For teams running credit, hiring, claims or other automated decisions, and AI agents
When someone asks why the system decided, show what it did, unchanged.
An applicant refused, a candidate screened out, a claim declined, an agent that took an action nobody expected. The question that follows is always the same: what exactly happened, which version of the system did it, and can you show the log was not edited once the question was asked?
- When it is tested
- A person contests an automated decision about them; a supervisor or market-surveillance authority asks for your logs; an internal review after an incident; a court.
- What you will be asked for
- For high-risk AI systems, the events logged automatically (AI Act, Article 12), kept by the provider (Article 19) and by the deployer for at least six months (Article 26(6)). For a decision based solely on automated processing, the person’s right to human intervention and to contest it (GDPR, Article 22(3)), which you can only honour if you can say what was decided and why.
- What goes wrong today
- Logs rotate, are overwritten, or sit in the same infrastructure the team can edit. Model versions move on. Months later nobody can say which version decided, and an honest log is no easier to believe than an edited one.
- What changes
- One sealed record per decision or agent run: the inputs’ fingerprint, the model version, the outcome and any human review, kept in your own systems with only a salted fingerprint sent to us. Drop-in recorders for your existing logs, LangChain and the OpenAI Agents SDK do it without changing your code’s logic. AI agents and assistants
Sealing shows that a log existed unchanged from a given moment. Whether your system is high-risk, and whether its logs are enough, is yours to establish. What the AI Act asks · GDPR Article 22
If you build an AI-governance or deployment platform
Your customers already keep their inventory, assessments and logs in your platform. What they cannot show from inside it is that none of it was rewritten after a decision was contested, because your platform, like any system, can be edited by whoever runs it.
- What you add
- One call when a log segment, an assessment or a model card is finalised: its salted fingerprint goes to the evidence log, the content stays in your platform. Each customer can then prove, to a supervisor or a court, that the record is the one from the day.
- Why it is not a competitor
- We hold no models, no logs and no assessments, and we build no governance features. One open format, free verifiers, no lock-in. Our boundaries
- How to start
- A free sandbox key with no form, the recorders for Python and Node, or the hosted MCP endpoint. A first integrating platform talks directly to the person who builds the API.
If you train or fine-tune models: seal each checkpoint without showing it
The question after an incident is often not what the model decided but which model it was: the exact weights, and the evaluation they passed before release. Weights are the last thing a lab will hand over, and an internal registry can be rewritten by whoever runs it.
- What you seal
- When a checkpoint is saved or released, a small record naming it: the SHA-256 of each weights file, the training configuration’s fingerprint, and the evaluation report’s fingerprint with its headline scores, if you choose to include them. The record stays with you; only its salted fingerprint is sent.
- What you can show later
- That the weights you hand an auditor, a customer or a court are byte for byte the ones you sealed on that date, with that evaluation, without having shown them to anyone before, us included. Anyone can check it offline with the free verifiers.
- How
- The recorders or the release-sealing GitHub Action in a training pipeline, one entry per checkpoint; or /seal on a folder of files by hand. Large files are fingerprinted where they are; nothing is uploaded.
- And the data it learned from
- The
ai.training-datarecord seals each training, validation or test data set by its manifest, with your records of where the data came from and how it was prepared (AI Act, Article 10(2)) and, for a general-purpose model, of your copyright check and published training summary (Article 53(1)(c) and (d)). The data never leaves you.
A seal proves which files existed and when, not that the evaluation was sound or that a model is safe.
Try it in two minutes
- Choose a decision log export or an agent run's output of your own. It is read in your browser and never uploaded.
- Seal it on Seal a file. A free sandbox key is issued on the page; only a salted fingerprint reaches the log.
- Make a copy, change one character in it, and check both against the receipt at Check. The copy fails; the original passes.
In production, the recorders seal every decision as your system writes its log.