Views & templates

A template is an organization-specific or catalog-level blueprint: an ordered set of blocks, each declaring what slice of patient state it reads and how it is allowed to write its output. Templates are configuration: what should be built, not a specific patient’s output. Your team selects and customizes them in the Console; see View configuration. A view is the timestamped result for one view category on one patient: block-level content, validity metadata, and pointers to persisted block bodies. Categories include, among others, medication, symptom, functional activity, engagement, emotional state, recent highlights, physiological monitoring, clinical assessments, care context, and protocol-specific views where your organization enables them.

​

How blocks are built

Each block chooses among complementary construction modes, often mixed inside one template, so you can trade latency, cost, and interpretability:

• Deterministic (pipeline) blocks apply fixed operations to resolved inputs: extract, normalize, aggregate, merge, without an open-ended model turn. They shine when the mapping from state to output is well-defined and you want repeatable, inspectable behavior.
• Agentic blocks send filled inputs and often the prior version of that block to a model that drafts or revises narrative or structured content under prompt and schema constraints. They handle fuzzy synthesis, prioritization, and language when deterministic rules would be brittle.
• Supervised pipeline combines both: a deterministic pass establishes a factual scaffold, then a constrained agent step reviews, tightens, or explains; useful when you want machine-checked structure plus nuanced wording.

First-time generation tends to synthesize from scratch; later runs often patch the prior block so updates stay incremental and cheaper. Schedulers skip blocks that are not yet due so work stays proportional to need. Cadence. Each block carries a minimum spacing between runs, for example daily refresh vs weekly narrative. Schedulers compare that to the last successful run so agentic and deterministic paths alike are not re-fired unnecessarily.

​

Confidence metrics

Blocks can attach structured scores so downstream workflows know when to trust, flag, or escalate, not just whether text exists.

• Coverage: did required inputs resolve? Are critical sections populated vs intentionally empty?
• Freshness: how stale are the underlying logs and modules feeding this block relative to the run?
• Certainty: model-reported confidence where an agent produced or revised content.
• Consistency: internal agreement across sections or against upstream structured facts.
• Evidence density: how much corroborating signal supports the narrative versus thin inference.
• Overall: a rollup for dashboards and automation gates when a single threshold is enough.

Those scores are computed from explicit signals such as missingness, gaps in required inputs, freshness of source data, and agentic certainty where a model assigns its own reliability, so clinical and engineering teams can tune policies per block or per view category. Provenance. Values in patient state carry lineage back to contributing logs where applicable. View blocks likewise carry attribution: snapshot-level rollups and, for structured outputs, finer-grained IDs on rows or fields. That helps observability answer “why does this say this?” TEMP highlights are a lightweight, append-only slot on a view for very recent highlights; templates optionally gate which log types or memory tags may populate it. They sit beside main block outputs and can be cleared after a generation pass so staging does not grow without bound.

​

Next steps