Model-development corpus
Historical consultation material is prepared under a documented data-governance process and used for medical fine-tuning. It is not treated as a live or authoritative source of current clinical guidance.
Evidence infrastructure
Training data and live evidence are not the same thing.
Diagnify describes two separate evidence systems: governed historical consultation material used in model-development workflows, and a maintained library of guidelines and source documents retrieved at inference. The second can inform a live answer without becoming part of the model’s learned parameters.
Current documentation status: Diagnify describes an internal guideline-retrieval layer, but a public source-level register and completed evidence briefs are not yet available on this page. No claim of comprehensive coverage should be inferred.
Two evidence systems
Learn from clinical history. Retrieve current guidance at the point of use.
Keeping these systems distinct makes provenance, currency and failure modes easier to inspect.
Inference-time guideline library
When the product retrieves evidence, the source document remains separate from the model weights. The intended record includes publisher, title, version, publication date, jurisdiction and retrieval date.
Citation and clinician review
Retrieved material must be checked against the cited source and the patient’s context. A citation does not establish that the passage is current, complete or applicable.
Guideline source register
Every source needs an identity and an update state.
A public register should make the grounding library auditable without exposing licensed source text or confidential system details.
- Publisher, title and stable source link
- Jurisdiction, population and clinical scope
- Publication date, version and last-checked date
- Access or licence basis and permitted use
- Current, superseded, withdrawn or under-review status
- Topic tags, exclusions and known coverage gaps
Selection and maintenance controls
- Documented source hierarchy and eligibility rules
- Scheduled review plus urgent safety-update triggers
- Jurisdiction matching rather than silent generalisation
- Conflicting recommendations retained and identified
- Retrieval, quotation and citation-fidelity testing
- Version history and correction route
Status: a source-level register and retrieval-evaluation report have not yet been published on this page.
Retrieval limits
Grounding is a control, not a guarantee.
Retrieval can reduce unsupported output, but it creates its own risks that need separate evaluation.
Relevant evidence can be missed
The search may fail to retrieve the most relevant source or may over-rank a less useful passage.
A source can be wrong for the case
A retrieved guideline can be outdated, jurisdiction-specific, incomplete or inapplicable to the individual patient.
A citation can still mislead
The generated interpretation may overstate, misquote or detach a passage from important context even when a source link is present.
No zero-hallucination claim: neither fine-tuning nor retrieval eliminates hallucinations. Diagnify output still requires source inspection and qualified clinical judgment.
Publishing standard
A visible trail from question to conclusion.
The template is designed so a reader can understand what was asked, what was searched, how evidence was judged, and where judgment or uncertainty remains.
Defined clinical question
Each brief begins with the population, intervention or exposure, comparator and outcomes. Scope exclusions are stated before the search is interpreted.
Reproducible search record
Search date, named sources, search terms, filters and eligibility decisions are recorded so readers can see the evidence window and repeat the work.
Explicit appraisal
Study limitations, risk of bias, effect uncertainty, applicability and any conflicts in the evidence are separated from the final interpretation.
Brief method
What every reviewed brief will contain.
A brief is a structured evidence record, not an unattributed summary. These fields are required before publication.
Frame the PICO and intended use
Define the population, intervention or exposure, comparator and outcomes. State the clinical context, intended professional audience, jurisdiction and questions the brief does not answer.
Record the search
Provide the date searched, databases and guideline sources, complete search strategy, limits, inclusion and exclusion criteria, and the date through which the search can be considered current.
Extract and appraise
Record study design, population, intervention and comparator details, outcomes, effect estimates, measures of uncertainty, follow-up, funding and relevant conflicts. Name the critical-appraisal and certainty framework used, then show the judgments that produced the rating.
Interpret benefits, harms and disagreement
Separate statistical results from clinical interpretation. Present important harms, absolute effects where available, conflicting findings, guideline disagreement and remaining uncertainty without smoothing them into a single confident answer.
Test applicability
Describe whether the studied population, setting, baseline risk, intervention and outcomes fit the intended use. Identify groups for whom the evidence is indirect, sparse or not applicable.
Review, version and correct
Name the author and clinical reviewer, their relevant credentials and declared conflicts. Show the publication date, last search date, next review trigger, version number, material changes, superseded versions and a correction route.
Publication template
The fields readers should expect to see.
The template keeps evidence, interpretation and maintenance information distinct.
| Section | Required record | Why it matters |
|---|---|---|
| Question | PICO, outcomes, scope, audience and jurisdiction | Prevents a broad conclusion from being applied to a different question. |
| Search | Date, sources, complete strategy, limits and eligibility criteria | Shows how current and reproducible the evidence search is. |
| Evidence | Study characteristics, effect estimates, uncertainty, harms and conflicts | Lets readers inspect the findings rather than accept a summary on trust. |
| Appraisal | Named method, risk-of-bias judgments and certainty rationale | Makes the strength of the conclusion traceable to explicit judgments. |
| Applicability | Setting, population, baseline risk, indirectness and exceptions | Clarifies where a finding may or may not transfer to practice. |
| Governance | Author, reviewer, conflicts, review dates, version and change log | Shows accountability and how the brief is maintained over time. |
Interpretation
Certainty is not a badge.
A certainty label is useful only when the framework, outcome and reasons for each judgment are visible.
Outcome-specific first
Certainty can differ across outcomes. A brief should not let a strong finding for one outcome imply equal confidence in every benefit or harm.
Uncertainty retained
Wide intervals, sparse data, indirect populations, selective reporting and inconsistent findings remain visible in the conclusion and are not hidden by confident language.
Evidence and recommendation separated
What studies found is distinct from what a clinician might consider. Patient values, resources, feasibility, baseline risk and local guidance may change a decision.
Jurisdiction identified
Guidelines, medicine availability, regulatory status and standard practice vary. A brief identifies the setting to which each source and interpretation apply.
Clinical context still governs. Evidence briefs are educational resources for qualified health professionals. They do not provide patient-specific medical advice and do not replace local guidance, clinical judgment or verification of primary sources.
Maintenance
A change log, not a silent rewrite.
Each brief will retain enough history for readers to understand when and why its interpretation changed.
Version record
Version number, original publication date, last clinical review and last evidence search appear together.
Material changes
New evidence, changed conclusions, corrected errors and altered scope receive a plain-language change note.
Superseded material
Outdated versions are clearly marked and connected to the current brief rather than silently removed from context.
Continue exploring