What Burebi practises from Ontopsis
Ontopsis is the discipline behind Burebi (see the foundation, in Romanian). The link to the term in the lexicon is our interpretation; the fact next to it is on record, with its date.
Evidence
Every claim, with its date and basis.
On this site we write nothing about Burebi that is not in the table below. A fact older than 30 days is rechecked before it is used, and a claim that becomes false is corrected within one working day.
| Claim | Basis | Date |
|---|---|---|
| A board of seven automated roles works on the product. Each has one measurable truth in hand, and a proposal must quote at least 12 characters from its own measurement or from an existing file. | decided | 18 Oct 2026 |
| Copyleft (18 Oct 2026): none of the product's 878 dependencies is under GPL or AGPL without a permissive alternative; a few are under MPL-2.0 (weak, file-level copyleft); PDFs are printed through an external Chromium. | measured | 18 Oct 2026 |
| The product serves a single customer today. Work on the product is done separately from the customer's business. | decided | 18 Oct 2026 |
| Distribution to a new machine is not solved: PDF export needs Chromium (to be shipped or downloaded on first run), and that condition is still open. | decided | 18 Oct 2026 |
| The own modules for a company without an ERP (projects, scheduled processes, documents, reports) do not include accounting, stock or invoicing. | decided | 18 Oct 2026 |
| The pilot lasts 4 weeks, on one company workstation installed by us together with the company's IT; the company provides a contact person (about 2 hours a week) and a read-only ERP user; the success criteria are written on day one; at the end the company decides: continue, change or stop. | decided | 18 Oct 2026 |
| No licence fee is charged during the pilot; the AI model is paid by consumption, and the cost of each turn is shown on the project. | decided | 18 Oct 2026 |
| A person answers contact messages within 2 working days. | decided | 18 Oct 2026 |
| A live presentation of the application takes 30 minutes and uses the data of our own marketing post, not a customer's. | decided | 18 Oct 2026 |
| Web server logs are kept for 14 days, and traffic reports contain no IP addresses. | decided | 18 Oct 2026 |
| The contract that follows the pilot is individual; as a guide, about one national minimum wage per month (Republic of Moldova: 6,300 lei from 1 January 2026, per gov.md; Romania: 4,325 lei gross from 1 July 2026, HG 146/2026, per press reports). | decided | 18 Oct 2026 |
| Burebi runs on Claude Haiku 5.5: $0.10 / $0.50 per million tokens (input / output) for requests up to 100,000 tokens, against $4 / $20 for Opus 5.5, about 40 times less. Above 100,000 tokens, Haiku 5.5 costs $0.50 / $2.50 (public prices on claude.com/pricing, checked on 17 Oct 2026). | measured | 17 Oct 2026 |
| All of a company's rules sit in one view, “The company's rules”: on the production instance, 445 rules in 34 families (17 Oct 2026). When a concept is defined in several places, the drift is shown. | measured | 17 Oct 2026 |
| The same application serves different companies in separate instances: each has its own data, knowledge graph, secrets, interface storage and assistant memory; apart from the code, nothing is shared. | measured | 17 Oct 2026 |
| No customer data is stored on our servers: the database, the knowledge graph, the log and the secrets stay on the customer's machine. In the production setup (17 Oct 2026), the gateway to our server is switched off and answers pass only through the AI model (Claude, Anthropic). | measured | 17 Oct 2026 |
| For an answer, the AI model receives the question, the necessary context and the result already computed by code, in order to word it; the company's database and definitions are not copied to it. | measured | 17 Oct 2026 |
| Where the catalogue has no definition yet, the model formulates a hypothesis: it writes the query and checks it, and the answer says it does not go through definitions. The hypothesis becomes a company definition only after code validates it, runs the formula on data and compares it with the verified value; a learned definition stays “measured” until a person confirms it. | measured | 17 Oct 2026 |
| Burebi's price is set in the contract, individually. As a guide: about one national minimum wage, for an assistant that works alongside the top manager and for a continuous process of building up the company's capability. | decided | 17 Oct 2026 |
| On 17 Oct 2026, the founder decided that the site puts Burebi first, with Ontopsis as its foundation. Until 16 Oct 2026, the site put the discipline ahead of the product. | decided | 17 Oct 2026 |
| Under an answer with figures from the catalogue, Burebi states what the figure leaves out: until when the source has data, what is declared outside the definition, what was assumed. “Where does the figure come from?” and the report sheet carry the same list. | measured | 16 Oct 2026 |
| Every command that changes data leaves a row in the audit log, chained by hash; the contract requires 0 commands without a trace, and a test checks it. | measured | 16 Oct 2026 |
| Every action taken at a workstation is signed with the name of the person working there; a generic role (“System”) is not accepted as a name. | measured | 16 Oct 2026 |
| The company's rules and knowledge are not deleted, they are withdrawn: with a version, a reason and who did it, and restoring returns to the same hash. Personal data can be deleted, leaving a trace without content. | measured | 16 Oct 2026 |
| A company threshold changes through a single door: hypothesis, demonstration on today's data, a code-based check and the click of a named person, with withdrawal possible. Irreversibility ceilings change only through code. | measured | 16 Oct 2026 |
| On every project, Burebi shows how many answers with figures came only from definitions, how many times it abstained instead of guessing, what the figures leave out and what the turns cost. | measured | 16 Oct 2026 |
| In a test of 12 turns (16 Oct 2026), the AI model cost 0.0118 $ in total. The figure comes from a test project, not an estimate for a company. | measured | 16 Oct 2026 |
