LLM governance: versioned behavior protocols built from human corrections
Turns human reviewers' corrections into approved, versioned rules that produce the system prompt, the evaluator prompt, the regression tests and an exportable package that can be verified.
When a company puts a language model to work, its reviewers' corrections end up in loose spreadsheets or in prompt changes nobody documents. This LLM governance software turns those corrections into behavior rules that a person approves, groups them into versioned protocols and generates from them the system prompt, the evaluator prompt and the regression tests. The tests run against your company's LLM, and every version can be exported as a package you can verify.
Who it is for
- AI teams that maintain assistants or agents on an LLM and need to explain how they behave.
- Compliance and risk management teams, who need to trace the path from a human correction to the rule and the version in use.
- Consultancies that tune models for their clients and want to deliver a documented protocol with its tests.
What LLM governance software solves
- Scattered corrections. Human evaluations are grouped into locked batches, and candidate rules come out of them with a reproducible method.
- Reviewers who contradict each other. The system detects when the same thing has been corrected in opposite directions and suggests reviewing it by hand, splitting the rule or discarding it.
- Changes with no control. Each protocol has numbered versions, explicit approval and comparison between versions.
- Regressions. Each version has its own tests, which run against the model.
What you can do with it
Use cases and evidence
- Register use cases with their goal and base model, and move them from draft to active.
- Link each case to the conversations, documents and human evaluations that justify it.
- Lock batches of evaluations so they work as a fixed, reproducible input.
Rules from corrections
- Group similar corrections and calculate each rule's confidence from the amount of evidence behind it and how consistent it is, with adjustable thresholds.
- Tell proposed rules from weak rules, and see contradictions with a suggestion for resolving them.
- Always get the same result from the same batch and the same parameters. AI plays no part in this step.
Human review
- Approve, reject, mark as weak or retire each rule.
- Trace each rule back to its evidence and forward to the versions that include it.
Protocols and versions
- Create protocols with numbered versions, approve them and compare two versions line by line.
- Generate the system prompt and the evaluator prompt from the approved rules, with a token estimate.
- Generate and regenerate the regression tests for each version.
Tests and export
- Run the tests against the configured LLM and see each response as passed, warning or failed, with a recommendation (accept, regenerate or hand to a person) and its history.
- Export each version as a package organized by content level, with instructions and a digital fingerprint (hash) to check it has not been altered.
Administration
- Use a dashboard, a chat with the configured model and a document manager.
- Browse the audit log with filters and manage users and roles.
Benefits
- Your reviewers' corrections stop getting lost and become rules with an owner and a date.
- Before you publish a change, you find out whether the model has stopped doing something that used to work.
- You have documentation ready for an internal audit or for answering a client.
Industry context
General industry data with its public source. These are not results of this product.
- In 2025, 19.95% of EU enterprises with 10 or more employees used AI technologies, 6.47 percentage points more than in 2024. Source: Eurostat, "Use of artificial intelligence in enterprises", 2025 data.
What you get with each license
Price upon request, depending on license type and scope. Maintenance and support are contracted separately for every license type.
- The software installed on your server or in your cloud, under a license to use.
- The connection to your LLM, database setup and initial configuration.
- Initial training for administrators.
- The complete source code, to build it into your AI platform or adapt it.
- Installation and configuration documentation.
- A technical handover session with your team.
- A warranty against defects for the agreed period.
- The right to embed the asset in a commercial product of your own, identified in the contract.
- The right to distribute it to your customers as part of that product, without asking for authorization for each installation and with no limit on end customers, unless otherwise agreed.
- It does not include reselling or sublicensing it on its own.
- It is not exclusive, unless otherwise agreed.
- Maintenance, support and adaptations are contracted separately.
Technical specifications
Language and runtime: TypeScript 5.9 on Node.js. Minimum supported version: Node.js 20 (the one the file storage client requires in its configuration).
Frameworks: Next.js 14.2 and React 18.3, with Prisma 5.22 for data access and Tailwind CSS 3.4.
Database: PostgreSQL.
Operating systems: A server with Node.js.
Interface languages: Spanish.
Required infrastructure:
- PostgreSQL.
- File storage through Supabase Storage (in the Supabase cloud or in a self-hosted installation).
- An LLM with an API compatible with OpenAI Chat Completions, in the cloud (requires your own account) or running on your own infrastructure.
Private deployment: Can run fully on its own if you use a self-hosted Supabase installation for documents and a model running on your infrastructure with the Chat Completions interface. If you use cloud services, it needs a connection to them.
Main libraries:
| Library | Version | License |
|---|---|---|
| next | 14.2.35 | MIT |
| react | 18.3.1 | MIT |
| @prisma/client | 5.22.0 | Apache-2.0 |
| @supabase/supabase-js | 2.109.0 | MIT |
| jszip | 3.10.1 | MIT OR GPL-3.0-or-later |
| jsonwebtoken | 9.0.3 | MIT |
| bcryptjs | 2.4.3 | MIT |
| zod | 3.25.76 | MIT |
Dependencies with specific license terms:
- jszip 3.10: MIT OR GPL-3.0-or-later.
Compatibility and integrations: OpenAI Chat Completions API (with streaming responses), compatible with OpenAI, Azure OpenAI and your own models that expose that interface, and Supabase Storage.
Redistribution terms
The asset's own code can be embedded in the licensee's commercial products under the source code and OEM licenses, on the terms of the contract. With the installed license, the software is used and not redistributed. The only dependency with specific license terms is jszip (MIT OR GPL-3.0-or-later); the details are in the Technical specifications.
Installation requirements
Installed on your server or in your cloud, with a PostgreSQL database and file storage compatible with Supabase Storage, in the cloud or self-hosted, for documents. It needs an LLM that offers OpenAI's Chat Completions interface, either a cloud service or a model running on your own infrastructure. Model usage costs are on you.
Integrations
Any service compatible with OpenAI's Chat Completions interface, such as OpenAI, Azure OpenAI or your own models that expose it.
Frequently asked questions
Does it use AI to decide which rules get created?
No. Candidate rules come from a fixed calculation on the corrections, and they only enter a version once a person approves them. The LLM is used in the chat and when the tests run.
How are reviewers' corrections loaded?
They come in as input data from the evaluation records you already have, and the way they are loaded is agreed according to their format. The software does not include a screen for reviewers to correct responses.
How are responses evaluated in the tests?
They are compared with the expected response by their text, not by their meaning. It is a basic regression check: doubtful responses are flagged for a person to review.
How does it help with the EU AI Act?
It provides documentation and traceability: which correction gave rise to each rule, who approved it, which version is in use and which tests it has passed. It does not replace any conformity assessment that may apply under the EU AI Act (Regulation (EU) 2024/1689).
Can one installation serve several clients?
It is designed for one installation per company. To serve several clients from the same installation, data separation has to be strengthened during adaptation.
Is maintenance included?
No. Maintenance (updates and bug fixes) and support (questions, training and adaptations) are contracted separately.
Request
Request information on P-IA-01
Fill in the form with the asset reference and the license type you're considering. I reply in writing. If you need to review code, tick the NDA box and we'll sign one before anything is shared.