Hyperscribe: EHR-Embedded Clinical AI Agent

• Hyperscribe is an AI system integrated within Canvas Medical’s EHR that transforms ambient clinical audio into structured, schema-defined chart updates.
• It employs a multi-stage pipeline—from transcription to executable EHR actions—that ensures high documentation fidelity and auditability.
• Continuous governance through rubric-validation, clinician feedback, and controlled experimentation drives systematic improvements in clinical documentation.

Hyperscribe is an EHR-embedded clinical AI agent developed by Canvas Medical that converts ambient clinical audio and patient context into structured chart updates. In the current literature, it is characterized as an ambient clinical documentation system embedded in the EHR action space, designed around structured outputs, explicit intermediate reasoning, bounded chart actions, and continuous governance rather than around unconstrained free-text note generation alone (Shah et al., 30 Apr 2026, Shah et al., 27 Apr 2026).

1. Definition and scope

Hyperscribe is intended for clinician-facing documentation workflows inside the Canvas Medical EHR. Its function is to observe an encounter, interpret clinically meaningful content, and generate discrete chart updates such as Diagnose, Prescribe, Assess and Plan, and Lab Order, while also contributing to narrative sections including History of Present Illness (HPI), Mental Status Examination, and Assessment and Plan. The system also references updates to medication statements, allergies, past medical history, and questionnaires (Shah et al., 30 Apr 2026).

A central architectural distinction is that Hyperscribe emits typed, schema-defined objects rather than only prose. This makes the system inspectable at multiple intermediate stages and supports stage-specific logging, audit, retry logic, and rubric-based evaluation. The literature therefore treats Hyperscribe not simply as an ambient scribe, but as a chart-aware documentation and chart-update agent operating within an EHR-bounded command space (Shah et al., 27 Apr 2026).

The evaluation corpus used for Hyperscribe spans primary care, psychiatry, oncology, behavioral health, and medication management. The benchmark contains 823 cases, of which 736 are real-world and 87 synthetic, with 35 specialty categories represented. The evaluation target is primarily documentation fidelity: whether the system accurately captures and structures what was said or implied in the encounter using chart context, rather than diagnosis quality or downstream care outcomes (Shah et al., 27 Apr 2026).

2. Processing architecture

Two complementary descriptions of the pipeline appear in the literature. The evaluation paper describes a five-stage process: transcription into timestamped speaker-attributed utterances; conversion of transcript into clinical instructions; conversion of instructions into structured parameters; translation into executable EHR actions; and logging of stage-specific outputs for audit and evaluation. The governance paper presents the operational core as a four-stage cyclical process over successive audio chunks—audio to transcript, transcript to instructions, instructions to parameters, and parameters to commands—while separately emphasizing tiered logging and post-session audit capability (Shah et al., 27 Apr 2026, Shah et al., 30 Apr 2026).

Operationally, Stage 1 records clinical audio and produces timestamped utterances with speaker labels. It is a model-agnostic transcription stage and uses a transcript tail from the previous chunk to preserve continuity across segments. Stage 2 converts speaker-attributed transcript into instructions: clinically meaningful intents mapped to discrete EHR action classes. Each instruction contains a unique identifier, detection order, instruction type, and a free-text information field. The action space is dynamically constrained by patient context, and complex encounters can invoke hierarchical detection by note section rather than a single LLM call (Shah et al., 30 Apr 2026).

Stage 3 translates each instruction into a structured parameter object defined by the target command type. This stage uses ontology mapping services including ICD-10 for diagnoses, RxNorm for medications, and SNOMED CT for clinical findings. It injects minimal patient demographics while deliberately excluding broader chart history at this point. Stage 4 converts parameter objects into executable commands using the richest chart context, including the full patient chart as well as intake questionnaires, vitals entered by medical assistants, and pre-populated note templates (Shah et al., 30 Apr 2026).

The system is designed for near-real-time chunkwise processing and progressive context injection. The literature describes this as graduated introduction of context: raw audio in Stage 1, transcript with speaker attribution in Stage 2, minimal demographics in Stage 3, and full chart context in Stage 4. The same work introduces quadratic revision, the observation that as the number of processing rounds grows, the number of potential revisions increases quadratically because each new round can update documentation produced by any prior round (Shah et al., 30 Apr 2026).

