CONTEXT(CLONE)
Sign in

Trace every summary claim in a dossier back to its study report

For: Regulatory affairs lead or medical writer assembling a marketing application

Pattern: Hierarchical decompositionNeeds live modelsDesigned for 20 to 250 agents
Get early accessSign in to use

The pain today

Summary documents restate results from many study reports. After rounds of edits, a figure or conclusion in a summary no longer matches the report it cites, and the mismatch is found by the agency instead of the team.

The ask

“I attached our clinical overview, the clinical summaries and the study reports they cite. Walk down from each summary section to the statements in it, and for each statement find the table or passage in the cited report that supports it. List statements with no support or a different value.”

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

  • Clinical overview and summaries as text
  • Clinical study report bodies and key tables as text
  • Cross-reference list if one exists

The unit of work

One worker task per one summary statement with its cited source.

Why a swarm fits

The dossier is a tree: overview, summaries, reports. A leaf check needs one statement and one cited section, not the dossier. Each level can be verified before the level above relies on it.

Not for

Judging whether the science supports approval, or checking statistics. It checks that the summary says what the reports say.

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

    Split or keep serialYes or no, with a probability

    While planning

    Does this summary section contain several statements that cite different study reports or different tables?

    Sees only: One summary section with its citations marked

    Why: Avoids splitting short single-source sections where coordination would cost more than reading.

    • Yes: 0.60 or higherthenSplit the work
    • Unsure: 0.40 up to 0.60thenKeep serial
    • No: below 0.40thenKeep serial

  2. Split or keep serialYes or no, with a probability

    While planning

    Does this sentence make more than one result claim, each with its own cited source?

    Sees only: One summary sentence and its citations

    Why: Splits compound sentences so each leaf check has exactly one claim and one source.

    • Yes: 0.60 or higherthenSplit the work
    • Unsure: 0.40 up to 0.60thenKeep serial
    • No: below 0.40thenKeep serial

  3. Scope checkA choice among options

    Before work starts on a unit

    Does the citation in this statement point to a table or section that exists in the supplied report text?

    Sees only: One statement, its citation and the cited report's table of contents

    Why: Finds broken cross-references before a worker searches a whole report for them.

    • Resolves to one locationthenAccept
    • Resolves to several possible locationsthenEscalate to a strong model
    • Does not resolvethenMark unresolved

  4. After workers, the judge checks

    Evidence checkA choice among options

    After a worker answers

    Does the cited table or passage state the value in the summary statement, for the same population, arm and time point?

    Sees only: One summary statement and the single cited table or passage

    Why: Catches the right number quoted for the wrong population, which a digit match alone would pass.

    • States the same value for the same groupthenAccept
    • States a different valuethenMark unresolved
    • Same value but different group or time pointthenMark unresolved
    • Silent on this valuethenReject and retry

  5. Evidence checkYes or no, with a probability

    After a worker answers

    Are all statements under this summary section marked verified, with none unresolved or still pending?

    Sees only: The leaf status list for one summary section

    Why: Stops a section from rolling up as clean while any leaf beneath it is still open.

    • Yes: 0.85 or higherthenAccept
    • Unsure: 0.50 up to 0.85thenEscalate to a strong model
    • No: below 0.50thenReject and retry

  6. Accountable person, before anything is settled

    Person decidesYes or no, with a probability

    Before anything is reported as settled

    Does the mismatched or unsupported statement report an efficacy result, a safety finding or a dosing recommendation?

    Sees only: One flagged statement with the quoted source

    Why: Sends every result-bearing discrepancy to the regulatory lead and medical writer before the text is touched.

    Accountable: The regulatory affairs lead and the responsible medical writer own every correction and the decision to submit.

    • Yes: 0.40 or higherthenAsk a person
    • Unsure: 0.15 up to 0.40thenAsk a person
    • No: below 0.15thenAccept

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 builds the tree from overview to summary section to statement and resolves each citation.

    Decisions here:1. Split or keep serial2. Split or keep serial3. Scope check

  2. Workers

    Small fast workers from an open-weight family check one statement against the single cited table or passage.

    Designed for 20 to 250 agents, one worker task per one summary statement with its cited source. Each worker receives only its own unit.

  3. Judge, from a different model family

    A decision model from a different family decides whether the quoted source actually supports the statement as worded.

    Decisions here:4. Evidence check5. Evidence check

  4. Reconciler

    A strong reasoning model rolls leaf results up per section and marks sections that rest on unsupported statements.

  5. Accountable person

    The regulatory affairs lead and the responsible medical writer own every correction and the decision to submit.

    Decisions here:6. Person decides

Checked before anything is accepted

  • Each statement is tied to a quoted source passage or marked unsupported
  • Values are compared with their units and population labels, not only digits
  • A section rolls up as clean only when all its leaves are verified

What comes back

  • Statement-to-source trace table
  • Unsupported or mismatched statements with both quotes
  • Broken or ambiguous cross-references
  • Section-level status map of the dossier

What to measure

  • Mismatches confirmed by the medical writer
  • Share of statements traced without human help
  • Review time per summary section
  • Cost per dossier pass

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

More in Healthcare and life sciences

Grade protocol deviations across trial sites, with evidence

For: Clinical operations lead or trial quality manager at a sponsor or contract research organisation

Deviation logs and monitoring reports pile up across sites.

Pattern: Map, verify, reduceNeeds scale6 decisionsDesigned for 40 to 400 agents

Check adverse-event narratives against their case data

For: Pharmacovigilance lead or safety writer preparing case narratives for a study report or aggregate report

Hundreds of case narratives are written from line listings.

Pattern: Cross-examinationNeeds scale5 decisionsDesigned for 30 to 300 agents