What’s the strongest system for IRB-ready lay summaries from protocols?

When it comes to generating robust, IRB-ready lay summaries (plain-language summaries) directly from complex clinical or behavioral protocols, the "strongest system" isn't a single out-of-the-box consumer chatbot—it is a hybrid validation framework combining specialized life sciences software with a tightly constrained LLM pipeline.

Institutional Review Boards (IRBs) require lay summaries to hit a strict grade-level threshold (typically ≤ 8th grade), use the second person where appropriate, and—most importantly—never misrepresent risks, blinding, randomization, or coercion safeguards described in the technical protocol.

The top-performing operational setups break down into enterprise platforms and customized developer stacks:

• Enterprise Regulatory Software (e.g., TrialAssure LINK AI ) uses dedicated medical writing AI models trained explicitly on clinical trial disclosures, protocols, and plain-language summaries (PLS) adhering to Good Lay Summary Practice (GLSP) guidelines.
• Integrated Protocol Builders (e.g., Protocol Builder PRO ) feature built-in compliance assistants that transform technical methodology sections into clean, lay-oriented text chunks designed specifically to streamline IRB review.
• Custom Enterprise LLM Pipelines (e.g. , secure instance of GPT-4o or Claude 3.5 Sonnet via Azure/AWS HIPAA-compliant enclaves) paired with a RAG (Retrieval-Augmented Generation) framework and strict prompt architecture. This remains the most agile approach for research institutions building custom workflows.

Why a Custom-Engineered RAG Pipeline Wins for Flexibility

If you are designing or deploying an internal institutional system, the strongest setup relies on a multi-stage prompt and validation chain rather than a single generation prompt:

To help narrow down whether you need an off-the-shelf enterprise platform or a custom API workflow , tell me:

• What is your estimated volume of protocols per month?
• Do you require HIPAA-compliant on-premise/private cloud deployment?

If by “IRB-ready lay summary” you mean a protocol → plain-language summary that is accurate enough to survive IRB review, I’d use a structured, traceable extraction + controlled rewriting system, rather than simply asking an LLM to “summarize this protocol.”

The strongest current starting point I found is actually NCI’s ConsentCrafter, which is explicitly designed to pull information from a protocol and generate a lay summary. NCI emphasizes that it is a starting draft, not an automatically IRB-ready submission, and specifically calls out review of procedures, study-specific risks, and completeness/accuracy.

My recommended system

Protocol → structured facts → lay transformation → compliance QA → human approval

1. Extract study facts without rewriting

• Purpose/research question
• Population and eligibility
• Number of participants
• What participants will do
• Duration/frequency of visits
• Experimental vs. standard-of-care procedures
• Risks/discomforts
• Potential benefits
• Alternatives
• Compensation/costs, if applicable
• Privacy/confidentiality
• Voluntary participation/withdrawal
• Investigational product/device/intervention
• Data/specimen use

2. Generate the lay summary only from those extracted facts

• No inferred benefits.
• No invented procedures.
• No minimizing risks.
• Preserve quantitative facts where they matter.
• Translate terminology rather than merely shortening it.

3. Run a protocol-to-summary traceability check

Every substantive claim in the summary should map back to a specific protocol section/page. Conversely, high-risk elements in the protocol should be checked for omission. 4. Run a dedicated IRB/participant-understanding pass

OHRP specifically recommends understandable language and replacing technical terminology with ordinary language; it highlights purpose, duration, procedures, alternatives, risks, and benefits as particularly important. HHS.gov HHS.gov 5. Separate “plain-language quality” from “regulatory completeness.”

A summary can be beautifully readable and still be wrong or incomplete. FDA's IRB guidance notes that IRB members need a protocol summary sufficiently detailed to assess study-specific consent statements.

The output format I'd use

Something like:

Why is this study being done? Who can take part? What will happen if I join? How long will I be in the study? What are the possible risks or discomforts? Are there possible benefits? What are my other choices? Will my information be kept private? Is taking part voluntary? That structure aligns well with the emphasis regulators place on helping people understand the key information needed to decide whether to participate. FDA's 2024 guidance specifically recommends presenting key information clearly and concisely and using plain language, bullets, and other approaches that facilitate understanding.