The live system used non-reasoning foundation models—Anthropic Sonnet 4.5 and OpenAI GPT-4.1—because reasoning models were considered too slow and unpredictable for the real-time pipeline. Structured-output validation is backed by a two-layer retry mechanism: up to 3 HTTP retries for transient API failures and up to 3 JSON-level retries for syntactically invalid or schema-nonconforming output (Shah et al., 30 Apr 2026).

3. Case-specific rubric evaluation

The central evaluation methodology defines a case as

$C=(T,N,L),$

where $T$ is the transcript, $N$ is the point-in-time note, and $L$ is the longitudinal patient context. For each case, a rubric is defined as

$R=\{(c_i,w_i)\mid i=1,\ldots,k\},$

where each $c_i$ is a natural-language documentation requirement and each $w_i>0$ is a numeric weight indicating clinical importance. The normalized score for note $n$ under rubric $R$ is

$$S(n,R)=\frac{\sum_i w_i\cdot s_i(c_i)}{\sum_i w_i}\times 100,$$

with $T$0 produced by an LLM-based scoring agent (Shah et al., 27 Apr 2026).

Rubrics are checklist-like, weighted, and criterion-based. Each criterion is phrased as a natural-language statement prefixed with “Reward for”. Two criteria are mandatory in every rubric: one addressing overall completeness relative to the transcript, and one addressing non-repetition of information already documented in the chart. Clinician-authored rubrics contained 1 to 15 criteria, whereas LLM-generated rubrics were constrained to 4, 5, or 6 criteria (Shah et al., 27 Apr 2026).

For each case, two clinicians independently created rubrics. The workflow was to review the transcript and longitudinal context, review Hyperscribe outputs generated under different configurations alongside the point-in-time note, compare 3–5 candidate notes, label the single best and single worst note using holistic clinical judgment, and author rubric criteria spanning clinical accuracy, appropriateness of clinical actions, workflow alignment, and safety. Clinicians could begin either from an empty rubric or from a starter rubric pre-populated from case context, but retained final authority over rubric content (Shah et al., 27 Apr 2026).

Each clinician-authored rubric was then validated against the clinician’s own best-versus-worst preference. A rubric was accepted only if the maximum score across three scoring runs for the clinician-labeled worst note remained below the minimum score across three scoring runs for the best note:

$T$1

This is a strict separation criterion rather than a loose preference correlation test (Shah et al., 27 Apr 2026).

The dataset contains 823 cases, 736 of them real-world and 87 synthetic. The real-world cases came from 168 unique patients. Reported case characteristics include a median transcript length of 419 words, a median of 37 turns, transcript range of 28–9,402 words, and turn range of 1–742. Demographics were available for 100% of cases; point-in-time notes were available for 90.4%. Encounter tags included 14 encounter-type categories, and the case mix included short, medium, and long encounters; single-topic and multi-topic visits; and low-, moderate-, and high-acuity cases (Shah et al., 27 Apr 2026).

Twenty clinicians participated. The literature repeatedly emphasizes 1,646 accepted rubrics, two per case, but also reports 5,797 total rubrics created and 3,060 rubrics achieving validation. A separate cost section reports 3,108 accepted rubrics, which is an internal inconsistency in the papers rather than an external correction (Shah et al., 27 Apr 2026, Shah et al., 30 Apr 2026).

4. Controlled experimentation and empirical results

Seven experimental versions of Hyperscribe were evaluated: Control baseline, Next branch, JSON schema, Hierarchical detection, Model updates, Model switching, and Prompt minimization. For each experiment, 10 outputs per case were generated—5 from OpenAI models and 5 from Anthropic models—across all 823 cases. Each output was scored against three rubric sources, namely clinician rubric 1, clinician rubric 2, and one LLM-generated rubric, with two scoring runs per rubric. The paper states that this yielded over 216,000 scored note-rubric pairs (Shah et al., 27 Apr 2026).

The reported median scores show a marked discontinuity between the earlier four configurations and the later three. The first four experiments cluster around 83–84%, whereas experiments 5–7 are approximately 94–95%. The papers attribute the largest jump primarily to foundation model updates.

