Limitations
Where Opponentura can get it wrong
We sell a review of decisions, not infallibility. If you are about to trust this service with an important decision, this page is worth reading before you pay, not after a mistake.
Separate passes, different roles and multiple model providers reduce some failure modes of a single AI conversation. They do not create independent human expertise or make agreement automatically true.
See real reports Back to plans
The short map of limitations
Seven lines; each is taken apart below.
- Poor input, poor review.
- Different models can share the same error.
- Search does not see everything.
- A system built to find problems can overstate them.
- Signs of bias in a text are neither a diagnosis nor a measurement.
- A report does not replace a professional opinion.
- A verdict is a snapshot, on one date and on the data available.
The review is no stronger than what you told us
We can only review well a decision we managed to understand. If a key document is missing, a number is approximate or a counterparty is left unnamed, that limitation must be visible in the report itself rather than implied.
The panel works with your account of the matter and your documents. A fact you did not mention does not exist for the hearing: stay silent about a personal guarantee and it will not appear in the report. Volume is capped too — up to 80 A4 pages on the top plan; beyond that we say what we did not accept instead of pretending to have read it. So the cheapest upgrade in quality is not a pricier plan but a full answer at intake.
Models can be wrong too
We separate critics by role and by vendor: part of the panel runs on Anthropic engines, part on OpenAI, and each side checks the other's facts. That lowers the risk of a shared mistake without removing it — the models trained on overlapping data and share some of the same misconceptions. Hence our wording: independent passes, different roles and models from different vendors — not “independent expertise”.
Web search does not guarantee completeness
Before printing, the report passes a fact review by another vendor's engine with live search. It has three outcomes: confirmed, corrected, not confirmed. The third is a legitimate result, not a defect: a flag beats a handsome link to a document that does not exist. Fresh news, a closed registry or a paper that never reached the web stay outside the review.
We are likelier to overstate a risk than to miss it
This reduces severity inflation; it does not prove the decision is safe. The separate pass answers three questions: what genuinely changes the decision, what is ordinary business risk, and what the panel did not find.
A panel of critics is built to look for threats, and that has a price: caution is easily mistaken for quality. Worse, we have a vested interest here — the refund guarantee depends on the number of severe findings. So “critical” or “high” is assigned only when all four conditions hold: a concrete damage scenario, a stated basis (a fact, a document, a calculation or your own words), a stated effect on the choice, and a verifiable next step. Miss one and the severity drops. The share of severe findings in every report is counted by the service itself and reported to the owner when it exceeds a threshold. And a separate pass does re-check whether the panel overstated things: after the synthesis the report is read by a proportionality reviewer running on another vendor's engine. It names what genuinely changes the decision, what is an ordinary cost of doing business, and lowers the severity of findings that fail the four conditions. It may lower, never raise, and it cannot touch the verdict — that is the panel's work. In the document its output is the “Proportionality” section, right after “The essentials”.
You can see what a finding rests on
A finding carries an origin tag, set by its weakest link. Honestly about where we are today: not every finding is tagged. The service measures the tagged share against a threshold rather than promising full coverage — and in our first reports external facts more often end in “not confirmed” than in a link. We would rather say so here than leave you expecting footnotes. The tags: [customer's words] — as you told us, unverified; [document] — from a file you attached; [public source] — from a public source with a link; [calculation] — arithmetic on your own numbers; [assumption] — a critic's supposition; [panel's inference] — a conclusion drawn from the above. The tag reflects the weakest link. It is inconvenient for us and useful for you: without it a model's guess and a fact backed by a document look equally weighty on the page.
We do not measure your mind
The decision-bias reviewer writes “your words show signs of such and such a bias — here is the quote”. He does not write “we measured your level of bias”: measurement requires standardized instruments, which we do not administer mid-hearing. This is an assessment of a text, not a diagnosis of a person.
The report does not replace a professional
A hearing is not legal, medical, tax or investment advice and does not replace a qualified opinion. The opposite is true: the document is built so that you can walk into a lawyer's office with ready questions instead of a vague “please take a look”. Some matters — medical, narrowly industrial, forensic — require a live specialist, and we say so inside the report.
In regulated or expert domains the useful output is often not a substitute for the professional but the questions, documents and assumptions worth taking to them.
What we do about these limitations
None of them goes away. Some can be made visible and checkable — that is what the pipeline is for:
- mandatory intake: five questions are always asked before the review;
- every finding carries the origin of its grounding — your words, a document, a public source, a calculation, an assumption or the panel’s inference;
- a separate fact check on another vendor’s engine;
- a separate proportionality review of the criticism itself;
- if the hearing ran with a reduced panel, the report says so;
- after delivery we come back and ask about the outcome;
- the issued file is immutable and verifiable by its hash.
This does not remove the limitations. It makes some of them visible and checkable.
A verdict is a snapshot dated to the hearing
The world keeps moving after the seal goes on. That is why the report states the conditions for revision, and why follow-up comes after: reminders about the action items and a repeat hearing with the same panel if circumstances changed. We will never quietly rewrite old reports — a new document is issued and the old one stays as it was.