Indication reconstruction

The order says “trauma, fall.” The chart says a great deal more.

ARKA rebuilds the clinical indication from the record — problem list, labs, vitals, the triage note, prior reports, pathology — and puts a source link on every element. Then it scores that, instead of the two words someone typed at 3am.

For Radiology, CMO/ACMO, and IT: incomplete requisitions cost interpretation time and put the wrong text under the score. ARKA returns a provenance-bound indication the engine can score and the radiologist can trust.

Published evidence How it works Scopes & FHIR Free baseline

Published evidence

The evidence that this is a real problem

Four published findings — assembled in one place, fully cited, and labelled. A radiologist reading this section learns something; a committee can take the DOIs into the room.

Who this is for: Radiology and CMO leaders who need hard published numbers for a committee packet.
What it costs them today: Incomplete or discordant requisitions drive re-work, delayed reads, and CDS that scores the wrong text.
What ARKA returns: Four peer-reviewed anchors in one place — with resolvable DOIs — so a meeting starts from evidence, not anecdote.

Four published findings on requisition quality, RI-RADS distribution, LLM extraction, and automated RI-RADS scoring
Finding	Statistic	Provenance	Citation
Requisition integrity vs provider notes n = 315 orders relative to the provider's own notes	81% incomplete · 42% discordant	MeasuredMeasured, published (JAMIA 2018)	Lacson R, et al. Integrity of clinical information in computerized order requisitions for diagnostic imaging. JAMIA. 2018. (opens in new tab)
Inpatient RI-RADS grade distribution n = 762 inpatient imaging referrals	A 1% · B 7% · C 31% · D 53% · X 8%	MeasuredMeasured, published (Insights Imaging 2024, n = 762)	Parillo M, et al. Assessment of RI-RADS in inpatient diagnostic imaging referrals. Insights Imaging. 2024. (opens in new tab)
GPT-4 extraction of 10 oncologic-imaging parameters Generated histories preferred by radiologists over original requisitions	F1 0.983	MeasuredMeasured, published (Radiology 2025)	Bhayana R, et al. Leveraging large language models to generate clinical histories for oncologic imaging requisitions. Radiology. 2025. (opens in new tab)
Automated RI-RADS scoring (external validation) External n = 18,887; F1 ≤ 0.49 on rare A/B grades — the region reconstructed indications land in	F1 0.93 · κ 0.88	MeasuredMeasured, published (Eur Radiol 2025, external validation)	Büyüktoka RE, et al. Applying large language model for automated quality scoring of radiology requisitions using a standardized criteria. Eur Radiol. 2025. (opens in new tab)

Full synthesis and related citations live in the evidence registry entry on requisition quality.

Architecture

How it works

Six stages. Two phases. Latency is not hidden behind a slide — it is the product contract.

Who this is for: Clinical informatics and CDS engineers who have shipped inside the ~500 ms order-entry envelope.
What it costs them today: Any architecture that pretends note-reading fits on the synchronous path will be rejected by the first Epic analyst.
What ARKA returns: A two-phase design they can audit — 350 ms structured on order-select; narrative pre-warmed on chart open.

≤ 350 ms p95

Phase A — structured reconstruction on the synchronous order-select path (prefetch-only; no free text)

MeasuredBudget target, enforced in ire:latency / CI acceptance gates

≤ 4,000 ms

Phase B — narrative reconstruction, pre-warmed on chart open (patient-view). Never blocks order-select.

MeasuredBudget target; off the critical path by architecture

Retrieve
Fetch the nine resource classes under a hard budget and record exactly what could not be fetched. Missing classes narrow the reconstruction; they never fail the card.
Extract
Propose span-anchored clinical elements from structured fields (Phase A) and narrative notes/reports (Phase B), under the closed nine-type schema and provenance gate.
Normalise
Code concepts, enforce verbatim span grounding, and dedupe so each admitted element is unique, coded where possible, and span-valid.
Assemble
Build the StructuredIndication — elements, completeness, degradation, diagnostic question — from admitted elements only.
Grade
Score RI-RADS on typed vs reconstructed indications and compute Documentation Concordance (completeness gap and discordance, separately).
Emit
Map the StructuredIndication to consumers (AIIE context, PA packet, worklist, clinical scenario) without altering deterministic decision logic.

Anyone who has deployed CDS knows the ~500 ms constraint. Phase A is sized for that envelope (house: 800 ms API p95). Note reading is pre-warmed when the chart opens so the card at order-select can already carry a fuller reconstruction when the cache is warm — and states “structured sources only” when it is not.

Implementation dossier

The nine resource classes

Exact FHIR searches. Exact SMART scopes. All .rs (read/search only). This is half of the Implementation Dossier — the answer to the integration objection before it is raised.

Who this is for: Epic analysts and IT security reviewing scope expansion.
What it costs them today: Integration objections stall before clinical value is visible — bulk export, write scopes, opaque data pulls.
What ARKA returns: Nine read-only SMART scopes, exact FHIR searches, one hook to disable. Half of the Implementation Dossier, on the page.

