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