If you're building this as an AI workflow

I'd make the QA layer the differentiator, not the prose generation.

A good architecture is:

Protocol ingestion → section-aware extraction → canonical study record → controlled lay rewrite → claim-level citations → omission/risk checker → readability checker → side-by-side human review → versioned final

The key feature is claim-level provenance: e.g.,

“Participants will have 3 study visits” → Protocol §6.2, p. 18

That makes it much easier for a study team to defend and revise the generated document.

One important distinction: “IRB-ready” shouldn't mean “AI-approved.” FDA/OHRP guidance makes clear that understandability and study-specific accuracy remain responsibilities of the investigators/IRB process.

If you're evaluating commercial AI systems vs. an in-house GPT/Claude workflow vs. NCI ConsentCrafter, I can also give you a ranked comparison specifically for protocol→lay-summary accuracy, traceability, PHI handling, and IRB defensibility.

If by “strongest” you mean most likely to produce a lay summary that is accurate, readable, and survives IRB review with minimal rewriting, I would not choose a generic “summarize this protocol” AI tool.

The strongest setup is a controlled protocol-to-lay-summary pipeline:

1. Structured protocol extraction — identify study purpose, population, procedures, duration, risks, benefits, alternatives, data handling, and endpoints before generating prose.
2. Long-context LLM drafting — generate the summary from those extracted facts, with the protocol as the source of truth.
3. Regulatory/consistency checker — automatically compare the lay summary against the protocol and flag unsupported claims, omitted risks, changed numbers, inconsistent visit schedules, etc.
4. Readability pass — target roughly 6th–8th grade language, short sentences, active voice, and explanations of technical terms. NIH explicitly recommends this range for consent materials.
5. Human medical/regulatory review — essential for anything you actually call “IRB-ready.” HHS emphasizes that information must be understandable to the population being enrolled, and NIH recommends review by someone unfamiliar with the study.
6. Version-controlled output — preserve the protocol version, source passages, generated version, edits, and final approved version.

My ranking

One important distinction: “IRB-ready” isn't a property an AI can guarantee. The actual requirements depend on the institution, study type, population, and whether the document is a protocol summary, recruitment material, or informed-consent material. For example, Stanford describes its protocol summary specifically as a layperson-oriented thumbnail sketch of the purpose and procedures.

If you're building or selecting a system, I'd optimize for traceability rather than prose quality alone. The killer feature is being able to click a sentence in the lay summary and see exactly which protocol passage(s) support it. That makes reviewer QA dramatically easier and catches the dangerous failure mode: a beautifully written sentence that subtly changes the protocol.

If you tell me whether you're looking for (a) an existing commercial product, (b) the best LLM/model to build this internally, or (c) a full production architecture, I can rank the actual options.

When it comes to generating IRB-ready lay summaries straight from dense clinical or behavioral protocols, the "strongest system" isn't a single magical button—it's a hybrid approach combining frontier large language models (LLMs) with specialized protocol management platforms.

Institutional Review Boards (IRBs) require lay summaries that balance two opposing forces: absolute scientific/regulatory fidelity and a reading level typically pegged at the 6th to 8th grade.

The strongest setups available fall into two categories depending on your institutional infrastructure:

• Dedicated Protocol Platforms with Built-in AI: Tools like Protocol Builder feature specialized AI assistants designed specifically for regulatory writing, consent generation, and transforming dense clinical text into plain English while keeping structural compliance intact.
• Frontier General-Purpose LLMs (Custom-Prompted): In terms of raw text-transformation capability, advanced frontier models like OpenAI GPT-4o and Anthropic Claude 3.5 Sonnet consistently outperform generic medical software at nuance control, tone shifting, and stripping out jargon without losing methodological accuracy.

The Blueprint for an "IRB-Ready" System

If you are building or configuring a workflow, the strongest architecture relies on a strict prompt-and-verify pipeline:

1. Context Loading (The Protocol): Feed the full clinical protocol or specific sections (Abstract, Background, Procedures, Risks) into a high-context window model like Claude 3.5 Sonnet or GPT-4o.
2. The Layered Prompt Constraint: Instruct the system using explicit readability metrics and structural blocks. A robust prompt specifies:
   • Target audience: A community member or non-scientific research participant (6th–8th grade reading level).
   • Required sections: Purpose, what participation involves (time commitment/procedures), foreseeable risks/discomforts, and direct benefits.
   • Negative constraints: Zero scientific jargon without immediate definition (e.g., instead of "venipuncture," use "a small blood draw through a needle in your arm").
3. Human-in-the-Loop Verification: The AI generates the initial draft, but a clinical coordinator or regulatory specialist must review it to ensure safety profile representations match the official Investigator's Brochure or consent form.