IRE resource class tiers with FHIR search templates and SMART scopes
Tier	Resource	FHIR search	SMART scope	Phase
T0	Patient	Patient/{{id}}	(from existing app launch / context)	A
T1	Condition Active conditions define the differential the order is working — core structured indication material.	Condition?patient={{context.patientId}}&clinical-status=active&_count=100	patient/Condition.rs	A
T1	Encounter ED vs inpatient vs ambulatory changes the decision frame for the reconstructed indication.	Encounter?patient={{context.patientId}}&_sort=-date&_count=5	patient/Encounter.rs	A
T1	Observation — labs Recent labs (WBC, lactate, CRP, troponin, D-dimer, LFTs, lipase, INR, hCG, eGFR) change study territory and urgency framing.	Observation?patient={{context.patientId}}&category=laboratory&code=6690-2,26464-8,2524-7,32693-4,1988-5,30522-7,10839-9,6598-7,67151-1,48065-7,48066-5,48067-3,7799-0,1742-6,1920-8,6768-6,1975-2,3040-3,6301-6,34714-6,2106-3,80384-1,33914-3,69405-9,62238-1,77147-7,2160-0&_sort=-date&_count=50	patient/Observation.rs	A
T1	Observation — vitals Vitals (HR, BP, SpO₂, temp, RR, FiO₂) ground acuity and exam-finding elements for Phase A.	Observation?patient={{context.patientId}}&category=vital-signs&_sort=-date&_count=20	patient/Observation.rs	A
T2	AllergyIntolerance Contrast and other allergies support premedication or modality-change context in the chart.	AllergyIntolerance?patient={{context.patientId}}&clinical-status=active&_count=50	patient/AllergyIntolerance.rs	A
T2	Device Implanted devices and MRI conditionality matter to indication completeness and modality choice context.	Device?patient={{context.patientId}}&_count=20	patient/Device.rs	A
T2	MedicationRequest Active meds (anticoagulation, metformin, nephrotoxics) enrich mechanism and prior-treatment context.	MedicationRequest?patient={{context.patientId}}&status=active&_count=100	patient/MedicationRequest.rs	A
T2	Procedure Recent procedures and surgeries frame implant/post-op status relevant to the indication.	Procedure?patient={{context.patientId}}&_sort=-date&_count=20	patient/Procedure.rs	A
T3	DocumentReference Clinical notes are the richest narrative source for Phase B (mechanism, red flags, functional impact).	DocumentReference?patient={{context.patientId}}&category=clinical-note&_sort=-date&_count=10	patient/DocumentReference.rs	B
T3	DiagnosticReport — pathology Pathology reports (known malignancy) change staging-protocol framing for the indication.	DiagnosticReport?patient={{context.patientId}}&category=http://loinc.org\|LP7839-6&_sort=-date&_count=10	patient/DiagnosticReport.rs	B
T3	DiagnosticReport — radiology Prior radiology reports change protocol context (e.g. delayed-phase follow-up) for the reconstructed indication.	DiagnosticReport?patient={{context.patientId}}&category=http://loinc.org\|LP29684-5&_sort=-date&_count=10	patient/DiagnosticReport.rs	B

Nine read-only scopes. No write access. No bulk export. You turn it off by disabling one hook.

Deduplicated SMART scopes required for IRE prefetch (every entry ends in .rs):

patient/AllergyIntolerance.rs
patient/Condition.rs
patient/Device.rs
patient/DiagnosticReport.rs
patient/DocumentReference.rs
patient/Encounter.rs
patient/MedicationRequest.rs
patient/Observation.rs
patient/Procedure.rs

Hard gate

Provenance

Every element carries a resource, a date, and the exact characters it came from. Click any of them.

Who this is for: Compliance, ACMO, and any buyer who will not adopt a black-box extraction story.
What it costs them today: Unsourced 'AI summaries' cannot support an audit, a payer contract, or a liability defense.
What ARKA returns: Every admitted element carries a resource, a date, and the exact characters it came from — or it never enters the indication.

100%

of reconstructed elements in an ARKA indication carry a resolvable SourceSpan — resource type, id, path, effective date, start, end, and quote

MeasuredArchitectural guarantee — provenanceCompleteness === 1.0, enforced by test before deploy

An element without a resolvable source never enters an ARKA indication. That is a hard gate in the code, enforced by a test that must pass before we deploy. Verbatim grounding: span.quote must equal sourceText.slice(start, end) — mismatch drops the element.

Scenario

Vol II §15.3 opener · synthetic chart · offline fixture

SIMEHR — SIMULATED PATIENT CHART · IRE beats 1–3

SimEHROrdersDemo VolTwo

What arrives.

MRI lumbar spine — back pain

Two words. This is what actually arrives.

D — deficient

D: one category

impressionabsent
findingspresent
questionabsent

What the chart already contains.

