LZLZL/Prediction markets/Research method
FREEBUILD IT F · MethodFlagship

Adversarial verification
how to kill your own idea

2026-08-21 · default to "overturned unless it survives"

You have just had an idea. The next step is not to verify it but to try to destroy it. The difference is not attitude, it is the default: verification assumes "sound unless a problem appears", adversarial verification assumes "unsound unless it survives an attack".

1Why "looking for evidence" is broken

You are doingWhat follows
looking for supporting evidenceyou will find some — with enough data there always is
finding it, believing moreconfidence grows independently of truth
meeting a counter-example, explaining itan explanation always exists too

The problem: "supporting evidence can be found" is true of sound and unsound ideas alike, so it carries no information. Exactly one thing distinguishes them: you attacked it seriously and it did not die.

2Separate the roles

RuleNote
Assume it is wrongthe burden is on the proposer, not the reviewer
"Cannot be shown" counts as overturnedyou need not prove it false, only that it was not shown true
The reviewer does not fix itraising the problem is the whole job
Attack from separate anglescorrectness, cost, execution, basis — each sees different faults

The second row is the core and the hardest to accept: presumed guilty is correct here, because when money is involved an unproven idea and a wrong idea have identical consequences.

3Measured: a pre-launch review

One real-money line went through an adversarial review before launch: 13 defects found, all fixed before it ran. Two were fatal:

DefectConsequence
the cancel method called did not exist in the SDK, and the error was swallowed by a broad exception handler the wallet's only cancellation defence would never run, silently
the processed-set lived only in memory a restart would re-order the same window; a crash loop compounds it

Both share a property: invisible in normal operation. The first surfaces only when you actually need to cancel; the second only on restart — and those are precisely the moments you least want a surprise.

Three general checks these produced

① Any process with automatic restart: for every piece of state ask "is it still there after a restart?" In-memory state resets and the process continues as though nothing happened.

② A broad exception handler turns a hard error like "method does not exist" into silent failure. A defence that never raises is identical to no defence — except you believe you have one.

③ A capital floor's "unknown" must not be wrong in either direction. Unknown treated as zero halts falsely; unknown treated as fine runs blind. In-flight funds must be modelled explicitly.

4It applies to conclusions, not just code

One of my own trading proposals was overturned by adversarial verification and archived without a cent being spent.

The questions used are plain and hard to talk past:

QuestionBlocks
What is the denominator?quoting a subset rate as a whole-sample rate
Is that causal link in the source, or did you join it?presenting a plausible guess as a record
How many subsets did you cut before this one?picking the largest number in the noise
What is this approximation's bias direction?treating an optimistic bound as the value
Does it survive a different basis?a conclusion that came from the ruler, not the data

The last one caught me hardest: a beautiful pattern that "survived every cut" turned out to be entirely a label artefact. Anything that "holds under every condition" should first be suspected of living in the instrument rather than in the data — full account in the broken ruler.

5Doing it alone

ApproachWhy it works
Sleep on ittime partially simulates "a different person"
Write "how it will die" firstsee pre-claimed failure modes
Explain it to someone unfamiliarthe parts you cannot explain are the parts you have not thought through
Force yourself to write three objectionsfailing to reach three means you have not seriously considered being wrong
Cost first, upside seconda great many proposals die at the cost calculation

6One line to keep

"Supporting evidence can be found" is true of sound and unsound ideas alike, so it carries no information. The only thing that does: you attacked it seriously and it survived. Where money is concerned, an unproven idea costs the same as a wrong one — so the default has to be "unsound".

EvidenceCheck it yourself

Pre-launch review an adversarial review of one real-money line found 13 defects, all fixed before launch; two were fatal (a non-existent cancel method swallowed by a broad exception handler, and an in-memory processed-set causing duplicate orders on restart).
A basis discrepancy the deployment document's defect table lists 8 while another record says 13, with no reconciliation between them. This page uses 13 and notes the discrepancy.
A proposal overturned one of my own trading proposals was archived after adversarial review, unfunded.
The ruler artefact a pattern surviving every cut was ultimately judged a label artefact — see the broken ruler.
Checked 2026-08.

NextWhere to go

F · METHOD
Pre-claimed failure modes
F · METHOD
Null models: the most basic adversary
F · METHOD
Patterns that survive every cut are suspicious
F · METHOD
Replays must state their bias
This is an educational and research record. It is not investment advice, promises no returns, and offers no personalised trading recommendations. Rules and API behaviour are per the official documentation; this page states when it was checked and both can change without notice. Prediction markets are restricted or unavailable in some jurisdictions — confirm your own before taking part.