By Heath Muchena · Sources reviewed 2026-11-13 · Original DN analysis and explicitly illustrative examples
What Matters
Participation, AI use, prevalence, causation and collusion are different claims. Evidence for one should not silently become evidence for the others.
- Experiments do not count live-market incidents.
- A majority claim needs a population and attribution method.
- Known counterevidence survives missing fields.
The DN thesis
The most useful claim is the strongest one the actual evidence supports, with its boundary still attached. DN’s map helps reviewers identify the next missing question without producing a certainty score or an accusation.
1. Write the claim before collecting impressive charts
“Bots moved the market” can mean that automated accounts traded, that they dominated volume, that their actions helped cause a price move or that operators coordinated unlawfully. Those are different claims. A chart that is relevant to one may say little about the others.
Specify the outcome, market, interval and actors. A claim about one trading pair during a brief disruption is narrower than a claim about an entire exchange or all of decentralised finance. Keep the original scope visible when a headline is shortened or a result is shared.
The DN instrument makes the claim type the first substantive choice. Its evidence requirements change accordingly: a prevalence claim needs a population boundary, a causal claim needs an analysis of contribution and a collusion claim needs more than a shared direction of travel. This is a map of questions, not a scoring system that makes weak inputs add up to certainty.
For a concrete review, write the strongest statement the evidence currently supports, then record what would be needed for the next stronger statement. “Automated activity was observed” can be useful without being inflated into “AI caused the crash.”
2. Separate automation from AI attribution
Fast execution, repeated order shapes and activity outside office hours can motivate an investigation. None is a complete explanation of the software or the people behind it. An ordinary execution algorithm, a scripted strategy, a market maker and an AI-assisted operator can leave superficially similar traces.
Keep the evidence of automation separate from the evidence of AI use. Ask whether the source inspected the system, relied on operator disclosures, inferred a behavioral pattern or reused a third party’s labels. Each route has limitations that should travel with the claim.
A useful historical warning comes from the SEC’s report on algorithmic trading, which describes the 2010 Flash Crash through interacting order flow, liquidity conditions and automated trading behavior. [2] That account is not evidence that contemporary generative-AI agents caused the event. Replacing “algorithmic” with “AI” changes the claim.
DN’s crypto-hack attribution ledger separates actor attribution from an asserted AI role. This article applies the same discipline to market outcomes without reconstructing an exploit or identifying a person from a wallet label.
3. Inspect the denominator behind a majority claim
A claim that AI accounts generate most activity needs a numerator and a denominator measured in compatible units. Orders, trades, quote notional, base volume, unique accounts and messages are not interchangeable. One account can submit many orders, while one participant can control many accounts.
Define the measured venue, products, interval and data coverage. Explain how activity was classified and what remained unclassified. If only a subset of accounts has labels, the labelled subset is not automatically representative of the whole market.
Also ask what the denominator excludes. An exchange-wide statement cannot be established from one visible pool without a justified aggregation method. Gross turnover should not quietly become net economic exposure. A burst of cancellations should not silently become executed volume.
The local instrument does not manufacture a percentage from a transaction count. It records whether the population and attribution method are defined, and preserves a declared missing denominator as an inference gap. That is more honest than ranking venues by an unverified “AI share.”
4. Keep experiments inside their evidence boundary
A simulation can reveal a mechanism that deserves attention. It can also make strong assumptions about objectives, information, agent populations, constraints and the market environment. A result under those assumptions is not a count of real-world incidents.
The CMA’s 2026 discussion distinguishes theory, experiments and simulations from a smaller set of empirical studies, and describes an evolving field with open questions. [1] This supports careful investigation, not a fabricated list of confirmed AI collusion episodes.
Record what the experiment actually varied and what it held fixed. If the simulated agents had identical objectives or shared information that real participants do not possess, explain why the mechanism might or might not transfer. A convincing demonstration of possibility does not estimate prevalence.
The instrument therefore preserves simulation-only status even when every reference field is filled. It does not penalize experiments for being experiments. It prevents the evidence category from being silently changed when someone writes the market claim.
5. Ask what would have happened without the alleged cause
A price movement that follows a bot’s activity is not automatically a movement caused by it. Both could respond to the same news, liquidity change or trading opportunity. The direction may also run the other way: the bot may be responding to the movement.
A causal review needs an account of the mechanism and a credible comparison. Depending on the study, that might involve a design using a change in access, a documented intervention, a carefully constructed comparison group or another defensible source of variation. Naming a method is not enough; the identifying assumptions and limitations matter.
Ask whether the claimed effect is a contribution, an amplification, a trigger or the sole cause. A multi-factor event can include an automated feedback mechanism without being explained entirely by it. The wording should follow the actual estimand and evidence, not the most dramatic possible description.
The worksheet distinguishes “association only” from a reported causal-design reference. The latter is still not a validated causal finding. DN does not inspect the data, reproduce a study or calculate an effect size through this local tool.
6. Do not turn common behavior into a legal conclusion
Several systems can respond similarly because they observe the same market. They might also share a service, exchange information, implement an agreement or interact in ways that soften competition. Those possibilities require different evidence and may raise different legal questions.
The CMA distinguishes explicit agreement implemented with software, information exchange through a hub and other forms of algorithmic coordination. Its discussion also explains that the applicable UK legal analysis depends on the conduct and legal route. [1] It should not be converted into a universal rule that matching trades prove an unlawful agreement.