| Burebi is a product for business processes, project management and analysis of a company's data; it is not an application builder. | decided | 14 Oct 2026 |
| Burebi does not claim that a small model thinks like a large one: the heavy thinking is paid for once and kept in the structure of the product (a deterministic semantic engine, a knowledge graph, recipes, truth guards). The model picks identifiers from the company's catalogue, fills in validated fields and words the answer. | decided | 14 Oct 2026 |
| Autonomy grows on measured evidence, scored by code (not by an AI model), and can be withdrawn: Manual → Shadow → Approval → Automatic. Sending anything outside the company and writing to the ERP stay permanently at the “Approval” level. | decided | 14 Oct 2026 |
| The Hub is the console for a company's workstations: unified audit, cost, issues, signed rule packages, backup. By default only metadata leaves a station, and each station works fully without it. It runs locally today; putting it on the company's network is not decided. | decided | 14 Oct 2026 |
| The database and the knowledge graph stay local, on the user's machine. For an answer, the question and the necessary context reach the AI model (Claude, by Anthropic), through Claude Code or the API. The phrase “the data never leaves” would be false in the absolute, which is why we say exactly what stays and what goes. | decided | 12 Oct 2026 |
| Portability (12 Oct 2026): 0.33% of the product's engine touches the customer's installation; the percentage must not grow. | measured | 12 Oct 2026 |
| Quality in live use (12 Oct 2026): 12 of 500 real turns failed (2.4%). A green test suite and the live application are two different measurements, and the public figure states both. A turn is a question and its answer. | measured | 12 Oct 2026 |
| Licences (12 Oct 2026): one rejected licence, 17 security advisories from dependencies (RUSTSEC) and no licence file of our own. The licence of the product is the founder's decision. | measured | 12 Oct 2026 |
| In the customer's ERP, Burebi's actions are add-only, with confirmation on every run and a dedicated user with explicit rights. | decided | 12 Oct 2026 |
| For a company without an ERP, Burebi has its own modules: projects (roadmap, tasks, risks and decisions, status report), scheduled processes, documents with versions and visual reports from data. Mail and teamwork are not yet connected to real servers. | decided | 12 Oct 2026 |
| We have not measured competitors: the role that was meant to do it is suspended, and declared so. Any comparison with other products first requires web research with cited sources, labelled as such. | decided | 8 Oct 2026 |
| Readiness for a new customer (7 Oct 2026): 3 of 3 pillars of the method are specific to Oracle databases; a second, end-to-end data engine is needed before we promise portability. | measured | 7 Oct 2026 |
| Usability (7 Oct 2026): median turn 26 s, p90 99 s (nine in ten turns finish within 99 s); 89 of 496 turns took over a minute; error messages are written for the founder, not for a stranger. | measured | 7 Oct 2026 |
| A business question goes through: statement → request with identifiers from the company's catalogue → SQL query composed by code → rows with provenance → verified answer. The figure is computed by code, on the company's definition; the AI model recognises the question and words the answer. | measured | 7 Oct 2026 |
40 facts. “Measured” = a measurement of the product; “decided” = a decision of the founder, on record.
| Ontopsis term | What Burebi does | Measured on |
|---|---|---|
| Judge with no stake | Autonomy grows on measured evidence, scored by code and not by an AI model, and stays reversible. | 14 Oct 2026 |
| Gradual application and irreversibility threshold | Autonomy has four levels: Manual, Shadow, Approval, Automatic. Sending anything out and writing to the company's ERP stay permanently at the Approval level. | 14 Oct 2026 |
| Demonstration, not only validation | The test suite and the application used live are reported as two different measurements. In live use, 12 of 500 real turns failed (2.4%). | 12 Oct 2026 |
| Residue, declared | What is not measured is said: the comparison with other products was not measured, and the role that should have done it is declared suspended. | 8 Oct 2026 |
| Honest disclosure | It says exactly what stays local and what goes. The database and the knowledge graph stay on the user's machine. For an answer, the question and the necessary context reach the AI model (Claude, by Anthropic). | 12 Oct 2026 |
| Constitutive commitments, explicit | The heavy thinking is paid for once and crystallised in structure: a deterministic semantic engine, a knowledge graph, recipes, truth guards. The AI model picks identifiers, fills in validated fields and words the answer. | 14 Oct 2026 |
| Proposal tied to evidence | A board of seven automated roles works on the product. A proposal must quote at least 12 characters from its own measurement or from a file that exists. | 28 Sep 2026 |
What it does not prove
A single case. Burebi serves a single customer today. A discipline is not founded on one case, and Burebi is not presented as its proof.
Generality is not shown. For a new customer, 3 of 3 pillars of the method are specific to Oracle databases (7 Oct 2026). Until it runs end to end on a second engine, the practice is not shown to be general, not even inside the product.
The foundation came afterwards. The principle “life precedes the model” was formulated afterwards; it is not written into Burebi's architecture. Reversibility in Burebi was born of operational prudence. Ontopsis gives it a foundation, and rebuilding it on that foundation is work not yet done.
Comparisons are missing. Competitors are not measured, so we do not claim that other products do not practise the same.
Why Burebi is now in front
Until 16 October 2026, this site put the discipline ahead of the product, so that Ontopsis would not become Burebi's packaging. On 17 October, the founder decided to put Burebi first (the decision is on record in the table above), arguing that a current of thought needs a viable product to live. Ontopsis remains the discipline behind it and is not reduced to it: it will be credible when it is practised outside it too.