Systems Layer

From intelligent answers to proof-carrying action.

SOVEREIGN is the product layer behind the research program: a local-first digital twin and decision cockpit for users who need AI to finish real work, respect boundaries, preserve evidence, and learn without self-deception.

observe warrant act learn cleanly

Product Map Technical Layer Proof-Carrying Action Evidence Map

Evidence contract ledger visual for proof-carrying action systems.

Product Position

One core system, several application fronts.

The core product is not a chatbot and not a single trading or robotics module. It is a proof-carrying cognitive action infrastructure: a system that turns goals, observations, uncertainty, constraints, evidence, failed attempts, and user trust into auditable work orders and bounded actions.

Finance and robotics are separate product fronts because they stress different parts of the same substrate. Finance tests proof, restraint, regret, and execution evidence. Robotics tests embodiment, recovery, perception, and action under physical uncertainty. The personal digital twin is the user-facing cockpit that binds them into a usable service for high-knowledge and high-net-worth users.

Core

Digital Twin Cockpit

A private operating layer for plans, decisions, memory, workflows, evidence gates, task closure, and value ledgers.

Finance

Proof Action Lab

Trading is treated as a cost-bearing testbed for whether an AI system has earned the right to act.

Robotics

Embodied Evidence

Robot learning papers stress perception, recovery, provenance, physical uncertainty, and when not to act.

Public

Research Archive

The website, Zenodo, GitHub, and Hugging Face expose reproducible slices without exposing private systems.

Proof-Carrying Action

The first positive result can be qualified restraint.

The latest finance evidence pack does not claim profit. It shows a stricter result: the system blocks itself when execution proof, denominator evidence, and regret computability are not closed.

AxisCurrent StateMeaningProduct Use

GateNO_GO_PAPER_EXECUTION_BLOCKEDQualified restraintPrevents premature action

Denom.money_denominator_n = 0No money evidence admittedBlocks fake edge claims

WarrantActionWarrant pass = 0No action earned authoritySeparates advice from action

Leaksauthority leak = 0; credit leak = 0Repair work did not become rewardProtects clean learning

Queue122 repair items; 18 receipt gapsThe system knows why it cannot actTurns no-go into work orders

Operating Loop

Every action needs a proof trail.

The operating chain is:

goal -> observation -> relation field -> thesis -> falsifier -> warrant -> receipt -> regret -> scar -> clean learning

A suggestion without an execution certificate is not an action. A repair queue item is not progress credit. A pretty report is not reward. A no-action decision is still evidence if it has a wait-policy receipt and a counterfactual baseline.

Open-source boundary map for public and private layers.

Release Boundary

Open enough to verify. Private enough to survive.

Public releases should include papers, evidence maps, claim boundaries, schema, small runnable demos, and reproduction notes. Private layers should keep customer workflows, trading runtime details, credentials, unpublished execution traces, and product-specific decision logic out of public repositories.

Public: research pages, DOI links, non-sensitive schemas, toy demos.
Delayed public: non-public artifacts only after public-boundary clearance.
Private: execution authority, customer memory, proprietary workflows, and live system logs.

Claim Boundary

We are not training AI to sound more confident. We are building systems that know when they are not yet allowed to act.

This boundary is not weakness. It is the difference between a persuasive model and a responsible action system.