Keep the operational observation separate from the legal conclusion. Record what communications, shared rules, data flows, contractual arrangements or other evidence are actually available. Do not use a cluster of wallet addresses as a substitute for that analysis.
The tool’s collusion branch flags common behavior alone as a gap. Even when additional evidence is reported, the result remains evidence for review. It does not identify an offender, find manipulation or decide which jurisdiction’s competition law applies.
7. Preserve corrections and counterevidence in the final record
A useful review includes evidence that could weaken the preferred explanation. A missing record is not evidence that an event did not happen, but a contradictory record should not disappear because other fields remain unknown.
The CMA’s Ticketmaster investigation illustrates the value of precise correction: it did not find evidence of the assumed real-time algorithmic pricing model and described staged releases at different prices instead. [3] The example concerns ticket sales, not cryptocurrency trading; its lesson is about testing the mechanism behind a public claim.
Save the claim’s wording, the source versions, the unresolved alternatives and the reviewer responsible for the next question. If later evidence changes the conclusion, retain the prior version and explain the change. Do not rewrite an old uncertain claim as though the new result had always been known.
The exported DN record preserves the current selections and references. It does not store market data, identify accounts or validate a source. For performance accounting, use the separate live-results reconciliation guide. A profitable bot, a busy bot and a bot that caused a market move are three different assertions.
DN Market Claim Evidence Map
Choose one claim and classify the evidence you actually have. The map identifies claim-specific gaps; a complete record is not proof. Inactive fields are retained in the download but do not drive the selected claim.
Use non-sensitive aliases, never credentials, personal data or confidential documents. Blank and unknown entries stay unresolved. The tool does not open evidence references.
No current record. Load a labelled example or enter your own information.
This instrument makes no network requests and uses no persistent browser storage. Its download contains the current inputs, calculation or review findings, source references and provenance. The website’s other services remain subject to its privacy policy.
Illustrative cases
These cases are invented to expose the method’s boundaries. They are not observed customers, deployments, trading outcomes or legal findings.
Illustrative simulation promoted to a market claim
The result remains a simulation-only evidence boundary. A complete study record does not establish that the outcome occurred in a live market.
Illustrative automation with unknown AI use
The automation declaration is retained while the AI attribution required by this claim remains unknown.
Illustrative majority claim without a denominator
The population gap remains visible. Transaction activity alone cannot supply a market-wide majority percentage.
Illustrative common behavior called collusion
Similar behavior does not settle the coordination mechanism or legal classification. The tool records the inference gap.
Illustrative scoped official-case analysis
Source, scope, attribution and design references are reported. The output is a documented claim only, with causation and legal violation still unestablished by the tool.
Illustrative counterevidence with other unknowns
The known counterevidence survives the missing AI and causal-design fields. Unknowns are not used to hide a reported contradiction.
Methodology, assumptions and limits
The selected claim activates different evidence requirements. All claims retain automation evidence and market/source scope. AI-majority and AI-causation claims additionally require AI attribution; a majority claim needs a defined population, a causal claim needs an analysis basis, and a collusion claim needs a coordination-evidence basis.
Simulation-only evidence, association-only causality, an undefined majority denominator and common behavior presented as collusion are explicit inference gaps. A reported contradiction remains visible even when other fields are unknown. Reported designs and official-case material are not automatically accepted as true or final.
The instrument reads only the reader’s classifications and aliases. It does not open sources, calculate prevalence or effect sizes, authenticate identities or apply competition law. Edits invalidate the previous result and JSON download. Current raw inputs, inactive fields, source references and example provenance are exported.
Frequently asked questions
Does a rapid transaction pattern prove AI involvement?
No. Automation and AI use are separate attribution questions. This tool records the evidence declaration and its reference; it does not classify accounts from behavior.
Can a simulation prove that bots colluded in a real market?
A simulation can demonstrate an outcome under its assumptions. It does not by itself establish that the outcome occurred in the claimed market or that conduct was unlawful.
What does a complete evidence map mean?
Only that the declared claim has the required references for this review structure. It is not authentication, a validated causal analysis or a legal finding.
Why retain inactive fields?
Changing the claim should not delete the reader’s earlier entries. The export preserves them while identifying which fields did not drive the current result.
Primary sources
Reviewed 2026-11-13. A later export timestamp does not update the sources.
- CMA: AI and collusion—frontiers, opportunities and challenges — Distinguishes mechanisms and evidence settings; jurisdiction-specific legal discussion, not proof of crypto-market incidents.
- SEC: Report to Congress on Algorithmic Trading — Algorithmic trading and market-stress mechanisms; not an attribution to contemporary generative AI.
- CMA: Why clear and accurate pricing matters — Ticketmaster mechanism example illustrates evidence correction; not a trading-market analogy with proven identical effects.
Turn an evidence gap into a focused question
If a market claim depends on an unclear source, denominator or causal design, explore a focused DN Research Desk enquiry. Research scope is agreed separately; no attribution, enforcement or investment outcome is promised.
Disclosure: this article links to DN’s own research service. A commercial enquiry does not purchase a favorable conclusion. This instrument is educational documentation support and does not provide personalized financial, legal or security approval.