A — adequate

A: three categories, rich detail

impressionpresent
findingspresent
questionpresent

What changes.

duration
Typed
248
Reconstructed
6
elements: 6b8726a58215e55c, c95c6ad84cbde38e
redFlags
Typed
[{"flag":"Cancer history","present":false},{"flag":"Neurological deficit","present":false},{"flag":"Fever with back pain","present":false},{"flag":"Saddle anesthesia","present":false},{"flag":"Bowel dysfunction","present":false},{"flag":"Bladder dysfunction","present":false},{"flag":"Thunderclap headache onset","present":false},{"flag":"Worst headache of life","present":false},{"flag":"Immunocompromised state","present":false},{"flag":"Anticoagulation therapy","present":false},{"flag":"Trauma","present":false}]
Reconstructed
[{"flag":"Cancer history","present":false},{"flag":"Neurological deficit","present":false},{"flag":"Fever with back pain","present":false},{"flag":"Saddle anesthesia","present":false},{"flag":"Bowel dysfunction","present":false},{"flag":"Bladder dysfunction","present":false},{"flag":"Thunderclap headache onset","present":false},{"flag":"Worst headache of life","present":false},{"flag":"Immunocompromised state","present":false},{"flag":"Anticoagulation therapy","present":false},{"flag":"Trauma","present":false},{"flag":"Fever","present":true}]
elements: a6a6f77630716ff3
symptoms
Typed
—
Reconstructed
Low Back Pain, Back Pain
elements: 112515b883be40fe, b6686ca8eb617312, fc682f7082be61f3

Typed rating

usually appropriate

Chart rating

usually appropriate

Δ +3

Documentation Concordance Score

13.6%

1 − completeness gap (charted decision-relevant weight missing from the typed indication). Hover for the published definition.

Synthetic charts only. Click a reconstructed element to open its source span.

Stochastic / deterministic

The firewall

A language model may propose a coded element with provenance. It may never produce, influence, or be an input to a rating, decision, protocol, safety gate, or payer-contract field.

Who this is for: Technical buyers who have watched LLM demos collapse under payer contracts and audit review.
What it costs them today: A stochastic model on the decision path is an un-contractable artefact — and a regulatory story that does not close.
What ARKA returns: A machine-readable firewall: models may propose elements; they may never rate, approve, protocol, gate, or write the audit trail.

Permitted

extract candidate clinical elements from free-text notes
draft the human-readable indication paragraph
draft the medical-necessity narrative
draft the patient-facing shared-decision script

Forbidden

the appropriateness rating
the approve/deny decision
the protocol selection
any safety gate
anything written to an audit trail or a payer contract

This table is rendered from docs/arka-ire/firewall.json (v1.0.0) — the same file npm run lint:ire and the firewall tests enforce. Enforcement: import-boundary test; decision-surface test; runtime assertNoModelDerivedInput; npm run lint:ire.

Full AI boundary page — where we use AI, and where we refuse to

Anti-gaming

You can't type your way past the chart

Under ARKA, the scored artefact is not what was typed — it's what the record supports. To manufacture an appropriate verdict you would have to fabricate documentation, which is a materially different act with a materially different consequence, and it leaves an audit trail.

Who this is for: Any buyer who has seen keyword-stuffed indications beat rule engines.
What it costs them today: Gaming the typed text manufactures an appropriateness verdict without changing the chart — and leaves no honest audit trail of the fabrication.
What ARKA returns: The scored artefact is what the record supports. Stuffing keywords into the order does not move the rating unless the documentation is also fabricated — a different act, with a different consequence.

Use Show the gaming attempt on the demo and watch the appropriateness rating not move when the typed indication is stuffed with keywords the chart does not contain.

Scenario

Vol II §15.3 opener · synthetic chart · offline fixture

SIMEHR — SIMULATED PATIENT CHART · IRE beats 1–3

SimEHROrdersDemo VolTwo

What arrives.

MRI lumbar spine — back pain

Two words. This is what actually arrives.

D — deficient

D: one category

impressionabsent
findingspresent
questionabsent

What the chart already contains.

A — adequate

A: three categories, rich detail

impressionpresent
findingspresent
questionpresent

What changes.

duration
Typed
248
Reconstructed
6
elements: 6b8726a58215e55c, c95c6ad84cbde38e
redFlags
Typed
[{"flag":"Cancer history","present":false},{"flag":"Neurological deficit","present":false},{"flag":"Fever with back pain","present":false},{"flag":"Saddle anesthesia","present":false},{"flag":"Bowel dysfunction","present":false},{"flag":"Bladder dysfunction","present":false},{"flag":"Thunderclap headache onset","present":false},{"flag":"Worst headache of life","present":false},{"flag":"Immunocompromised state","present":false},{"flag":"Anticoagulation therapy","present":false},{"flag":"Trauma","present":false}]
Reconstructed
[{"flag":"Cancer history","present":false},{"flag":"Neurological deficit","present":false},{"flag":"Fever with back pain","present":false},{"flag":"Saddle anesthesia","present":false},{"flag":"Bowel dysfunction","present":false},{"flag":"Bladder dysfunction","present":false},{"flag":"Thunderclap headache onset","present":false},{"flag":"Worst headache of life","present":false},{"flag":"Immunocompromised state","present":false},{"flag":"Anticoagulation therapy","present":false},{"flag":"Trauma","present":false},{"flag":"Fever","present":true}]
elements: a6a6f77630716ff3
symptoms
Typed
—
Reconstructed
Low Back Pain, Back Pain
elements: 112515b883be40fe, b6686ca8eb617312, fc682f7082be61f3

Typed rating

usually appropriate

Chart rating

usually appropriate

Δ +3

Documentation Concordance Score

13.6%

1 − completeness gap (charted decision-relevant weight missing from the typed indication). Hover for the published definition.

