What is Stratl?

Stratl is the system of record for consequential AI decisions. This page explains every word of that sentence, then shows what Stratl actually produces.

The problem, without jargon

A company lets software decide things that matter to people: whether a refund is issued, whether an insurance claim is paid, whether a loan is approved, whether an account is closed. More and more of that software is an AI agent: a program that reads a situation, calls a language model such as Claude or GPT, looks things up, sometimes asks a person for permission, and then acts.

Months later, someone asks what happened. A customer disputes a decision. An auditor samples a hundred cases. A regulator asks how an automated system treated a category of people. A court asks for everything about one person.

Today the answer is scattered. The model provider has a log. The tracing tool has a trace. The ticketing system has a note. None of them agree on what “the decision” was, none of them say who was allowed to act, all of them can be edited by whoever runs them, and none of them can be checked by an outsider. That is not evidence. It is a story the company tells about itself.

What Stratl does

Stratl does four things, always in this order.

1. Ingest. It collects what the AI did from the places that already see it: a few lines of code in the agent, an OpenTelemetry trace, an export from a model provider or a tracing tool. Nothing has to be rebuilt.
2. Normalise. It turns that into one record per decision, filed under the person or thing the decision was about (the subject: a claim, an applicant, an order). The record says what the AI saw, what it called, which policy applied, who approved, what action ran, and how it ended. It never contains the prompts or documents themselves, only their fingerprints, so the record cannot leak what it describes.
3. Custody. Each record is signed with a key that nobody, not even Stratl, can copy; linked to the record before it so nothing can be removed quietly; and, every hour, sealed into a checkpoint that an independent timestamp authority dates. The records live in storage the customer controls, where the Stratl service has no permission to delete.
4. Serve. From those records Stratl produces what people ask for: an evidence bundle for one decision, a subject pack with everything about one person, a period pack against a named regulation, and a notice to the affected person when a decision went against them.

The last piece is the one to remember: anyone can verify a record with a free command-line tool, offline, with no Stratl account. If Stratl disappeared tomorrow, every bundle already exported would still prove what it proves.

A decision, end to end

A support agent at an online store receives a ticket about a damaged parcel. Here is what Stratl records, in the order it happened. The words in bold are the event types you will see everywhere in these docs.

Step	What happened	Event in the record
1	The ticket text arrived	input (a fingerprint of the text)
2	The agent asked a model what to do	model_call (provider, model, fingerprints of prompt and answer)
3	The agent looked up the order	tool_call (tool name, fingerprints of the request and the result)
4	The refund policy said amounts over 100 need a person	policy_eval (policy id, version, result)
5	A support lead approved it	approval (who, with what authority, what they decided)
6	The refund was issued	action (name, target, when, result)
7	The outcome for the customer	outcome (completed, not adverse, category: refund)

The record is signed and chained in under a second. If the customer later says “I was never refunded”, the bundle shows the refund was issued at 14:32:02 by an agent after a named support lead approved it, and a verifier confirms nothing in that account has been changed since.

For people who run the business

You do not need to understand hashes or signatures to use Stratl. You need to know that a Stratl record is something you can hand to a regulator, and that its proof does not rest on anyone’s word, including ours. Continue with the one-page brief.

For compliance and legal

The record answers the nine questions a supervisory authority asks about an automated decision. See the decision record, explained and which frameworks the packs cover.

For engineers

The record format is an open specification with a JSON Schema and cross-language conformance fixtures. Start with local setup, then the record format walkthrough.

What Stratl is not

Stratl is deliberately narrow, and the boundaries are part of the design.

• Not a gateway or a firewall. Stratl does not sit between your agent and the model, and it cannot block or slow a request. It records.
• Not an approvals engine. People approve things in your own systems. Stratl records that they did, and with what authority.
• Not an observability dashboard. It does not show latency, cost or token graphs. Your tracing tool does that better.
• Not a security detector. It does not look for prompt injection or data exfiltration.
• Not a promise of compliance. Stratl’s packs say which record fields help demonstrate which control. They never say “you are compliant”. Nobody honest can say that from records alone.

Words we do not use

You will not find “explainable AI”, “reasoning”, “compliance guaranteed”, “immutable forever” or “impossible to tamper” anywhere in Stratl. Records capture observable facts, not a model’s private reasoning. Storage keeps every version and refuses deletes; that is a strong control, not a law of physics. The docs say what the system does and what a verifier can check, and stop there.

Where things live

Thing	Where	Who controls it
The signed records, checkpoints and any original content	Object storage in the custody region the workspace chose, or the customer’s own bucket	The custodian: Stratl-managed, or the customer
The signing key	A key service that signs but never releases the private key (AWS KMS, Google Cloud KMS or Azure Key Vault)	Same
A searchable index of fingerprints and metadata	Stratl’s PostgreSQL database	Stratl
The public keys	Inside every bundle, and at a public URL that needs no account	Everyone

Next: who each part of these docs is for, or jump straight to the concepts.