From question to figure,with proof at every step.

The eight steps

1. You ask, in your own words. “How much did we sell last month in each shop?”
2. Burebi recognises the question on the company's definitions. A word the company has not defined is not guessed: Burebi abstains and flags it, so that the definition is given once, by a person.
3. Code composes the query and computes the figure. The AI model picks identifiers from the company's catalogue and fills in validated fields; it does not write the query.
4. You get the figure with its provenance. “Where does the figure come from?” shows the definition, the query and the scope: period, source, filters.
5. You get its limits. Until when the source has data, what stays outside the definition, what was assumed.
6. Every change leaves a trace. A row in the audit log: what, when and who. The rows are chained (hash): each depends on the one before, so none can be changed in secret.
7. A rule changes through a single door. Hypothesis, demonstration on today's data, a check made by code (the judge is not an AI model) and the click of a named person.
8. What was changed can be withdrawn. With a version, a reason and who did it; restoring returns to the same hash.

When there is no definition yet: the hypothesis

1. The hypothesis. The catalogue abstains; the model searches for the answer with a query it writes and checks. The answer says it does not go through definitions.
2. The demonstration. The proposed definition is applied on a copy of the catalogue, validated and run on data; the result is compared with the verified value.
3. The definition. Only then does it enter the catalogue, as “measured”. A person confirms it, with their name.

This is how the company's catalogue grows: from real questions, checked at every step.

Why a small model

Burebi runs on Claude Haiku 5.5, about 40 times cheaper than Opus 5.5: $0.10 / $0.50 against $4 / $20 per million tokens (input / output), for requests up to 100,000 tokens; above that, Haiku 5.5 costs $0.50 / $2.50 (claude.com/pricing, checked on 17 October 2026). We do not claim a small model thinks like a large one: the heavy thinking is paid for once and crystallised in structure, meaning a deterministic semantic engine, a knowledge graph, recipes and truth guards. The model chooses, fills in and words.

What stays local, what goes

Stay with you: the database, the knowledge graph, the company's definitions, the log and the secrets. No customer data is stored on our servers, in today's production setup (full diagram).

Reach the AI model (Claude, by Anthropic), through Claude Code or the API: the question, the necessary context and the result already computed by code, so it can be worded. The model carries out the operations, but it does not receive the database and does not guess what the data means: the meaning sits in the definitions. What Anthropic does with what it receives is set by its own terms, which differ for a subscription and for the API.

Different companies run in separate instances: each has its own data, knowledge graph, secrets and assistant memory.

What we measure about ourselves

The test suite and the application used live are two different measurements, so we state both.

• In live use, 12 of 500 real turns failed (2.4%), measured on 12 October 2026.
• A turn takes 26 seconds at the median; nine in ten finish within 99 seconds (7 October 2026).
• On every project: how many answers with figures came only from definitions, how many times it abstained, what they cost.

Correctness against an answer adjudicated by a person is not yet measured. When it is, we will publish it here.