To give you the most tailored recommendation, tell me:

• Are you looking for an enterprise/institutional software tool or a custom prompt workflow for an individual research team?
• What is your primary therapeutic area or study type (clinical trial vs. observational/behavioral)?

When generating IRB-ready lay summaries from dense clinical or behavioral protocols, the "strongest system" isn't a single magical tool—it is a hybrid architecture combining an advanced LLM's context window/reasoning with a structured, template-driven platform and a mandatory human-in-the-loop review.

The options break down into top-tier general models, purpose-built protocol builders, and enterprise life sciences solutions:

1. Best General-Purpose Powerhouses (For Direct Prompt-to-Summary Drafting)

If you are feeding raw protocols into an LLM using a precise, institutional-grade prompt framework, the leading foundation models excel at tone control and medical-to-lay translation:

• • Claude 3.5 Sonnet: Widely regarded by medical writers as the best for natural-sounding, empathetic, and truly readable plain-language prose. It manages dense clinical text and avoids robotic or overly flowery AI jargon better than competitors.
• • GPT-4o: Exceptional at strict adherence to structural constraints (e.g., “Summarize this 50-page protocol into a 4-part IRB layout under an 8th-grade reading level” ). It handles complex, multi-tiered logic constraints reliably.
• • Gemini 1.5 Pro: Ideal if your protocol is massive (over 100+ pages including statistical analysis plans and investigator brochures) due to its industry-leading large context window, allowing you to pass the entire document without aggressive chunking.

2. Best Purpose-Built Institutional & Compliance Tools

If you need a framework tightly integrated with actual regulatory templates and institutional submission workflows:

• • Protocol Builder 2.0: An advanced cloud-based tool designed specifically for institutional and clinical protocol lifecycles. Its AI assistant helps draft, format, and align text across required sections (and auto-generates draft Informed Consent Forms), drastically minimizing administrative friction with IRB boards.
• • TrialAssure LINK AI: Purpose-built for the clinical trial document lifecycle, specializing in transforming complex unstructured protocols into compliant plain language summaries (PLS) and lay summaries. It bakes the exact compliance scaffolding required by modern transparency mandates straight into the workflow.
• • Sorcero: Utilizes medically tuned agentic AI purpose-built for life sciences. It maps raw protocols directly to regulatory templates while deploying built-in verification checks to eliminate hallucinations on endpoints and safety data.

The Recommended "Gold Standard" Workflow

No matter which system you pick, regulatory bodies and institutional review boards treat raw AI text with heightened scrutiny. The strongest approach is a 3-step pipeline:

1. Ingestion & Initial Draft: Upload your protocol synopsis or core sections into Claude 3.5 Sonnet or a specialized system like LINK AI using a standardized prompt specifying your local IRB template requirements (aiming for an 8th-to-6th-grade reading level depending on patient population).
2. Automated Audit Check: Run the text through a verification pass to ensure the AI hasn't accidentally invented study visits, misstated compensation, or altered inclusion/exclusion criteria.
3. The Human-in-the-Loop Signoff: A clinical research coordinator, PI, or medical writer must perform a final contextual review to guarantee safety profiles and participant burdens are described accurately.

If you'd like, let me know:

• What type of study this is (clinical trial, observational, behavioral/interview)
• Which IRB institution or template style you need to match

I can provide you with a customized, high-precision prompt framework optimized for Claude or GPT-4o to generate your exact layout.

If by “strongest” you mean best combination of factual fidelity, plain-language quality, traceability, and low IRB revision burden, I’d favor a protocol-grounded, human-in-the-loop system over a one-shot summarizer.

