Clinical Decision Support and AI Transparency

The Platform is clinical and financial decision support software. It reads documents a facility already produces. These may include referral packets, MDS assessments, charts, orders, and claims. The software returns a structured, scored result in a fixed schema, with each finding quoted back to the page and section it came from.

It is not an electronic health record, not certified health IT under 45 C.F.R. Part 170, not a diagnostic device, and not a determinant of medical necessity or coverage.

2.1FDA

Orenda designs and offers the Platform to fall within the clinical decision support exclusion at section 520(o)(1)(E) of the Federal Food, Drug, and Cosmetic Act, 21 U.S.C. § 360j(o)(1)(E), as interpreted in FDA’s guidance Clinical Decision Support Software (Sept. 28, 2022). Four design properties support that position:

• the Platform does not acquire, process, or analyze a signal from an in vitro diagnostic or a signal acquisition system;
• it displays, analyzes, and prints medical information about a resident from records the facility already holds;
• it supports or provides recommendations to a health care professional about prevention, diagnosis, or treatment; and
• it is designed so that the professional does not rely primarily on it and every finding travels with its source quote, page, and section, so the basis of the recommendation is independently reviewable.

The Platform has not been cleared, approved, or authorized as a medical device. Customer’s use of it in a manner inconsistent with Section 5.1 of the Agreement — for time-critical decisions, as the primary basis for a clinical determination, or in a way that forecloses independent review of the cited source — can defeat the exclusion, and Customer will not do so.

2.2California AB 3030

California Health and Safety Code § 1339.75, added by AB 3030 (2024) and effective January 1, 2025, requires a health facility, clinic, physician’s office, or group practice that uses generative AI to generate written or verbal patient communications pertaining to patient clinical information to include (a) a disclaimer that the communication was AI-generated and (b) instructions for reaching a human health care provider. Skilled nursing facilities are health facilities licensed under Health and Safety Code § 1250(c).

The obligation runs to the facility, not to Orenda. Two points matter operationally:

• The statute exempts a communication read and reviewed by a licensed or certified health care provider before it is sent. The human signature requirement in Section 5.3 of the Agreement is designed to place Clinician-Signed Documentation within that exemption.
• The exemption depends on review having actually occurred. Customer is responsible for confirming that it did, and for the disclaimers on any resident-facing communication that was not so reviewed.

The statute does not apply to communications unrelated to clinical information, such as scheduling or billing.

2.3Other jurisdictions

State AI law is unsettled and moving. Colorado’s SB 24-205 was enjoined from enforcement in April 2026 and then repealed and replaced by SB 26-189, which adopts a narrower notice-based framework for automated decision-making technology effective January 1, 2027. Texas HB 149 (TRAIGA) took effect January 1, 2026. Utah’s Artificial Intelligence Policy Act imposes disclosure duties on regulated occupations. Orenda will monitor developments applicable to the Platform and will notify Customer of a change that materially affects either Party’s obligations. Neither Party should treat this Exhibit as a fixed statement of the regulatory landscape.

3.1Inference environment

Model inference runs on Amazon Bedrock inside Orenda’s own AWS account, in United States regions, under an executed business associate agreement with AWS. Orenda does not route Customer Data to any model provider not covered by a BAA. Prompts and completions are not retained by the model provider and are not used to train, fine-tune, or evaluate any model (Agreement § 7.3).

3.2Minimization before inference

Direct identifiers are stripped before a clinical assessment is sent to the model. Orenda’s assessment tools are built to operate on age, sex, and admission date. Medical record numbers, names, and dates of birth are removed from the material the model receives, and the system prompt instructs the model not to reproduce identifiers it may nonetheless encounter in the body of an uploaded document. This is a minimization control, not de-identification under 45 C.F.R. § 164.514, and material sent to the model remains PHI and remains protected as PHI under the BAA.

3.3Document handling

Text-native documents are read complete, layout included, without chunking, so that a finding on page thirty-four is available to the same pass that reads page four. Scanned pages route through optical character recognition first. OCR introduces transcription error, and a finding drawn from an OCR’d page carries the OCR text’s fidelity, not the original’s. Orenda records which documents were read through OCR.

3.4Structured output

Output is forced to a fixed schema on every run. Nothing is parsed out of freeform prose. The schema for the Admission Fit Assessment returns: fit score (0–100); recommendation (accept, review, decline); risk factors with severity and clinical note; strengths; staffing impact (nursing hours, therapy minutes, special requirements); cited evidence (quote, page, section); and a stated confidence value, not an implied one.