Documentation Concordance Score

Where asserted indications and charted evidence part company

By service line, with n and a confidence interval. Completeness gap and discordance rate are reported separately — they are different failure modes with different remedies (JAMIA 2018 anchor).

Who this is for: Associate CMO and compliance — the documentation buyer Baugh named.
What it costs them today: Without a cohort number, documentation gaps stay anecdotal and never become a service-line program.
What ARKA returns: Where asserted indications and charted evidence part company, by service line, with n and a confidence interval — and small-cell suppression you can show legal.

Sample report · redacted · synthetic extract

Baseline Capture — Documentation Concordance

Illustrative sample output from the offline baseline kit. Not a customer result.

IllustrativeIllustrative — synthetic cohort kit sample-output

16.7%

Mean DCS (n = 63)

Illustrative95% CI 13.7% – 19.9% (bootstrap)

0.0%

Discordance rate

IllustrativeWilson 95% CI 0.0% – 5.7%

mechanism

Element most often missing from requisitions (100.0% of orders)

Illustrative63 of 63 orders in sample

n < 11

Small-cell suppression threshold

MeasuredEnforced rule: strata with n < 11 never appear in public tables

Mean DCS by service line (unsuppressed strata only)

Service line	n	Mean DCS	95% CI
Emergency	28	14.0%	0.10–0.19
Medicine	22	18.0%	0.13–0.24
Surgery	13	21.0%	0.14–0.29

Small-cell suppression is active and visible

Any stratum with n < 11 is suppressed from clinician and service-line tables and listed for auditability — never rendered as a rate.

service_line: pilot_lab — n = 4 < 11 (small-cell suppression)
clinician_hash: 03424174937a3431 — n = 1 < 11 (small-cell suppression)

For radiologists

The reconstructed indication ships into the read

You open a study and the clinical question, the relevant labs, the priors, the medications and the pathology are already assembled, with source links. You stop reading the chart by hand.

Who this is for: Radiologists who currently rebuild the chart by hand before every complex read.
What it costs them today: Minutes of chart archaeology per study — priors, labs, meds, pathology — for a two-word indication that rarely helps.
What ARKA returns: The reconstructed indication ships into the read: clinical question, labs, priors, meds, pathology, source links. You stop reading the chart by hand.

Clinical question first
The diagnostic question is derived from the order plus the highest-weight charted elements — not from the requisition free text alone.
Labs and vitals that change the frame
WBC, lactate, eGFR, anticoagulation — the same signals a careful radiologist would dig for, already linked to Observation resources.
Priors and pathology
Prior radiology and pathology reports are narrative Phase-B classes. They surface with dates and source DocumentReference / DiagnosticReport ids.
Clickable grounding
Every chip is a receipt. A false or missing element cannot hide behind a fluent paragraph — it either has a span or it is not in the indication.

Honest status

What is built, and what is not

A pre-revenue company that publishes its own build status survives a technical reference call. One that implies everything ships today does not.

Who this is for: Technical reference callers who will hang up the moment a demo overclaims.
What it costs them today: Implied 'we ship everything today' fails the first deep diligence question.
What ARKA returns: A published build ledger — built, in build, not complete — with links to protocol and evaluation methodology.

Built
Appropriateness engine (AIIE)
Deterministic scenario-level rating with evidence links and audit trail.
Built
CDS Hooks / SMART / Da Vinci integration
order-select, patient-view pre-warm, SMART launch, PA packet emission paths.
Built
Auto-approve-only architecture & federated analytics
Approve path only; analytics de-identified. Denials stay human.
In build
Indication reconstruction
In build against CT abdomen/pelvis; seeking one design-partner site for shadow evaluation.
Not complete
Prospective clinical validation
Protocol and statistical analysis plan drafted; enrollment and outcomes not claimed as complete.

Prospective clinical validation is not complete. Review the evaluation ladder, protocol posture, and planned pilot design on How ARKA validates and the pilot plan (protocol & measurement). SAP and enrollment milestones track there — not as completed outcomes on this page.

Next step

Measure your baseline — free.

Send a 90-day extract of imaging orders. We return your RI-RADS distribution, your Documentation Concordance Score, and the single element most often missing from your requisitions. No PHI, no integration, no commitment. If the number isn't there, we'll tell you and we'll stop.