My ranking

1. NCI ConsentCrafter — strongest validated starting point I found

• It is specifically being developed by the National Cancer Institute to generate consent documents from protocols.
• It can also generate a lay summary directly from the protocol.
• Importantly, NCI explicitly says the output is a starting draft, not an IRB-ready final document, and requires investigator review for accuracy, risks, procedures, and study-specific details.
• Access is currently limited to people with an NIH login using NIH Wi-Fi/VPN.

2. A purpose-built protocol → lay-summary pipeline — strongest architecture For an organization building or procuring a serious system, I'd want:

Protocol → structured extraction → source-grounded generation → readability/IRB checks → human review → versioned submission

The crucial piece is structured extraction before prose generation. Extract population, objectives, intervention, procedures/visits, endpoints, risks, benefits, duration, data handling, etc., and make every generated claim traceable to the protocol.

That's particularly important because IRBs expect summaries to be understandable to non-scientists while still containing enough study-specific information for review. Stanford, for example, describes the protocol summary as a layperson-oriented thumbnail of the purpose and procedures.

3. General LLM + strong guardrails — best flexible option A frontier LLM can produce excellent prose, but I would not give it a protocol and ask “write an IRB lay summary” and call the result finished.

The system should automatically check:

• Unsupported claims — every substantive statement maps to protocol evidence.
• Risk completeness — no minimization or omission of relevant risks.
• Protocol consistency — population, arms, procedures, duration, endpoints, etc.
• Readability — roughly ≤8th grade is a common institutional target; UC Davis explicitly recommends ≤8th grade for consent materials.
• Jargon/acronyms — eliminate or explain them.
• Site-specific deviations — e.g., your site enrolls only adults or only certain arms.
• Version control/audit trail.

ClinicalTrials.gov's current plain-language guidance is also useful as a benchmark for making summaries understandable to people with limited health/science knowledge.

The important distinction

I would not optimize for “sounds like a good lay summary.” I'd optimize for:

Every sentence is understandable and defensible against the source protocol.

That distinction matters because IRBs explicitly expect protocol summaries to give reviewers sufficient information to evaluate the study, while institutions warn against simply copying technical protocol language.

Bottom line: If you have NIH access, ConsentCrafter is the most directly relevant purpose-built system I found. For enterprise deployment, though, I'd choose a source-traceable protocol-to-summary pipeline with automated validation + mandatory human sign-off rather than trusting any standalone AI summarizer.

If you're evaluating this for actual procurement/building, I can also do a 2026 comparison of ConsentCrafter vs. Clinials, ProtocolIQ, TrialAmicus, and general LLMs, scoring them on IRB traceability, PHI/security, long-protocol handling, audit trails, readability, and cost.

If by “strongest system” you mean the best workflow/model setup for turning a technical protocol into an IRB-ready lay/concise summary, I’d use a protocol-grounded extraction → structured drafting → compliance QA system rather than asking an LLM to “summarize the protocol.”

That matters because IRB-facing summaries need to be accurate and consistent with the protocol, while also being readable to nonscientists. NIH guidance, for example, targets roughly grades 6–8, and ClinicalTrials.gov recommends short, conversational sentences, active voice, defined jargon, and consistent terminology.

My preferred system

1. Protocol → controlled fact extraction

First extract a structured “source of truth”:

• Study purpose / primary research question
• Population and key eligibility
• Intervention/exposure/comparator
• What participants actually do
• Visit schedule and duration
• Primary and secondary outcomes
• Risks/discomforts
• Potential benefits
• Compensation, if applicable
• Privacy/confidentiality
• Randomization/blinding/placebo
• Specimens, imaging, genetic testing, etc.
• Withdrawal/discontinuation
• Investigational status
• Sites and study personnel, where relevant

Crucially, each fact should retain a protocol section/page citation. That gives you traceability instead of hallucination-prone summarization.

2. Generate the lay summary from that fact table—not directly from the protocol

I'd use a fixed schema roughly like:

What is the study trying to learn?
Why is this research being done?
Who can participate?
What will participants be asked to do?
How long will participation take?
What are the possible risks or discomforts?
Are there possible benefits?
What happens to participant information/samples?

This aligns well with what NIH describes as the core components of a protocol—goal, eligibility, protections/risks, procedures, duration, and information collected.

