AI policy
What Ask the ledger does with a model, what the model may not do, and how to report an answer that is wrong.
What Ask does
Ask is retrieval first. A question is matched against chunks of text that are each tagged with a dataset and a source in the registry. The model composes an answer from those chunks only. It does not answer from its training data.
Every pound in an answer must appear in a retrieved chunk. If the model writes a figure that is not in a chunk, the answer is discarded and a deterministic drafter writes the answer from the chunks instead.
Every citation is a chunk. Each chunk names its dataset, its source, the published date and the retrieved date. An answer with no citation is not shown.
Each sentence carries its statement kind: fact, calculation, official statement, party statement, poll result, news report, academic finding, analysis or inference. When the evidence for a question is missing, the answer says not ingested. It does not guess.
Models
| Job | Model |
|---|---|
| Composing the answer | Workers AI kimi-k2.6 |
| Embedding chunks for retrieval | bge-base-en-v1.5 |
| AI Search index | bge-m3 |
Limitations
The model can misread a chunk, pick the wrong period, or describe a figure loosely. The pound check catches an invented number. It does not catch a wrong sentence about a real number. Read the citations, not only the answer.
Retrieval finds what is ingested. A series that is not in the registry cannot be found, and the answer will say so. A missing series is not zero.
No model is used to pick sources, to score claims, or to write the pages on this site other than Ask.
Reporting an error
Copy the question, the answer and the citations, and open an issue at inferrex-master/uktax-lol. A confirmed error is logged as a dated note under Corrections.