect the world.

Public Interfaces

Small runnable slices, clear boundaries.

LayerPublic entryWhat it provesBoundary

Action OSProof-Carrying ActionReference schemas for warrants, receipts, repair, and regretPublic slice only BenchmarkWisdomBenchLearning-from-failure scoring and sample evaluation surfaceRunnable public benchmark EvidenceEvidence mapClaim boundaries, evidence levels, and replay disciplineNo private logs SystemsSystems mapHow the public layers relate to the product architectureCommercial core protected

Release Rule

Open the discipline. Protect the engine.

The public project should let builders test the idea without exposing customer workflows, proprietary ledgers, private product logic, or unreleased operational systems. That boundary is part of the research.

The next community goal is practical: make enough public artifacts that skeptics can run examples, submit counterexamples, and understand the action boundary, while keeping the deeper system reserved for product deployment and future evidence releases.