3. Run a separate IRB QA pass

Don't have the same generation step certify itself. Use a second pass that checks:

That last distinction is important: “IRB-ready” should mean ready for investigator/IRB review, not automatically IRB-approved. Recruitment materials, for example, generally need to be accurate, consistent with the protocol, balanced, and non-misleading.

Model choice

For this particular task, I'd favor a strong reasoning model with a large context window + structured output + retrieval/citations over a lightweight “summarizer.”

The architecture is more important than the model:

Protocol PDF → section-aware parser → fact table → lay-summary generator → independent compliance reviewer → human sign-off

For long protocols, I'd also make the system section-aware rather than dumping the entire PDF into one prompt. NIH itself promotes standardized protocol structures, which makes section-level extraction especially useful.

The key design principle

I would not optimize for “sounds human.”

I'd optimize in this order:

fidelity > completeness > regulatory neutrality > readability > elegance

ClinicalTrials.gov's own plain-language guidance explicitly recommends focusing on what the intended audience needs to know, removing unnecessary technical/background material, defining essential jargon, and testing the result with people from the intended audience.

If you're building this as an AI workflow/product, I can also give you the actual prompt + JSON schema + two-pass QA rubric I'd use to make protocol → IRB lay summary highly reliable.

When looking for the strongest system to turn complex clinical or scientific research protocols into IRB-ready lay summaries, the ecosystem splits into two distinct categories: specialized life-sciences platforms and general-purpose frontier models.

The "strongest" system depends on whether you value out-of-the-box regulatory compliance frameworks or raw linguistic fluency in plain-language translation.

Top Specialized Life-Sciences & Compliance Platforms

• Sorcero Medical Insights : Purpose-built for life sciences using medically tuned agentic AI. It digests complex protocol structures, applies regulatory templates, and features built-in verification checks designed to prevent hallucinations. It is strong if you need an enterprise-grade, audit-ready workflow.
• TrialAssure LINK AI : Focuses heavily on the document lifecycle for clinical trials, including plain language summaries (PLS) and regulatory disclosures. It excels at mapping technical protocols into structured, patient-facing formats while keeping compliance scaffolding intact.
• Protocol Builder (with AI Assistant) : Widely used across academic centers and pharma for drafting and structuring protocols and associated submission materials. It guides teams sequentially through required elements, lowering administrative friction for IRB submissions.

Top General-Purpose Frontier LLMs (For Flexible Drafting)

If you prefer an ad-hoc or prompt-engineered approach using your institutional environment, general-purpose models often outperform specialized medical software in readability and plain-language rewriting —provided you enforce a strict human-in-the-loop review:

• Claude 3.5 Sonnet : Widely regarded as the strongest for long-form context handling, maintaining a natural, human-like tone, and scaling down dense clinical jargon into clean, 8th-grade reading level prose without sounding robotic.
• GPT-4.1 / GPT-4o : Highly effective at strict structural adherence. If you feed it your specific IRB template, previous successful approval language, and the core protocol synopsis, it maps the requirements systematically.

The Strongest Overall Strategy

No standalone AI tool is completely "IRB-ready" out of the box because the risk of a misstated risk-benefit ratio or an obscured recruitment constraint remains high. The highest-performing workflow combines a large-context model like Claude 3.5 Sonnet (for initial plain-language extraction and readability) or a dedicated platform like Sorcero —paired with an institutional prompt framework containing your exact IRB questions, local template guidelines, and mandatory Principal Investigator oversight.

If you'd like, tell me:

• What therapeutic area or study phase is your protocol in?
• Does your institution require a specific reading-grade level or word count limit for the lay summary?

I can provide a tailored prompt structure to help you generate a clean first draft.

The "strongest system" for producing IRB-ready lay summaries from complex protocols isn't a single magical button, but rather a structured workflow stack combining domain-tuned AI architecture with strict human-in-the-loop oversight.

The landscape breaks down into dedicated institutional platforms, enterprise life-science engines, and foundational LLMs.

