Test a fraud pattern hypothesis before it is acted on

For: Special investigations lead examining a suspected organised pattern across referred files

Pattern: Adversarial reviewNeeds live modelsDesigned for 10 to 200 agents

The pain today

A suspected pattern, such as the same repairer invoices across many files, is persuasive once stated. Acting on it wrongly delays honest customers, and nobody has time to argue the innocent explanation file by file.

The ask

I attached the documents from the claim files in my referral and my written hypothesis about a shared repairer invoice pattern. Build the evidence for it from the documents, then try to break it with innocent explanations the same documents support. Show what survives, with sources.

Plain words, as you would say it to a colleague. Edit it to fit your case before you send it.

What you attach or connect

  • Documents from the already referred claim files
  • The written hypothesis
  • Supplier invoices and estimates
  • Investigation guideline

The unit of work

One worker task per one referred file's documents against the hypothesis.

Why a swarm fits

Each file is read alone against one written hypothesis. A separate team with the opposite brief reads the same file, so the case for and the case against are built with equal effort.

Not for

Scoring or flagging individual claims or customers. It tests one written hypothesis against files an investigator already referred.

The decision tree

6 typed decisions, each with an action for every answer

At fixed moments in a run, the engine puts one narrow question to a decision model. The decision model never writes text: it answers yes or no with a probability, picks from listed options, or gives a score, about a small slice of the material. The engine then does exactly what this tree says, which is what makes the run auditable. The thresholds are the template's design values, not measured results.

  1. Planner, while planning

    Scope checkYes or no, with a probability

    Before work starts on a unit

    Is this statement of the hypothesis about what documents show, such as invoice details or dates, and free of any reference to a person's characteristics?

    Sees only: One statement from the planner's breakdown of the hypothesis

    Why: Statements about people rather than documents are refused before any file is read.

    • Yes: 0.85 or higherthenAccept
    • Unsure: 0.60 up to 0.85thenAsk a person
    • No: below 0.60thenSkip this unit
  2. After workers, the judge checks

    Evidence checkYes or no, with a probability

    After a worker answers

    Does the quoted invoice, estimate or file note contain the specific detail the supporting team says matches the pattern?

    Sees only: One hypothesis statement and the quoted document extract offered in support

    Why: A pattern built on misread invoices would wrongly hold up honest customers.

    • Yes: 0.85 or higherthenAccept
    • Unsure: 0.50 up to 0.85thenMark unresolved
    • No: below 0.50thenReject and retry
  3. Evidence checkYes or no, with a probability

    After a worker answers

    Does the quoted document extract support the ordinary explanation the breaking team offers for the same detail?

    Sees only: One hypothesis statement, the detail in question and the extract offered against it

    Why: The innocent explanation is held to the same evidential standard as the suspicion.

    • Yes: 0.85 or higherthenAccept
    • Unsure: 0.50 up to 0.85thenMark unresolved
    • No: below 0.50thenReject and retry
  4. Reconciler, while merging

    Conflict checkA choice among options

    While reconciling

    On the accepted extracts from both teams, how does this statement of the hypothesis stand?

    Sees only: One statement with the accepted supporting and opposing extracts

    Why: Keeps balanced and thin statements open rather than letting a persuasive story win.

    • Survives the objectionthenAccept
    • Broken by the objectionthenAccept
    • Evidence evenly balancedthenMark unresolved
    • Rests on a single filethenMark unresolved
  5. Run control, between rounds

    Another round?Yes or no, with a probability

    Between rounds

    Did the most recent batch of referred files change the standing of any statement of the hypothesis?

    Sees only: Statement standings before and after the last batch

    Why: Stops reading files once further evidence no longer moves the picture.

    • Yes: 0.60 or higherthenContinue
    • Unsure: 0.30 up to 0.60thenContinue
    • No: below 0.30thenStop
  6. Accountable person, before anything is settled

    Person decidesYes or no, with a probability

    Before anything is reported as settled

    Could this report be used to delay, decline, refer or report any claim or customer?

    Sees only: The drafted report of surviving, broken and undecided statements

    Why: Every report goes to the special investigations lead; the run takes no position on any file.

    Accountable: The special investigations lead decides any action. No claim is delayed, declined or reported on the run's output alone.

    • Yes: 0.10 or higherthenAsk a person
    • Unsure: 0.02 up to 0.10thenAsk a person
    • No: below 0.02thenAsk a person

The fleet: who does what

Model tiers by role, not brands: you choose the models. Strong reasoning models plan and reconcile, small fast models do the wide work, and the judge is a decision model from a different family, so it does not share the workers' blind spots.

  1. Planner

    A strong reasoning model breaks the hypothesis into specific, checkable statements about documents.

    Decisions here:1. Scope check

  2. Workers

    Small fast workers in two teams: one cites documents supporting each statement, the other cites innocent explanations.

    Designed for 10 to 200 agents, one worker task per one referred file's documents against the hypothesis. Each worker receives only its own unit.

  3. Judge, from a different model family

    A decision model from a different family decides whether each statement survives its objection on the documents.

    Decisions here:2. Evidence check3. Evidence check

  4. Reconciler

    A strong reasoning model reports surviving, broken and undecided statements without naming any file as fraudulent.

    Decisions here:4. Conflict check5. Another round?

  5. Accountable person

    The special investigations lead decides any action. No claim is delayed, declined or reported on the run's output alone.

    Decisions here:6. Person decides

Checked before anything is accepted

  • Every supporting point and every innocent explanation cites a document
  • No inference is drawn from names, addresses, nationality or other personal characteristics
  • The breaking team never sees the supporting team's reasoning
  • Statements that rest on a single file are marked as weak

What comes back

  • Statements of the hypothesis that survive, with sources
  • Statements broken by an innocent explanation, with sources
  • Undecided statements and what evidence would settle them
  • Documents that were unreadable or missing

What to measure

  • Share of surviving statements the investigator upholds
  • Hypotheses narrowed or dropped before action
  • Investigator hours per referral
  • Cost per referred file read

Names of measures only. No result is claimed for this template.

Templates open in the workspace chat with the ask filled in. Nothing runs until you send it.

Get early accessSign in to use

Map one exposure across the whole wording portfolio

For: Wordings manager or portfolio underwriter responsible for product wordings

When a new exposure appears, nobody knows which wordings, endorsements and legacy versions cover it, exclude it or say nothing.

Pattern: Map, verify, reduceNeeds scale6 decisionsDesigned for 40 to 800 agents

Prepare a coverage brief with clauses for and against

For: Claims handler or coverage specialist preparing a position on a complex commercial loss

The schedule, wording, endorsements and loss report each run long.

Pattern: Cross-examinationRuns today7 decisionsDesigned for 4 to 32 agents

Watch wording changes and recheck only what depends on them

For: Product governance lead or wordings manager at an insurer or managing agent

A clause edit in a base wording silently affects endorsements, customer summaries, claims guidance and broker templates built on it.

Pattern: WatchtowerNeeds a connector5 decisionsDesigned for 20 to 500 agents