Experiment	Median	IQR
Control baseline	84.0%	54.2%–90.0%
Next branch	83.9%	50.7%–90.0%
JSON schema	83.6%	50.0%–89.9%
Hierarchical detection	83.3%	58.5%–89.0%
Model updates	94.7%	81.0%–100.0%
Model switching	94.8%	81.0%–100.0%
Prompt minimization	94.3%	80.0%–100.0%

Rubric discrimination and stability were strong. Across 1,646 rubric-author pairs, the clinician-selected best note outscored the worst note by a mean of 68.51% and a median of 82.92%, with IQR 43.02–95.6%. Repeated scoring of clinician-authored rubrics had median range 0.00%, mean 1.54, and P95 8.00; LLM-authored rubrics had median range 0.75 percentage points, mean 2.01, and P95 8.50 (Shah et al., 27 Apr 2026).

Agreement analysis used Kendall’s $T$2 over rankings of the 10 outputs per case. In experiments 1–4, clinician-clinician agreement exceeded clinician-LLM agreement: for example, experiment 1 reported 0.55 versus 0.39, and experiment 3 reported 0.57 versus 0.44. In experiments 5–7, clinician-LLM agreement matched or exceeded clinician-clinician agreement: experiment 5 reported 0.45 versus 0.38, experiment 6 reported 0.46 versus 0.43, and experiment 7 reported 0.42 versus 0.42. The paper attributes this convergence to both ceiling compression and genuine LLM rubric improvement (Shah et al., 27 Apr 2026).

The later experiments exhibited compressed score distributions: standard deviation declined from about 33% to about 28%, separation margins compressed from 17–24 points to 8–9 points, and $T$3 reached 100%. The authors interpret this as making ranking harder even when outputs are high-quality. They therefore treat convergence in $T$4 as convergence in ranking behavior, not as proof of equivalence in clinical reasoning (Shah et al., 27 Apr 2026).

During clinician best/worst selection, evaluators were blinded to provider identity. Aggregated preference analysis reported Anthropic models with net preference rate +44.2 and selected as best 72.1% of the time, while OpenAI models had net preference rate +13.6 and were selected as best 56.8% of the time (Shah et al., 27 Apr 2026).

5. Continuous governance and deployment behavior

The governance paper distinguishes evaluation from governance in explicit terms: evaluation asks whether a system performs well at a point in time, whereas governance asks whether a system is being actively managed to perform better over time and across conditions. It defines continuous governance as the ongoing practice of monitoring, evaluating, iterating, and re-evaluating performance throughout deployment (Shah et al., 30 Apr 2026).

The governance framework integrates four channels plus a deployment gate: rubric validation, live clinician feedback integration, technical performance monitoring, cost tracking, and controlled experimentation before deployment. Changes motivated by any of these dimensions were required to pass the full benchmark before release. This makes the 823-case rubric benchmark not only an evaluation set, but also a predeployment gating mechanism (Shah et al., 30 Apr 2026).

Live feedback analysis covered 107 entries collected over roughly three months, from September to December 2025. The major themes were command generation failures in 42 entries (39.3%), speaker misattribution in 8 (7.5%), documentation granularity mismatch in 18 (16.8%), workflow or session-control problems in 22 (20.6%), and positive observations in 26 (24.3%). Because entries could contain multiple themes, percentages sum to more than 100%. Over time, feedback composition shifted from 78.6% error reports and 14.3% positive observations in September to 30.0% errors and 45.0% positive observations in December (Shah et al., 30 Apr 2026).

The most common failure category was command generation, including insufficient Assessment and Plan detail and missing information in questionnaires. Speaker misattribution was clinically important because family or clinician disclosures could be placed into the patient’s history. Documentation granularity mismatches reflected outputs that were either too close to verbatim transcription or overly compressed. Workflow failures centered on pause, resume, and stop controls as well as session-state visibility (Shah et al., 30 Apr 2026).

Engineering interventions were correspondingly specific. For command-generation failures, the team introduced a decision tree methodology for instruction detection, JSON schema refactoring, improved prompt guardrails, custom prompting for user-specific command customization, and customer education for SDK configuration issues. For speaker misattribution, the transcript-to-instructions prompt was rewritten to distinguish patient, clinician, and family speakers, and the system switched to a speech-to-text provider with stronger speaker recognition. For workflow failures, the interface moved from text buttons to persistent icon-based controls—Play, Pause, and Stop—added a real-time status indicator, and surfaced a Transcript tab for live verification (Shah et al., 30 Apr 2026).

