How the review works
The method belongs under the hood. What you get is a readable result.
You do not need to know the names of 18 procedures to use Opponentura. This page is for those who want to check how the service makes different roles look for different kinds of error, and what happens to the document before it is issued.
Not “seven opinions” — different review duties
Strengthen the decision
The strongest version of your position is stated first: attacking a weakened version proves nothing.
Find what it rests on
Assumptions, facts, causal links.
Test the alternatives
Other explanations of the same data, other moves, the other side’s position.
Model the failure
The worst case, rare severe events, the barriers in the way.
Check the evidence
What is known, what is taken on faith, what a source confirms.
Turn the conclusion into action
The first test, the revision conditions, dated action items.
18 procedures — different ways to be wrong
“O-18” is not one scientifically validated super-method, and we do not present it as one. It is a catalogue of analytical procedures drawn from decision analysis, intelligence analysis, risk management and argument logic.
4 procedures are mandatory and apply to all participants. The other 14 are handed out one each — to the role they fit best, and never twice. That is why critics look at the decision with different instruments instead of paraphrasing one another. A procedure counts as applied only if it produced a result or the reason it did not fit is stated.
The passport of all 18 procedures →
What these procedures are
All 18 are below. What we do not publish is the rest of the machinery: how lenses are distributed across roles, the wording of the tasks given to critics, and the order of synthesis — that is precisely the difference between a hearing and a chat prompt.
| Procedure | What it does | Who gets it |
|---|---|---|
| Steelman | first state the target position more strongly than the customer did; attacking a weak version does not count | every critic |
| Pre-mortem (Klein) | a year on, the decision has failed — write the story of that failure from the future, not a list of what might go wrong | every critic |
| Key assumptions | write out EVERY load-bearing assumption, including the unspoken ones, and what follows if each is false | every critic |
| Falsification (Popper) | state plainly the observation that would refute the position, and the cheapest way to obtain it | one of the critics |
| Toulmin warrant attack | break the argument into data → warrant → claim; attack the warrant, not the data and not the conclusion | one of the critics |
| Inversion (Munger) | invert the conclusion: what must be true for the decision to be right; test every such condition | one of the critics |
| Quality of information | walk the facts: source, age, interest, verifiability; flag what is unverified | one of the critics |
| What If? (the surprise has already happened) | take what the target calls impossible as a fact and explain how it came about | one of the critics |
| High impact, low probability | name 2–4 low-probability, devastating events and assess readiness for each | one of the critics |
| Good faith of the source | if the data comes from an interested party — what exactly it gains by showing this and hiding that | one of the critics |
| Competing hypotheses (ACH, Heuer) | list the alternative explanations and run the evidence through a matrix: which hypothesis survives the most attempts to refute it | one of the critics |
| Four points of view (stakeholders) | for each key player: their goal, their version of events, what they will do in response to this decision | one of the critics |
| Failure mode map (FMEA) | for a plan or process: failure modes × probability × severity × detectability; what covers the costliest ones | one of the critics |
| Bow-tie (ISO 31010) | on the principal risk: causes → event → consequences, preventive and mitigating barriers, and the gaps between them | one of the critics |
| Base rate and the outside view (Tetlock) | for any estimate or forecast: what is the base rate in a comparable class of cases, and why this case is not the exception | one of the critics |
| Scan for biases in the position | confirmation, anchoring, sunk cost, survivorship, motivated reasoning — each with the quote where it shows | one of the critics |
| Chain-of-Verification | put verification questions to your own key claims and answer them separately; on divergence, drop the claim | every critic |
| Code and engineering | edge inputs, races, boundaries, dependency failures; what breaks under load and in a partial outage | one of the critics |
One procedure is one critic’s work: the fewer critics, the fewer procedures physically fit into a hearing. A panel of 3 covers 16 of 18; a panel of 7 covers all 18. Whatever went to nobody we see in our own coverage ledger; the report states which procedures were applied and with what result.
How to read the names
Plain name first, term in brackets. You should not have to expand an acronym to understand a function.
- Strongest version of the decision (steelman)
- Failure scenario (pre-mortem)
- Key assumptions check
- Analysis of competing hypotheses (ACH)
- Causes → event → consequences (bow-tie)
- Re-checking key claims (chain-of-verification)
Why critics work separately first
If everyone sees the first answer, the first answer becomes the frame of the discussion. So where the procedure calls for it, positions are formed separately and only then compared.
Using different model vendors reduces dependence on one system, but it does not turn AI roles into independent human experts. Those are different things and we do not blur them.
Disagreement is not averaged out
After the separate passes, the real points of collision are worked through: where positions agreed, where they diverged, which fact could settle the argument, and who changed position under which argument. The goal is not a tidy consensus but knowing what survived the objection.
Even the analysis is reviewed
Facts
External claims are re-checked and get a status: confirmed, corrected or not confirmed.
Proportionality
A separate pass asks whether the critical machinery has turned ordinary business risk into a reason to panic.
Coherence
Before sealing, the document is edited for contradictions and errors.
If a fact is not confirmed, we say so. Not every risk is a reason to stop a decision.
Where a finding came from
In the full report the origin of an argument is marked: your own words, a document, a public source, a calculation, an assumption or the panel’s inference. The mark shows the type of grounding; it does not guarantee that the conclusion is true.
How it shows up in the report
Right after the verdict comes “The essentials” — the three results a hearing owes you regardless of plan: the three most dangerous assumptions your decision rests on (and what backs each), the cheapest check to run before the irreversible step, and the condition under which the decision stops being right. If any of the three is missing, that is a defect: the service detects it itself by checking the issued text, not the intention of a prompt.
Where the debate is part of the plan, its transcript is followed by a map of disagreement: where the panel agrees, where it splits, which fact would settle the dispute, and who changed position under an argument. A few lines that explain what the argument was for.
Beyond that, every report states the panel and why each member was invited, and the lens coverage: which procedures were applied and with what result. Before printing, the document passes two checks — a fact review by another vendor's engine and an editorial pass. A fact taken from the web enters the report only with a link to the primary source; “not confirmed” is better than a pretty invention.
If the client arrived with a situation and had no decision yet, the reports gain a “THE REPORT” section: every draft decision tabled, marking the one accepted and the ones the client rejected. The rejected drafts stay in the document on purpose — six months later, “why didn't we take the second route” is answered by the report rather than by memory.
Where this method is weaker than we would like it to be has a page of its own: where Opponentura gets it wrong.
Decision-bias review
On the plans that include it, a separate pass looks for signs of anchoring, confirmation-seeking, sunk cost and similar effects in how the decision is framed. The wording is honest — “the text shows signs of”, not “we measured your biases”. The standardized instruments that would actually measure them are not part of this service.
What the method cannot do
- it does not know the future;
- it depends on what you described and attached;
- models can share the same misconceptions;
- public-source search is incomplete;
- a professional conclusion requires a professional;
- an elaborate procedure does not make error impossible.
Where this comes from
The procedures are not our invention. The works they are drawn from, with links where the address has been verified, are listed in the passport.