Policies
How we use AI
First pass, plainly stated. Everything Strata Scope draws or writes, including anything an AI model read, is a first-pass suggestion for a competent engineer to verify. It is never a design value.
Strata Scope uses an AI model for one kind of job: reading documents that exist only as a PDF, a scan or a photograph, and turning what they say into rows an engineer can check. It does not use AI to interpret the ground, choose parameters, write report text or make decisions about people.
1Where AI is used
A log, a laboratory PDF or a core photograph is read only when you start the read yourself: uploading a core photograph stores it, and its board is read only when you press Read. The file goes from our server to the model; your browser never talks to the model, and the key that opens it never leaves our server.
| Task | What is sent | What comes back |
|---|---|---|
| Reading a borehole log | The log as you supplied it: a PDF, a scan or a photograph | The hole's header, strata descriptions and depths, and tests, in a fixed shape, each value marked as read |
| Reading laboratory results | The text of the laboratory's PDF, a few pages at a time. A scanned PDF with no text is refused rather than guessed at. | Test results in the same rows an AGS file gives, each marked as read from the certificate |
| Reading a core box board | One core photograph, reduced to 1,600 pixels across | Only what the board states: hole, box number, depth from and to. Anything not legible comes back blank, never guessed. |
2Where it is not used
- AGS files are read by ordinary code and checked against the AGS rules. No AI is involved.
- Correlations and parameters come from a fixed library of published correlations, each with its source, never from a model's recall.
- Report boilerplate, such as the factual-report statement and the limitations, is reproduced word for word. A model never writes it.
- Derived figures in the factual report are arithmetic on your data.
- The desk study is published survey data, shown with its source.
- Decisions about people. Nothing in Strata Scope uses AI to assess, score or decide anything about a person.
3The model and its provider
The reads use Claude Sonnet 5, a model made by Anthropic, PBC, a company in the United States, through Anthropic's commercial API. Anthropic acts as our processor under its Data Processing Addendum. It is listed with every other recipient on the sub-processor list.
4What is never done with your data
- No training. We do not train any AI model on your content. Anthropic's Commercial Terms state that it may not train models on customer content from its services.
- Short retention at the provider. Anthropic's published policy is to delete API inputs and outputs within 30 days, keeping them longer only where needed to enforce its usage policy or where the law requires (Anthropic privacy centre).
- No sale, no advertising. Your content is not sold or used to market anything.
- No reads behind your back. Dropping a log or a laboratory PDF on the intake screen, or uploading a core photograph, does not send it anywhere; it waits until you start the read, and the button that starts it says it is a billed call.
5How the engineer stays in control
- Every value a model read is marked as read, and drawn apart from values taken exactly from an AGS file, so the two cannot be confused.
- The model's answer is checked against a fixed shape on our server. If it does not fit, you see the model's raw answer and the reason, rather than a quiet gap.
- A read whose coordinates disagree with each other is held at a checkpoint until you have seen each proposed change and chosen.
- A later read adds to a job; it never replaces what the job already holds, and a hole from an AGS file keeps its exact position.
- A factual report cannot be issued without a named author, checker and approver.
6Limitations
An AI read can be wrong, and wrong with confidence. In particular it can:
- misread handwriting, faint scans, stamps, or photographs taken at an angle;
- misread a digit, a decimal point or a unit in a depth or a result;
- miss a row, or join two rows, in a dense table;
- copy a description faithfully that was itself wrong on the log.
That is why every read value must be checked against the source document by a competent engineer before it is used.
7The rules that apply
In the UK there is no law written for AI alone; data protection law applies, and we follow the Information Commissioner's guidance on AI and data protection, including on being open about how AI is used. Our privacy notice covers the personal data side.
The EU's AI Act sets transparency duties from 2 August 2026, mainly for systems that talk with the public or generate synthetic media. Strata Scope does neither: it transcribes documents for an engineer to check. [TO CONFIRM: solicitor, how the EU AI Act applies if the service is offered in the EU]
8Raising a concern
If a read got something wrong in a way the checks did not show, or you have a concern about how AI is used here, email mark@MTE-hq.co.uk with the job, the file and what you saw. We will acknowledge it, look into it, and tell you what we found and changed.