3.5Scoring basis

Scoring is against the Covered Facility’s own configured Facility Rules. These include acuity ceiling, accepted payers, isolation capacity, excluded diagnoses, and staffing parameters, not a generic rubric. Facility Rules are stored as a versioned record so that the rule set that produced a given result can be identified after the fact. A score is therefore only as sound as the rules it was scored against.

Orenda states these plainly because a compliance reviewer needs them and because Section 10.4 of the Agreement disclaims any warranty that Output is accurate.

• Large language models produce output that is plausible and wrong. Fluency is not accuracy, and confidence expressed in Output is a model-generated value, not a validated calibration.
• A material finding present in a source document can be omitted. Absence of a risk factor in Output is not evidence that the risk is absent from the record.
• A citation can be misattributed. The quoted text may not appear on the page cited, or may be drawn from a different section. This is precisely why the quote, page, and section are displayed.
• Scores are not clinically validated instruments. The fit score is a configured heuristic, not a validated risk-prediction model, and has not been subjected to prospective clinical validation or published external validation.
• Performance is not uniform across document types. Handwritten, faxed, degraded, or heavily scanned material performs materially worse than text-native documents.
• The Platform has no independent knowledge of the resident. It reads what it is given. A record that is incomplete at the source produces Output that is incomplete.
• Bias. Clinical documentation reflects the practices and populations that produced it, and models trained on general corpora can reproduce disparities present in that material. Orenda has not completed a formal fairness evaluation of Output across demographic groups. Customer should not use fit scores as the sole basis for an admission decision, and should monitor its own admission and decline patterns.

5.1Material Model Change

A “Material Model Change” means (a) a change to the underlying foundation model or model version used for a licensed Module; (b) a change to a system prompt, rule set, or scoring logic that is reasonably likely to change the recommendation or score a Module produces for materially similar inputs; or (c) a change to the Output schema.

5.2Notice

Orenda will give Customer at least thirty (30) days’ advance written notice of a Material Model Change, describing the change, the Modules affected, and the expected direction and magnitude of any effect on Output. Orenda may make a change on shorter notice where required to address a security vulnerability, a safety concern, or a model provider’s deprecation, and will notify Customer as soon as practicable and explain the reason.

5.3Version identification

Orenda records the model version that produced each Output, and keeps each version of the Facility Rules, so a decision made in one quarter can be interpreted against the configuration and model in effect at the time.

5.4Regression testing

Before a Material Model Change reaches production, Orenda evaluates it against a held-out internal test set covering each licensed Module, including schema conformance, citation fidelity, and rule adherence. This is internal testing, not clinical validation. Orenda will summarize the results of the evaluation to Customer on request.

5.5Customer’s right on a Material Model Change

If a Material Model Change materially and adversely affects Customer’s use of a licensed Module, Customer may notify Orenda within sixty (60) days. Orenda will work in good faith to remediate, and if it does not do so within thirty (30) days, Customer may terminate the affected Module and receive a refund of prepaid, unused Fees for it.

6.1The signature

Section 5.3 of the Agreement is the governing control and is restated here for the compliance reader:

No clinical documentation the Platform drafts enters a resident’s medical record unless a named, licensed clinician has read it, edited or accepted it, and electronically signed it. There is no automatic posting, no facility preference that disables the signature, no trusted-tool exemption, and no background process that can produce one. Orenda will not add a setting that changes this.

6.2Evidence of review

Orenda retains the model draft together with the clinician’s edits to it. An edit is meaningful only against what it changed, and that record is evidence that review occurred. The clinician signs in the facility’s own record system, which records the signer and the time of signature. The retained draft and edits are Customer Data, exportable by Customer, and retained for six (6) years alongside audit records.

6.3What the clinician is accountable for

The model drafts; the clinician authors. Documentation entering the chart is attributed to the signer, never to Orenda and never to a service account. Signing documentation one has not read is a violation of Exhibit D § 4 and may independently implicate the signer’s professional licensure.

Compliance, clinical, and model-governance questions: admin@axisorenda.com. A reproducible material Output error should be escalated as a Severity 2 incident under Exhibit B § 4.3; where Orenda determines an error is systematic, it will notify all affected customers within ten (10) business days and describe the scope, the corrective action, and the period affected.

This page is a reference copy. Where it differs from the version incorporated into a Customer’s signed agreement, the signed version controls.