Science
Four analytic products, and what each one may not be used for
These products are deliberately not interchangeable. Most harm in this class of system comes from reading one product as another, so each contract states its permitted and non-permitted interpretations.
Definitions
Product contracts
Phenotype detection — p_osa(x, t)
| Aspect | Statement |
|---|---|
| Question | How likely is it that an unrecognised condition is present now, given the governed record at time t? |
| Permitted | Prioritising evaluation capacity; describing detection likelihood at population level |
| Not permitted | Diagnosis; treatment initiation; denial of coverage; reporting as an event risk |
| Refuses when | Required features are UNKNOWN or CONTRADICTED, or the artifact is not permitted in this environment |
Event prognosis — r_18(x, t)
| Aspect | Statement |
|---|---|
| Question | What is the risk of a defined severe event within 18 months for a population with these characteristics? |
| Permitted | Population stratification; capacity planning; monitoring design |
| Not permitted | Individual determinism; automated denial; substituting for diagnosis or for phenotype likelihood |
| Refuses when | Observation history is insufficient, or temporal leakage protection fails |
Conditional intervention benefit
| Aspect | Statement |
|---|---|
| Question | For this indication and protocol, what benefit would this intervention be expected to produce? |
| Permitted | Comparing protocol-defined strategies under stated assumptions |
| Not permitted | A universal drug score; cross-indication transfer; unqualified benefit claims |
| Refuses when | No validated indication artifact exists — the current state in this build |
External clinical/control arm (ECA)
| Aspect | Statement |
|---|---|
| Question | Compared with a fit-for-purpose external comparator, what would outcomes have been under the alternative strategy for a protocol-defined population? |
| Permitted | Target-trial emulation with a frozen specification and prespecified analysis plan |
| Not permitted | Generating a member alert; acting as a risk score, spend calculation or generic comparison dashboard |
| Refuses when | Overlap is inadequate, positivity is not established, or the comparator is not fit for purpose |
Use cases
Worked examples
These are configured demonstrations over synthetic data. None of them reports a validated production result.
Configured examples
| Use case | What is configured | Limitation |
|---|---|---|
| OSA-18 | Direct-mining priority surface pairing detection likelihood against 18-month event risk | Illustrative synthetic surface only; no validated coefficients, no clinical finding |
| Cognitive / FINGER-informed | Multidomain trajectory framing with an endpoint selector and OSA-burden evidence | Domain framing only; no validated trajectory model and no cognitive diagnosis |
| Keytruda | Indication-specific protocol builder with comparator and estimand support | First configured example; refuses in the absence of a validated indication artifact |
No model in this environment is validated. Experimental models never auto-deny therapy, ECA never produces a member alert, and a missing value is never treated as a negative finding.
Technical note: ECA-Agent. ECA-Agent is the RISX Science method and packaging standard by which a validated analytic use case becomes a bespoke, governed analytic agent inside a client environment: a locked, versioned artifact combining a diagnostic model, a prognostic model, a conditional causal component where justified, and an economic decision policy, with monitoring and controlled retraining. It is not a chatbot and not an autonomous clinical decision-maker. ECA-Agent: Analytic Methodology and Deployment Lifecycle is restricted; authorized technical, reviewer and administrator accounts may sign in to read it.