Methodology
How observation from outside actually works.
Six steps, then the full register of what the method cannot see. The register is not an appendix: it is transcribed from the same source that prints it in every report we produce.
The method
Six steps.
01
Observe from outside
Prediction markets that settle on a public ledger leave a record of every matched trade: which accounts traded, in which market, at what price, at what moment. We read that record continuously. No integration with the platform is required, and no permission, which is what makes the observation independent. It also means we start from less than the operator has, not more.
02
Link accounts into participants
A single participant can hold many accounts. We build a funding and timing graph from the same public record and group accounts that move as one. This is the step everything else rests on: a pattern that looks unremarkable across nine accounts is often obvious once those nine are read as one participant.
03
Run the indicators
Twelve indicators scan the linked record on a schedule, each one answering a narrow question that a licence condition requires an operator to monitor. Detection is deliberately rule-based rather than learned: a rule can be read, argued with, and defended in a meeting, which a model's judgement cannot.
04
Record evidence, anonymised
Every detection carries the evidence that produced it: what was observed, over what window, at what scale. Account addresses never appear. They are replaced with stable anonymised references, so a reader can follow one participant across a whole report without ever being handed an identity we have no right to publish.
05
Map to an obligation
A detection that is not tied to a duty is trivia. Each indicator is mapped to the licence obligation it speaks to, across anti-money- laundering, market integrity, and consumer protection. We describe what was detected and which duty it engages. We do not assert that any breach occurred: that judgement belongs to the operator's compliance and legal function, not to an instrument.
06
Reconcile against declarations
This is the step that makes the rest worth having. Public data cannot distinguish authorised activity from the same behaviour left undeclared. Set the observed record against what the operator says is theirs, and the ambiguity resolves: either the behaviour matches the mandate, and the record becomes evidence of a working control, or something is doing the same thing without a mandate, and that is the finding.
Coverage
What each indicator can see.
Eight of the twelve are supported directly by the public settlement record. Three need a manual step or a second source. One is screening only, and we say so on its own page rather than reporting an absence of findings as a clean result.
- 01Wash tradingObservableThe public settlement record supports this directly.
- 02Linked walletsObservableThe public settlement record supports this directly.
- 06StructuringObservableThe public settlement record supports this directly.
- 07Fund flow and layeringPartly observableDetected, but a manual check or a second source is needed.
- 12New-wallet positionObservableThe public settlement record supports this directly.
- 03Pre-event spikePartly observableDetected, but a manual check or a second source is needed.
- 04Coordinated positioningObservableThe public settlement record supports this directly.
- 08SpoofingScreening onlyThe evidence largely does not reach the public record.
- 09Cross-platform divergencePartly observableDetected, but a manual check or a second source is needed.
- 11Resolution-window riskObservableThe public settlement record supports this directly.
- 05Position concentrationObservableThe public settlement record supports this directly.
- 10Markers of harmObservableThe public settlement record supports this directly.
Three of the twelve are qualified and one is screening only. We publish this because the alternative is an operator told their platform is clean by a system that could not have seen the problem.
Specimen figures. Not live data, and not drawn from any platform.
Output
What a period of observation returns.
Detections do not arrive evenly. One participant tripping several indicators at once produces a cluster, which is why the console and the reports group by participant rather than listing chronologically by rule.
- Critical
- High
- Medium
- Low
A specimen 30 day period. The cluster around days 17 to 21 is the shape a single participant's activity makes: many indicators tripping at once, not a steady trickle across all of them.
Specimen figures. Not live data, and not drawn from any platform.
Limits register
What the method cannot see.
These seven caveats print in every integrity report we produce. They are reproduced here in full, in the same wording, because a method described one way in a sales page and another way in the artefact would not be worth trusting in either place.
01
We see completed trades, not the order book. Orders that were placed and cancelled without trading are invisible to us on this platform.
Order-flow visibility: this platform's matching, pricing and market making are operated on exchange infrastructure whose order book is not public; the chain records matched fills only. Order-book indicators (spoofing and layering) therefore have limited or no visibility here, and quoted volumes reflect executed trades, not order intent.
02
The busiest group of accounts may simply be the platform's market maker. Public data alone cannot tell us; only checking against the operator's declared market-making accounts can.
Market-making attestation: sustained two-sided volume by linked wallets is consistent with authorised market-making. Public data cannot distinguish an authorised mandate from manipulative trading; classification requires reconciliation against the operator's declared market-maker accounts.
03
We have only been watching a platform for as long as we have been watching it, so 'new account' means new to us. An account may be older than our observation of it.
Observation depth: account age and behavioural baselines are measured from each account's first observed trade. Where observation of the platform is younger than a detector's look-back requirement, that detector under-reports and is stated as such rather than estimated.
04
When we say an account holds most of a market, we mean most of the money we saw put at risk during the period, not the market's full history.
Exposure measurement: position concentration is measured as net buy exposure observed in the review period, a conservative proxy for open interest, which requires position-level data to compute exactly.
05
Prices for the two sides of a yes or no market can share one record. We filter the false crashes that produces, at a small cost in edge cases.
Outcome resolution of prices: where trade records are keyed by market rather than by outcome token, complementary-outcome price artefacts are filtered, at the cost of masking a genuine move that lands exactly on the complement of the prior price.
06
We do not yet automatically check price moves against the news. A person still has to ask what happened at that moment.
Public-information cross-check: correlation of price moves with public news events is currently a manual review step, not automated.
07
Patterns that look like problem gambling can also be trading software. We say which looks likely, and never diagnose.
Behavioural signatures (markers of harm) have benign explanations, notably automated strategies; flags request behavioural review and do not assert harm.
Read the indicators.
Each of the twelve has its own page setting out the question it answers, what the record supports, and what it does not.