Skip to content

Simulate

Stress-test scientific interpretation before the world does.

A finding leaves your hands once. Trace Simulate models how different readers are likely to interpret it — where a caveat gets dropped, where an effect gets generalised, where a limitation disappears — so the sentence that causes the problem can be rewritten while it is still yours to rewrite.

Panel forming

150 synthetic respondents · 5 roles

SpecialistGeneral practicePatientPayerResearcher

27% read the relative effect as absolute (42 of 150)

Modelled response — not a fielded survey.

Demo workspace figures. 150 synthetic respondents resolve into five role clusters — Specialist, General practice, Patient, Payer, Researcher — and 27 percent are marked as having read the relative effect as an absolute one. Modelled response — not a fielded survey.

Simulation disclosure

How the panel is grounded

Modelled stakeholder response, calibrated against real published clinical discourse. Directional research intelligence — it complements human validation rather than replacing it.

Roles are built from how clinicians, reviewers, payers, and patients actually discuss published findings — peer review, editorials, and professional commentary — so a panel reflects the readings a paper genuinely meets rather than a guess at them. Each run reports the sources its roles were calibrated from, alongside its model, instrument version, seed, and sample size.

A panel is a model of readers, so it carries a role label rather than a name. It tells you where a sentence is likely to be misread; it does not tell you what any individual said.

Synthetic panels

A panel is a set of constructed readers, not a sample.

Each synthetic respondent is defined by an explicit profile: discipline, seniority, evidence threshold, and prior exposure to the topic. Profiles are visible, editable, and stored with the run, because a result you cannot inspect is not a result you can argue with.

Demo workspace

Synthetic respondent profile

Illustrative structure for the fictional TRACE-101 trial.

Simulated
Role
Treating specialist
Evidence threshold
Requires a registered primary endpoint
Prior exposure
Familiar with the phase 2 result
Reads
Abstract, then methods, then limitations
Risk posture
Conservative on off-label extension
Stated uncertainty
Reported with every response

Stakeholder presets

Six readers who will not read it the same way.

Presets are starting points, versioned alongside the methodology. You can add roles, remove roles, and change the evidence threshold that defines each one.

Treating clinician

Applicability to the patient in front of them.

Does this population look like mine, and does the effect size change what I do on Monday?

Methodologist

Design, power, and pre-registration.

Was the primary endpoint the one that was registered, and how were missing data handled?

Payer or HTA reviewer

Comparator choice and durability of effect.

Against what standard of care, over what horizon, and at what incremental cost?

Journal editor

Novelty, claim strength, and reporting completeness.

Do the conclusions stay inside what the data can support?

Science journalist

The headline the abstract invites.

What is the single sentence a general reader will take away?

Patient advocate

Burden, access, and harms.

What does this mean for people living with the condition, including those excluded from the trial?

Deliberation

Interpretation is social, so the panel is allowed to argue.

After independent responses are collected, panel members see each other's reasoning and may revise. Both rounds are kept. The interesting signal is usually the movement: which caveat survives contact with a confident summary, and which does not.

Recorded per run

What deliberation produces

  • Independent first-round responses, before any exposure to other roles
  • Revised second-round responses with the reason for each change
  • Positions that converged, and positions that did not
  • The misreadings that appeared in more than one role

Artificial societies

Larger populations, for questions a six-person panel cannot answer.

Where a panel models depth, an artificial society models spread: hundreds of synthetic respondents with varied specialties, evidence thresholds, and exposure sequences, run to see which interpretations become common and which stay marginal.

Population composition is declared before the run and reported with the result. Scale does not confer validity: a thousand synthetic respondents are still a model, and a wide agreement rate is a property of that model rather than of any real community.

Modelled response — not a fielded survey.

Validation

What we can and cannot yet demonstrate.

Validation work is ongoing and reported openly, including negative results.

Currently checked

  • Run-to-run stability of each panel's stated conclusions
  • Sensitivity of output to role wording and prompt order
  • Whether a known distortion inserted into an abstract is detected
  • Agreement between panel-flagged misreadings and reviewer comments on the same text

Not established

  • That agreement rates approximate the views of any real professional population
  • That panel output predicts adoption, prescribing, or coverage decisions
  • That the method generalises evenly across fields and languages

Known limits

  • Panels reflect patterns in language models' training distributions, which under-represent some specialties, regions, and languages.
  • Responses are not calibrated against a matched human sample, so agreement rates are not survey estimates.
  • The same input can produce different phrasings across runs; stability is reported, not hidden.
  • Simulation cannot detect a misreading that no role in the panel is constructed to notice.
  • Output must not be used as evidence of clinician opinion, market research, or regulatory support.

Life sciences

Where teams use it.

Within scientific exchange boundaries. Promotional use requires medical, legal, and regulatory review, and the workspace labels it as such.

Pre-submission review

Surface the objections a methodological reviewer is likely to raise while the manuscript can still be revised.

Abstract and plain-language summary drafting

Identify which sentence produces the overreaching headline before it is written by someone else.

Congress and publication planning

Compare how the same result reads to a specialist audience and to a general one, and plan materials accordingly.

Scientific exchange preparation

Rehearse the questions a medical information team is likely to receive, within a non-promotional frame.

Find the misreading while you can still fix the sentence.

Modelled stakeholder response, calibrated against real published clinical discourse. Directional research intelligence — it complements human validation rather than replacing it.