Measure your baseline — free Book a live walkthrough

Indication reconstruction

The order says “trauma, fall.” The chart says a great deal more.

Published evidence How it works Scopes & FHIR Free baseline

Published evidence

The evidence that this is a real problem

Four published findings — assembled in one place, fully cited, and labelled. A radiologist reading this section learns something; a committee can take the DOIs into the room.

Who this is for: Radiology and CMO leaders who need hard published numbers for a committee packet.
What it costs them today: Incomplete or discordant requisitions drive re-work, delayed reads, and CDS that scores the wrong text.
What ARKA returns: Four peer-reviewed anchors in one place — with resolvable DOIs — so a meeting starts from evidence, not anecdote.

Four published findings on requisition quality, RI-RADS distribution, LLM extraction, and automated RI-RADS scoring
Finding	Statistic	Provenance	Citation
Requisition integrity vs provider notes n = 315 orders relative to the provider's own notes	81% incomplete · 42% discordant	MeasuredMeasured, published (JAMIA 2018)	Lacson R, et al. Integrity of clinical information in computerized order requisitions for diagnostic imaging. JAMIA. 2018. (opens in new tab)
Inpatient RI-RADS grade distribution n = 762 inpatient imaging referrals	A 1% · B 7% · C 31% · D 53% · X 8%	MeasuredMeasured, published (Insights Imaging 2024, n = 762)	Parillo M, et al. Assessment of RI-RADS in inpatient diagnostic imaging referrals. Insights Imaging. 2024. (opens in new tab)
GPT-4 extraction of 10 oncologic-imaging parameters Generated histories preferred by radiologists over original requisitions	F1 0.983	MeasuredMeasured, published (Radiology 2025)	Bhayana R, et al. Leveraging large language models to generate clinical histories for oncologic imaging requisitions. Radiology. 2025. (opens in new tab)
Automated RI-RADS scoring (external validation) External n = 18,887; F1 ≤ 0.49 on rare A/B grades — the region reconstructed indications land in	F1 0.93 · κ 0.88	MeasuredMeasured, published (Eur Radiol 2025, external validation)	Büyüktoka RE, et al. Applying large language model for automated quality scoring of radiology requisitions using a standardized criteria. Eur Radiol. 2025. (opens in new tab)

Full synthesis and related citations live in the evidence registry entry on requisition quality.

Architecture

How it works

Six stages. Two phases. Latency is not hidden behind a slide — it is the product contract.

Who this is for: Clinical informatics and CDS engineers who have shipped inside the ~500 ms order-entry envelope.
What it costs them today: Any architecture that pretends note-reading fits on the synchronous path will be rejected by the first Epic analyst.
What ARKA returns: A two-phase design they can audit — 350 ms structured on order-select; narrative pre-warmed on chart open.

≤ 350 ms p95

Phase A — structured reconstruction on the synchronous order-select path (prefetch-only; no free text)

MeasuredBudget target, enforced in ire:latency / CI acceptance gates

≤ 4,000 ms

Phase B — narrative reconstruction, pre-warmed on chart open (patient-view). Never blocks order-select.

MeasuredBudget target; off the critical path by architecture

Retrieve
Fetch the nine resource classes under a hard budget and record exactly what could not be fetched. Missing classes narrow the reconstruction; they never fail the card.
Extract
Propose span-anchored clinical elements from structured fields (Phase A) and narrative notes/reports (Phase B), under the closed nine-type schema and provenance gate.
Normalise
Code concepts, enforce verbatim span grounding, and dedupe so each admitted element is unique, coded where possible, and span-valid.
Assemble
Build the StructuredIndication — elements, completeness, degradation, diagnostic question — from admitted elements only.
Grade
Score RI-RADS on typed vs reconstructed indications and compute Documentation Concordance (completeness gap and discordance, separately).
Emit
Map the StructuredIndication to consumers (AIIE context, PA packet, worklist, clinical scenario) without altering deterministic decision logic.

Implementation dossier

The nine resource classes

Exact FHIR searches. Exact SMART scopes. All .rs (read/search only). This is half of the Implementation Dossier — the answer to the integration objection before it is raised.

Who this is for: Epic analysts and IT security reviewing scope expansion.
What it costs them today: Integration objections stall before clinical value is visible — bulk export, write scopes, opaque data pulls.
What ARKA returns: Nine read-only SMART scopes, exact FHIR searches, one hook to disable. Half of the Implementation Dossier, on the page.