Operational metrics reported a median processing time per audio segment of 8.1 seconds. Stage-level timing included 1.1 seconds median for audio transcription, 1.6 seconds median for speaker diarization, and 5.9 seconds median for instruction detection, with instruction-detection P95 at 16.3 seconds. Across all system versions, 6.6% of note-generation attempts encountered at least one error; dominant failure modes were unexpected LLM output format and parameter type mismatches. Retries recovered 94.2% of these errors, yielding a 99.6% effective completion rate (Shah et al., 30 Apr 2026).

Cost tracking is treated as a governance channel rather than merely a budgeting concern. Reported figures include \$T$53 for synthetic case construction across 87 cases, 919 clinician hours for 1,646 clinician-authored rubrics, \$T$615,000 for note generation across 57k notes, \$T$725,334 total compute cost. Model-switching experiments reported 20–30% cost reduction depending on vendor with no measurable quality reduction—94.8% versus 94.7% median score—while prompt optimization reduced per-note prompt tokens by 4–6% (Shah et al., 30 Apr 2026).

6. Interpretation, limitations, and broader technical context

Several interpretive boundaries are explicit in the literature. First, Hyperscribe is evaluated primarily for documentation fidelity, not for downstream care outcomes, diagnostic reasoning quality, or patient safety outcomes in practice. Second, the studies are single-system and single-EHR: the governance and evaluation frameworks were validated only on Hyperscribe in Canvas Medical. Third, the methodology depends on an LLM scoring agent, and the papers do not fully characterize hidden scorer biases, scorer model drift, prompt sensitivity, or per-rubric failure modes for LLM-authored rubrics. Fourth, LLM-generated rubrics were not subjected to the same per-case best-versus-worst validation criterion used for clinician-authored rubrics. Fifth, approximately 10.6% of cases were synthetic, specialty coverage is concentrated in primary care, psychiatry, and oncology, and the papers contain internal inconsistencies in rubric counts and at least one internal date inconsistency around a “November 2024 release” versus the September–December 2025 feedback timeline (Shah et al., 27 Apr 2026, Shah et al., 30 Apr 2026).

A common misconception is to treat Hyperscribe as only a free-text note generator. The papers instead describe a system whose behavior is mediated by schema-defined instructions, parameter objects, and bounded command types. Another misconception is to treat high rubric scores as sufficient evidence of comprehensive safety. The literature does not make that claim; it presents rubric-based evaluation as a clinically grounded, economically viable method for repeated benchmarking and governance of documentation AI (Shah et al., 27 Apr 2026, Shah et al., 30 Apr 2026).

In adjacent research, the name Hyperscribe is also used hypothetically rather than descriptively. One paper states that HT boundaries are directly relevant to a system like Hyperscribe if it needs scoped views, modular authoring, overlapping subsystem perspectives, and safe local analysis over a shared Hypernetwork Theory model; another argues that scalable hypergraph embedding would be relevant if a hypothetical Hyperscribe stores, indexes, retrieves, analyzes, or predicts over higher-order relations; a third is relevant to a Hyperscribe-like serving layer if heterogeneous requests, TTFT/TPOT tradeoffs, elastic scaling, and prefill/decode disaggregation matter; and a fourth is relevant if stylus capture, handwriting recognition, ink generation, or pen-trace indexing are in scope. This suggests a broader design space around scoped views, hypergraph-native representations, multi-SLO serving, and digital-ink tokenization, but these works are not primary descriptions of the clinical Hyperscribe system itself (Charlesworth, 7 Dec 2025, Maleki et al., 2021, Yousefijamarani et al., 21 Aug 2025, Wang, 3 Mar 2026).

Taken together, the literature presents Hyperscribe as both a concrete clinical AI documentation agent and a focal example for a broader argument: deployable clinical AI systems should be architected for governability. In the Hyperscribe case, governability is realized through structured intermediate representations, rubric-validated evaluation, live feedback loops, technical monitoring, cost tracking, and controlled experimentation before deployment.