Real reviews
Real reports, not marketing reconstructions
These are the documents the customers received. The file was not rebuilt or polished for the website, so its hash still matches our registry — download it and check for yourself on the verification page. No names, company names or identifiers appear in them: that is how the service works, not the result of our editing.
These three reports are in Russian, and so is the one you would receive today: the panel deliberates in Russian and the sealed PDF is Russian. The English panel opens after our first ten paid hearings. Translating a sealed document is not a shortcut around that — a translation is a different file, and it fails our own verification. Below is what each case was about; the original is one click away.
Your report will not end up here
The two customer cases are published with their consent: the founder asked each of them personally, and the permission is recorded in our registry — without that record a report is never published, however good it looks. Only the owner of the service ever asks, and only once the case is closed; the bot never asks. The promise of confidentiality outweighs the value of a nice sample.
The documents themselves carry no identifiers by design: the service never asks for a company name, registration details or personal names, so the panel never receives them. We checked all three files — not a single company name, address, phone number, email or tax ID. You cannot tell whose company it is.
What happens to your materials and how quickly they are
deleted — in the privacy policy; you
choose the retention mode yourself with /privacy in the bot.
About the role format in reports issued before 19 August 2026
Those reports gave AI critics synthetic professional profiles such as “former operating director” or “15 years in due diligence”. Those were model role instructions, not biographies of real people and not a claim that human experts participated. They were not limited to the very first reviews: the panel builder itself worked that way, and we found it in our own audit on 19 August.
Since 19 August a panel member is a mandate — Mandate “Name” — described by what it tests and why it belongs in this case. A mandate has no lived experience, no years of practice and no past employers: it is a testing assignment, not a person.
We do not silently edit the old PDFs: changing a word changes the file hash. That is why the reports below keep the old format — explained here rather than rewritten there.
Should we launch Opponentura without demand validation?
Verdict: REVISE · panel: 7 critics · recess and a second round · substantive objections: 18 · orders: 8 · 12.08.2026
The founder brought his own service to the panel — the one you are reading about. The question was not «is the product good» but «is the demand proven»: the need was found in himself and his own circle, the first hearings are given to acquaintances for free, and not one stranger has paid. Seven critics attacked the design of the test, not the product, and returned REVISE: the hypothesis is stated so that it cannot end in a negative result.
First test. Take five founders in the target profile, do not show or describe the product, and ask one question: which decision in the last 12 months would have cost more than 300,000 roubles if wrong, and what did you do to check yourself? Zero cost, ninety minutes — it measures behavior that already happened and cannot be improved by polishing the product.
Outcome: not known yet. We will update this card when the customer reports the result.
Replacing a procurement director with an AI agent
Verdict: REVISE · panel: 3 critics · substantive objections: 3 · orders: 6 · 10.08.2026
An importer automated order calculation in his accounting system and wants to replace the head of procurement — salary 250k ₽ — with a junior at 100–120k within a month or two. The panel examined not the saving but whether it is proven: what is actually left for the human after automation, and whether the timeline fits the dismissal procedure the labour code requires.
First test. Audit the last 20–30 real purchases: what the system produced, what the manager then did, which deviations appeared in timing, price, batch, documents, payment and supplier, and whether those actions could have been carried out from the written procedure by someone else.
Outcome: not known yet. We will update this card when the customer reports the result.
Shrinking a restaurant chain — and which sites to close first
Verdict: REVISE · panel: 7 critics · tournament of 4 drafts · recess and a second round · substantive objections: 16 · orders: 8 · 10.08.2026
The owner of ten restaurants in Moscow loses money every month, owes wages and faces a rent payment in early September. No decision existed yet: close some sites, hand them to other operators, relaunch the chain in a cheaper format, or shut down entirely. The same panel went through every option in turn.
First test. A bank statement across all accounts and the cash balance on one date. Fifteen minutes and zero cost, and the result decides whether the owner has two weeks or four months — that is, which of the drafts is even on the table.
Outcome: not known yet. We will update this card when the customer reports the result.
Verdict, panel size, counts and dates come from our hearing log, not from prose written by hand.
What “original” means
A printout, screenshot or copied passage may be useful for reading, but file verification concerns the issued PDF. A matching hash proves that the file is unchanged. It does not prove that the analysis is correct.
Why the flaws stay visible
If an early report contains a formatting defect, a superseded method rule or an error we later identified, the original file remains unchanged and the correction is documented next to it. An immutable record should not become a polished retrospective.
Where Opponentura gets it wrong →
Want your own decision checked the same way?
Start a review — first verdict free First check whether you need one