IRE resource class tiers with FHIR search templates and SMART scopes
Tier	Resource	FHIR search	SMART scope	Phase
T0	Patient	Patient/{{id}}	(from existing app launch / context)	A
T1	Condition Active conditions define the differential the order is working — core structured indication material.	Condition?patient={{context.patientId}}&clinical-status=active&_count=100	patient/Condition.rs	A
T1	Encounter ED vs inpatient vs ambulatory changes the decision frame for the reconstructed indication.	Encounter?patient={{context.patientId}}&_sort=-date&_count=5	patient/Encounter.rs	A
T1	Observation — labs Recent labs (WBC, lactate, CRP, troponin, D-dimer, LFTs, lipase, INR, hCG, eGFR) change study territory and urgency framing.	Observation?patient={{context.patientId}}&category=laboratory&code=6690-2,26464-8,2524-7,32693-4,1988-5,30522-7,10839-9,6598-7,67151-1,48065-7,48066-5,48067-3,7799-0,1742-6,1920-8,6768-6,1975-2,3040-3,6301-6,34714-6,2106-3,80384-1,33914-3,69405-9,62238-1,77147-7,2160-0&_sort=-date&_count=50	patient/Observation.rs	A
T1	Observation — vitals Vitals (HR, BP, SpO₂, temp, RR, FiO₂) ground acuity and exam-finding elements for Phase A.	Observation?patient={{context.patientId}}&category=vital-signs&_sort=-date&_count=20	patient/Observation.rs	A
T2	AllergyIntolerance Contrast and other allergies support premedication or modality-change context in the chart.	AllergyIntolerance?patient={{context.patientId}}&clinical-status=active&_count=50	patient/AllergyIntolerance.rs	A
T2	Device Implanted devices and MRI conditionality matter to indication completeness and modality choice context.	Device?patient={{context.patientId}}&_count=20	patient/Device.rs	A
T2	MedicationRequest Active meds (anticoagulation, metformin, nephrotoxics) enrich mechanism and prior-treatment context.	MedicationRequest?patient={{context.patientId}}&status=active&_count=100	patient/MedicationRequest.rs	A
T2	Procedure Recent procedures and surgeries frame implant/post-op status relevant to the indication.	Procedure?patient={{context.patientId}}&_sort=-date&_count=20	patient/Procedure.rs	A
T3	DocumentReference Clinical notes are the richest narrative source for Phase B (mechanism, red flags, functional impact).	DocumentReference?patient={{context.patientId}}&category=clinical-note&_sort=-date&_count=10	patient/DocumentReference.rs	B
T3	DiagnosticReport — pathology Pathology reports (known malignancy) change staging-protocol framing for the indication.	DiagnosticReport?patient={{context.patientId}}&category=http://loinc.org\|LP7839-6&_sort=-date&_count=10	patient/DiagnosticReport.rs	B
T3	DiagnosticReport — radiology Prior radiology reports change protocol context (e.g. delayed-phase follow-up) for the reconstructed indication.	DiagnosticReport?patient={{context.patientId}}&category=http://loinc.org\|LP29684-5&_sort=-date&_count=10	patient/DiagnosticReport.rs	B

Nine read-only scopes. No write access. No bulk export. You turn it off by disabling one hook.

Deduplicated SMART scopes required for IRE prefetch (every entry ends in .rs):

patient/AllergyIntolerance.rs
patient/Condition.rs
patient/Device.rs
patient/DiagnosticReport.rs
patient/DocumentReference.rs
patient/Encounter.rs
patient/MedicationRequest.rs
patient/Observation.rs
patient/Procedure.rs

Hard gate

Provenance

Every element carries a resource, a date, and the exact characters it came from. Click any of them.

Who this is for: Compliance, ACMO, and any buyer who will not adopt a black-box extraction story.
What it costs them today: Unsourced 'AI summaries' cannot support an audit, a payer contract, or a liability defense.
What ARKA returns: Every admitted element carries a resource, a date, and the exact characters it came from — or it never enters the indication.

100%

of reconstructed elements in an ARKA indication carry a resolvable SourceSpan — resource type, id, path, effective date, start, end, and quote

MeasuredArchitectural guarantee — provenanceCompleteness === 1.0, enforced by test before deploy

Scenario

Vol II §15.3 opener · synthetic chart · offline fixture

SIMEHR — SIMULATED PATIENT CHART · IRE beats 1–3

SimEHROrdersDemo VolTwo

What arrives.

MRI lumbar spine — back pain

Two words. This is what actually arrives.

D — deficient

D: one category

impressionabsent
findingspresent
questionabsent

What the chart already contains.

A — adequate

A: three categories, rich detail

impressionpresent
findingspresent
questionpresent

What changes.

duration
Typed
248
Reconstructed
6
elements: 6b8726a58215e55c, c95c6ad84cbde38e
redFlags
Typed
[{"flag":"Cancer history","present":false},{"flag":"Neurological deficit","present":false},{"flag":"Fever with back pain","present":false},{"flag":"Saddle anesthesia","present":false},{"flag":"Bowel dysfunction","present":false},{"flag":"Bladder dysfunction","present":false},{"flag":"Thunderclap headache onset","present":false},{"flag":"Worst headache of life","present":false},{"flag":"Immunocompromised state","present":false},{"flag":"Anticoagulation therapy","present":false},{"flag":"Trauma","present":false}]
Reconstructed
[{"flag":"Cancer history","present":false},{"flag":"Neurological deficit","present":false},{"flag":"Fever with back pain","present":false},{"flag":"Saddle anesthesia","present":false},{"flag":"Bowel dysfunction","present":false},{"flag":"Bladder dysfunction","present":false},{"flag":"Thunderclap headache onset","present":false},{"flag":"Worst headache of life","present":false},{"flag":"Immunocompromised state","present":false},{"flag":"Anticoagulation therapy","present":false},{"flag":"Trauma","present":false},{"flag":"Fever","present":true}]
elements: a6a6f77630716ff3
symptoms
Typed
—
Reconstructed
Low Back Pain, Back Pain
elements: 112515b883be40fe, b6686ca8eb617312, fc682f7082be61f3

