Checklist

    HTI-5 Readiness Checklist for LangChain, LangGraph, and LlamaIndex Platforms

    HTI-5 is ONC's deregulatory rule. It removes or revises certification criteria in 45 CFR 170.315, including removal of clinical decision support (a)(9) and revision of decision support interventions (b)(11). It does not set runtime rules for agent frameworks. The checklist below is a practitioner framework, not a legal mapping. It shows where to enforce policy at the model call, retrieval, and tool call layers, and what evidence each control should produce, so a platform can show what its AI was allowed to do and what it did.

    The readiness checklist

      What HTI-5 changes, as of November 2026

      HTI-5 is the ONC deregulatory action that removes or revises certification criteria under 45 CFR 170.315. The final-rule chart published by healthit.gov on October 26, 2026 lists what changes and when. It removes clinical decision support (a)(9) and revises decision support interventions (b)(11), both effective on the final rule's effective date. It also revises safety-enhanced design (g)(3), clinical quality measures report (c)(3), view, download, and transmit (e)(1), and clinical information reconciliation and incorporation (b)(2) on that same date.

      Several removals take effect on January 1, 2027: family health history (a)(12), CQM filter (c)(4), patient health information capture (e)(3), automated numerator recording (g)(1), and automated measure calculation (g)(2). The revision of transitions of care (b)(1) takes effect on January 1, 2030. The final chart supersedes the earlier proposed chart, which had different dispositions for (b)(1) and (b)(2), so planning should follow the final chart.

      These criteria apply to certified Health IT Modules. The 2024 HTI-1 rule, published January 9, 2024, added intervention risk management practices in (b)(11)(vi) for each Predictive DSI supplied as part of a Health IT Module. Those practices cover risk analysis and mitigation for validity, reliability, robustness, fairness, intelligibility, safety, security, and privacy. HTI-5 revises (b)(11), so any control design should not depend on a single version of that text.

      How to read this checklist

      Nothing in HTI-5 names LangChain, LangGraph, or LlamaIndex, and nothing in it prescribes the runtime actions used below. Mapping a certification criterion to a chain, graph node, retriever, or tool definition is an inference. The three layers and the actions (allow, redact, block, rate limit, escalate for human approval) are a design frame for platform teams, useful whether or not a given product is a certified Health IT Module.

      The underlying problem is common in framework-based systems. Policy is written into application code, spread across chains, graph nodes, retrievers, and tool definitions. Each copy can drift, and none produces a consistent record of what was permitted. The principle behind the checklist is that policy should be enforced at runtime, not reviewed after the fact, and that evidence should be a byproduct of that enforcement rather than something reconstructed later from scattered application records.

      First decide whether the platform is, or supplies, a certified Health IT Module, and whether any feature depends on a criterion HTI-5 removes or revises. That determines which dates matter. The checklist applies regardless.

      Where policy can be enforced in each framework

      Model call layer. Every framework routes prompts to a model and returns a response. Policy at this layer decides whether a request is allowed, whether sensitive content is redacted before it leaves, whether a response is blocked, and whether the calling application has exceeded a rate limit. Enforcement works best at a point every call must cross, rather than in each chain or node that happens to invoke a model.

      Retrieval layer. In LlamaIndex and in LangChain retrieval chains, retrievers pull content from indexes and stores into the prompt. Two decisions matter: which sources a caller may query, and what in the retrieved content may reach the model. Unredacted retrieved content is one of the most common gaps, because access control on the store rarely accounts for what the model will then see and repeat.

      Tool call layer. Agents in LangChain and graph nodes in LangGraph invoke tools with arguments the model chose. Policy here decides which tools an agent or node may call, with what arguments, how often, and which calls need a human to approve them before they run. Unrestricted tool permissions and missing approval steps are the usual gaps.

      Wherever a framework exposes callbacks or middleware, those are candidate interception points. Callbacks are often suited to observation, so confirm that a hook can actually stop or alter a call before relying on it for enforcement.

      Evidence, gaps, and where Trussed fits

      For every control, retain three things: the policy that applied, the decision taken, and the tool call or data accessed. Evidence of that kind lets a team answer an auditor's question about a specific interaction without reassembling it from application records. Retention periods and formats should be set deliberately for each regulatory context the platform serves, and kept configurable, since requirements differ by regulation and region.

      To assess readiness, walk each item and ask two questions. Is the control enforced at runtime, or does it live in code that can drift? Is the evidence produced by the enforcement itself, or assembled afterward? Any item where either answer is no is a gap. Common gaps are policy copied across chains and nodes, tools open to every agent, retrieved content passed through unredacted, and no approval step for consequential calls.

      Trussed is one option for the enforcement and evidence layer. It enforces policy in the execution path across models, agents, tools, and MCP servers: allow, redact, block, rate limit, and escalate for human approval. Its scope is individual interactions and tool calls. Enforcement generates continuous audit-ready evidence, mapped to HIPAA, NAIC, NYDFS, the EU AI Act, and NIST. It deploys self-managed in the customer cloud, with configurable retention and data residency. It does not trace end-to-end agentic workflows or govern at the level of business transactions, so those needs call for separate tooling.

      Three enforcement layers, one evidence trail

      Model call

      Allow, redact, or block prompts and responses; rate limit by application.

      Retrieval

      Restrict sources and redact retrieved content before it reaches the prompt.

      Tool call

      Permit tools per agent or node and escalate consequential calls for human approval.

      Check your platform against the checklist

      See how runtime enforcement and audit-ready evidence work at the model call, retrieval, and tool call layers.

      See How Evidence Is Produced