Control Anchor / Reviewer demo
Keep the evidence connected.
AI-assisted work needs a history people can examine. Explore our public demonstration of evidence capture, attributable review, and linked corrections.
Concept illustration.
Public source · Offline component demo · Synthetic inputs · No model calls
The full product is under development in UVLM’s private repositories. This public repository provides a focused reviewer demo.
Start with what you can inspect
A record you can follow.
What went in? What was produced? What changed after review? The demo makes those relationships visible in a small, reproducible workflow.
Capture the context
Keep supplied source material, prompts, settings, and candidate output together. Inspect the deterministic checks applied to the evidence.
Follow the review
Trace a scripted review and a linked correction. The original candidate remains preserved alongside the separately recorded child revision.
Reopen the record
Explore exported JSON, JSONL, CSV, and local HTML views. Examine package verification and the rejection of a deliberately altered copy.
Recorded component validation
The repository reports 8 focused tests passed and 4 directory/ZIP verification checks on Windows with Python 3.12.14. These are bounded results from the documented run, not a certification or a new test performed by this webpage.
The wider product vision
Local judgment.
Shared possibility.
Our goal is to make AI-assisted work easier to examine, challenge, and improve while keeping human decisions and local permissions explicit.
- Architecture in development
Triadic Brain
The broader architecture for cooperating AI workers, grounding, advisory evaluation, and evidence-linked continuity.
- Workbench in development
Control Anchor
The intended workbench for locally defined internal controls, reviewable evidence, and attributable human decisions.
- Proposed network
Sophia P2P
A peer-to-peer network intended to share authorized compute and provenance between independently controlled local instances.
From family groups to institutions
The proposed design spans small family groups, community organizations, government departments, and global organizations. Local rules would govern what may be shared, with whom, and for what purpose.
Plain language, professional depth
We aim to support intuitively worded controls for non-specialists and structured mappings to institutional frameworks such as COSO ERM. Those mappings would support review; they would not confer compliance or professional assurance.
Choose your next step
Open the work.
Ask better questions.
For curious reviewers
Start with the README, recorded results, and limitations. You can inspect the project without running code.
For engineers
Follow the Windows/Python 3.12 quickstart. The demo requires no API key or dependency installation; filesystem requirements are documented.
For collaborators
Explore the integration plan, research questions, and proposed University Student Intern Opportunity. Participation and partnerships remain to be arranged.
Is this the complete Control Anchor product?
No. The full product is being developed in UVLM’s private repositories. This public repository exposes a focused, offline reviewer demo and supporting documentation so others can inspect one part of the design.
Does the demo call an AI model?
No. Its inputs and candidate output are authored fixtures, its review decisions are scripted, and its repair advisory is supplied. It does not execute the separate CoherenceLattice, Sophia, or Atlas cognitive components.
What does verification establish?
It checks bounded properties of retained evidence. It does not establish factual truth, independent custody, cryptographic signing, control effectiveness, or regulatory compliance. The deliberately altered-copy check does not establish detection of every possible alteration. See the documented verification limits.
Can I reuse the public code and materials?
Consult the repository’s license scope, license, and notices before reuse. Public access does not grant a blanket license to every document, image, or media asset.
Help shape the next step
Build with care.
Review with rigor.
We welcome engineering review, research conversations, and funding discussions grounded in work people can inspect. Help us turn this demonstrated component into a useful, human-governed product.
Email: info@ultraverbaluxmentis.org. Please do not include credentials, sensitive records, or private evidence in an initial inquiry.