Typed rating

usually appropriate

Chart rating

usually appropriate

Δ +3

Documentation Concordance Score

13.6%

1 − completeness gap (charted decision-relevant weight missing from the typed indication). Hover for the published definition.

Synthetic charts only. Click a reconstructed element to open its source span.

Stochastic / deterministic

The firewall

A language model may propose a coded element with provenance. It may never produce, influence, or be an input to a rating, decision, protocol, safety gate, or payer-contract field.

Who this is for: Technical buyers who have watched LLM demos collapse under payer contracts and audit review.
What it costs them today: A stochastic model on the decision path is an un-contractable artefact — and a regulatory story that does not close.
What ARKA returns: A machine-readable firewall: models may propose elements; they may never rate, approve, protocol, gate, or write the audit trail.

Permitted

extract candidate clinical elements from free-text notes
draft the human-readable indication paragraph
draft the medical-necessity narrative
draft the patient-facing shared-decision script

Forbidden

the appropriateness rating
the approve/deny decision
the protocol selection
any safety gate
anything written to an audit trail or a payer contract

Full AI boundary page — where we use AI, and where we refuse to

Anti-gaming

You can't type your way past the chart

Who this is for: Any buyer who has seen keyword-stuffed indications beat rule engines.
What it costs them today: Gaming the typed text manufactures an appropriateness verdict without changing the chart — and leaves no honest audit trail of the fabrication.
What ARKA returns: The scored artefact is what the record supports. Stuffing keywords into the order does not move the rating unless the documentation is also fabricated — a different act, with a different consequence.

Use Show the gaming attempt on the demo and watch the appropriateness rating not move when the typed indication is stuffed with keywords the chart does not contain.

Scenario

Vol II §15.3 opener · synthetic chart · offline fixture

SIMEHR — SIMULATED PATIENT CHART · IRE beats 1–3

SimEHROrdersDemo VolTwo

What arrives.

MRI lumbar spine — back pain

Two words. This is what actually arrives.

D — deficient

D: one category

impressionabsent
findingspresent
questionabsent

What the chart already contains.

A — adequate

A: three categories, rich detail

impressionpresent
findingspresent
questionpresent

What changes.

duration
Typed
248
Reconstructed
6
elements: 6b8726a58215e55c, c95c6ad84cbde38e
redFlags
Typed
[{"flag":"Cancer history","present":false},{"flag":"Neurological deficit","present":false},{"flag":"Fever with back pain","present":false},{"flag":"Saddle anesthesia","present":false},{"flag":"Bowel dysfunction","present":false},{"flag":"Bladder dysfunction","present":false},{"flag":"Thunderclap headache onset","present":false},{"flag":"Worst headache of life","present":false},{"flag":"Immunocompromised state","present":false},{"flag":"Anticoagulation therapy","present":false},{"flag":"Trauma","present":false}]
Reconstructed
[{"flag":"Cancer history","present":false},{"flag":"Neurological deficit","present":false},{"flag":"Fever with back pain","present":false},{"flag":"Saddle anesthesia","present":false},{"flag":"Bowel dysfunction","present":false},{"flag":"Bladder dysfunction","present":false},{"flag":"Thunderclap headache onset","present":false},{"flag":"Worst headache of life","present":false},{"flag":"Immunocompromised state","present":false},{"flag":"Anticoagulation therapy","present":false},{"flag":"Trauma","present":false},{"flag":"Fever","present":true}]
elements: a6a6f77630716ff3
symptoms
Typed
—
Reconstructed
Low Back Pain, Back Pain
elements: 112515b883be40fe, b6686ca8eb617312, fc682f7082be61f3

Typed rating

usually appropriate

Chart rating

usually appropriate

Δ +3

Documentation Concordance Score

13.6%

1 − completeness gap (charted decision-relevant weight missing from the typed indication). Hover for the published definition.

Documentation Concordance Score

Where asserted indications and charted evidence part company

By service line, with n and a confidence interval. Completeness gap and discordance rate are reported separately — they are different failure modes with different remedies (JAMIA 2018 anchor).

Who this is for: Associate CMO and compliance — the documentation buyer Baugh named.
What it costs them today: Without a cohort number, documentation gaps stay anecdotal and never become a service-line program.
What ARKA returns: Where asserted indications and charted evidence part company, by service line, with n and a confidence interval — and small-cell suppression you can show legal.

