Evidence engine
The synlayer Evidence Engine
Cense and Grovera run on one engine. It retrieves graded evidence, lets a language model write only from that evidence, checks every sentence against its source, and holds back any release that makes either product worse.
Architecture
Two products, five shared components, two foundations
CenseDifferential chain: extract → red flags → candidates → compare → rank → refer
GroveraPlan-review chain: profile → match interventions → dose evidence → measures → flags
Shared components
Knowledge-pack frameworkExpert-reviewed, versioned packs
Retrieval and citation checkerPassage-level sources; unmatched sentences removed
Reasoning orchestrationRules first, then retrieval, drafting and checks
Evaluation and release gatesHeld-out sets, blind expert grading
Outcomes engineBuilt for Grovera in 2027, then shared
Foundation modelsThird-party APIs, at least two providers, replaceable. They may cite only the evidence layer.
Compliance baseHIPAA business-associate agreements, credential verification, audit and version logs, institutional accounts.
One engine, two configurations
How each product uses each component
| Component | Cense | Grovera |
|---|---|---|
| Knowledge packs | One pack per rare disease: clinical pattern, red flags, labs, look-alikes, confirmation path, referrals. | Packs per condition, built around interventions: intervention × outcome domain × population. |
| Retrieval and citation checker | Shared. | Shared. |
| Reasoning orchestration | Differential chain ending in ranked candidates, missing evidence and referral. | Plan-review chain ending in evidence grades, dose evidence, measures to track, missing baselines and flags. |
| Evaluation and release gates | 100 held-out cases per disease and language; blind review by two specialists. | Evidence map checked against published reviews (NCAEP, Autism CRC, Cochrane); review by two experts. |
| Outcomes engine | Not used. | Reliable change, severity-adjusted expected trajectories and N-of-1 analysis. Deterministic; no language model. |
| Verification, institutions, audit | Shared. | Shared, plus COPPA requirements for parent-facing features. |
Rules the engine enforces
Five rules, in both products
- 1 No citation, no claimEach drafted sentence is matched to a retrieved passage. A sentence without one is removed before anyone sees it.
- 2 Rules before probabilitiesSpecialist-written red-flag and exclusion rules run first and override model output on safety-critical points.
- 3 Confidence is computedConfidence comes from evidence coverage and rule hits. The model does not grade itself.
- 4 One change, both test setsAny change to the engine runs the evaluation sets of both products. If either product gets worse, the change does not ship.
- 5 Every output can be reproducedPack, prompt and model versions are logged with each output.
Boundaries
What synlayer does not build
We sit between the guideline and the clinician. We don't write guidelines and we don't build foundation models.
Foundation models
Genetic testing or wet-lab work
Drug doses, in any product
Automatic entries into a diagnosis field in the health record
Processing of raw images or physiological signals