Sample report · redacted · synthetic extract

Baseline Capture — Documentation Concordance

Illustrative sample output from the offline baseline kit. Not a customer result.

IllustrativeIllustrative — synthetic cohort kit sample-output

16.7%

Mean DCS (n = 63)

Illustrative95% CI 13.7% – 19.9% (bootstrap)

0.0%

Discordance rate

IllustrativeWilson 95% CI 0.0% – 5.7%

mechanism

Element most often missing from requisitions (100.0% of orders)

Illustrative63 of 63 orders in sample

n < 11

Small-cell suppression threshold

MeasuredEnforced rule: strata with n < 11 never appear in public tables

Mean DCS by service line (unsuppressed strata only)

Service line	n	Mean DCS	95% CI
Emergency	28	14.0%	0.10–0.19
Medicine	22	18.0%	0.13–0.24
Surgery	13	21.0%	0.14–0.29

Small-cell suppression is active and visible

Any stratum with n < 11 is suppressed from clinician and service-line tables and listed for auditability — never rendered as a rate.

service_line: pilot_lab — n = 4 < 11 (small-cell suppression)
clinician_hash: 03424174937a3431 — n = 1 < 11 (small-cell suppression)

For radiologists

The reconstructed indication ships into the read

You open a study and the clinical question, the relevant labs, the priors, the medications and the pathology are already assembled, with source links. You stop reading the chart by hand.

Who this is for: Radiologists who currently rebuild the chart by hand before every complex read.
What it costs them today: Minutes of chart archaeology per study — priors, labs, meds, pathology — for a two-word indication that rarely helps.
What ARKA returns: The reconstructed indication ships into the read: clinical question, labs, priors, meds, pathology, source links. You stop reading the chart by hand.

Clinical question first
The diagnostic question is derived from the order plus the highest-weight charted elements — not from the requisition free text alone.
Labs and vitals that change the frame
WBC, lactate, eGFR, anticoagulation — the same signals a careful radiologist would dig for, already linked to Observation resources.
Priors and pathology
Prior radiology and pathology reports are narrative Phase-B classes. They surface with dates and source DocumentReference / DiagnosticReport ids.
Clickable grounding
Every chip is a receipt. A false or missing element cannot hide behind a fluent paragraph — it either has a span or it is not in the indication.

Honest status

What is built, and what is not

A pre-revenue company that publishes its own build status survives a technical reference call. One that implies everything ships today does not.

Who this is for: Technical reference callers who will hang up the moment a demo overclaims.
What it costs them today: Implied 'we ship everything today' fails the first deep diligence question.
What ARKA returns: A published build ledger — built, in build, not complete — with links to protocol and evaluation methodology.

Built
Appropriateness engine (AIIE)
Deterministic scenario-level rating with evidence links and audit trail.
Built
CDS Hooks / SMART / Da Vinci integration
order-select, patient-view pre-warm, SMART launch, PA packet emission paths.
Built
Auto-approve-only architecture & federated analytics
Approve path only; analytics de-identified. Denials stay human.
In build
Indication reconstruction
In build against CT abdomen/pelvis; seeking one design-partner site for shadow evaluation.
Not complete
Prospective clinical validation
Protocol and statistical analysis plan drafted; enrollment and outcomes not claimed as complete.

Next step

Measure your baseline — free.

Measure your baseline — free Book a live walkthrough

The order says “trauma, fall.” The chart says a great deal more.

The evidence that this is a real problem

How it works

Retrieve

Extract

Normalise

Assemble

Grade

Emit

The nine resource classes

Provenance

What arrives.

What the chart already contains.

What changes.

The firewall

Permitted

Forbidden

You can't type your way past the chart

What arrives.

What the chart already contains.

What changes.

Where asserted indications and charted evidence part company

Mean DCS by service line (unsuppressed strata only)

The reconstructed indication ships into the read

Clinical question first

Labs and vitals that change the frame

Priors and pathology

Clickable grounding

What is built, and what is not

Appropriateness engine (AIIE)

CDS Hooks / SMART / Da Vinci integration

Auto-approve-only architecture & federated analytics

Indication reconstruction

Prospective clinical validation

Measure your baseline — free.

The order says “trauma, fall.” The chart says a great deal more.

The evidence that this is a real problem

How it works

Retrieve

Extract

Normalise

Assemble

Grade

Emit

The nine resource classes

Provenance

What arrives.

What the chart already contains.

What changes.

The firewall

Permitted

Forbidden

You can't type your way past the chart

What arrives.

What the chart already contains.

What changes.

Where asserted indications and charted evidence part company

Mean DCS by service line (unsuppressed strata only)

The reconstructed indication ships into the read

Clinical question first

Labs and vitals that change the frame

Priors and pathology

Clickable grounding

What is built, and what is not

Appropriateness engine (AIIE)

CDS Hooks / SMART / Da Vinci integration

Auto-approve-only architecture & federated analytics

Indication reconstruction

Prospective clinical validation

Measure your baseline — free.