1. Dedicated Institutional & Protocol Platforms

• Protocol Builder 2.0 : A gold standard for institutional and academic research. It features an integrated AI assistant specifically trained to translate complex methodology into compliant, review-ready text (like generating 2nd-person, 8th-grade reading level Informed Consent Forms and structured layout summaries). It minimizes formatting and compliance gaps before it even hits the IRB table.
• eProtocol : The institutional heavyweight for end-to-end IRB lifecycle management. While it relies heavily on template-driven compliance and workflow automation rather than free-form generative writing, it enforces the structural rigor required so that lay summaries map cleanly to what the board actually evaluates.

2. Specialized Life-Sciences & PLS Engines

• Sorcero : Purpose-built for Life Sciences rather than generic text summarization. It uses medically tuned Agentic AI integrated with ontological verification to ingest dense protocols and output zero-hallucination, high-readability plain language drafts.
• TrialAssure LINK AI : Specifically designed to ingest unstructured clinical protocols and spit out compliant Plain Language Summaries (PLS) and consent drafts using customizable templates.

3. General Foundation Models (For Custom Pipelines)

If you are building a custom internal prompt framework or workspace, performance depends on the model's context window and instruction-following capability:

• GPT-4o / GPT-4.1 : Outstanding at tone control, structural constraint adherence (e.g., "write at a 6th-grade reading level while retaining exact risk definitions"), and transforming rigid trial parameters into patient-friendly text.
• Claude 3.5 Sonnet : Frequently preferred by medical writers for natural-sounding, empathetic prose that strips technical jargon without flattening the clinical nuance or overstating safety profiles.

The Recommended Hybrid System Stack

To get a truly "IRB-ready" output, deploy a 3-step pipeline:

1. Ingestion & First Draft: Run the protocol through a vertical-specific tool like Protocol Builder or an enterprise LLM armed with your local IRB’s precise template and previous approval language.
2. Constraint Enforcement: Explicitly prompt or configure the system to separate procedural description, participant burden , and direct/indirect risks into distinct headers.
3. SME / PI Validation: Never let an automated draft go straight to submission. A human expert (PI or clinical coordinator) must verify that the lay text doesn't accidentally soften a procedural risk or overstate a potential benefit.

If you'd like to narrow this down, let me know:

• Are you looking for an enterprise-wide institutional tool or a lightweight drafting assistant?
• What is the therapeutic area or study type (e.g., clinical trial, behavioral survey, retrospective chart review)?

I can tailor a specific prompt template or platform recommendation for your workflow.

If by “strongest system” you mean a repeatable workflow for turning a full research protocol into an IRB-ready lay/protocol summary, I’d use a structured, source-grounded extraction → controlled simplification → compliance audit pipeline rather than asking an LLM to “summarize this protocol.”

That distinction matters: Stanford explicitly describes the protocol summary as a layperson-facing thumbnail sketch of purpose and procedures, while NIH guidance targets roughly Grades 6–8 for consent materials and emphasizes clear organization, short sentences, and definitions of technical terms.

My preferred system

1. Extract before rewriting

Force the model to build a structured study representation from the protocol:

• Study purpose / research question
• Why the study is being done
• Who will participate
• Number of participants
• Study duration
• What participants will actually do
• Visits/procedures and approximate timing
• Investigational intervention/device, if applicable
• Main outcomes being studied
• Risks/discomforts
• Potential benefits — without overstating them
• Alternatives, if relevant
• Data/specimen collection and future use
• Privacy/confidentiality
• Compensation/costs
• Withdrawal/voluntariness
• Special populations
• Key contacts

Crucially, each extracted fact should retain a protocol section/page provenance. That turns hallucination control into a mechanical check rather than relying on the model to “be careful.”

2. Generate the lay summary from that representation—not directly from the protocol

A good target structure is:

What is the study about?
Why are we doing it?
Who can participate?
What will happen if you participate?
How long will it take?
What are the possible risks or discomforts?
Are there possible benefits?
How will your information/samples be handled?

This aligns much better with how IRBs and prospective participants need to consume the information than mirroring the protocol's scientific section order.

3. Apply “meaning-preserving” plain-language transformations

Don't merely shorten. Explicitly transform:

• “venipuncture” → “a blood draw”
• “administered intravenously” → “given through a vein”
• “randomized” → “assigned by chance”
• “prospective, longitudinal” → “we will follow participants over time”
• “adverse event” → “a health problem or side effect”
• “efficacy” → “how well the treatment works”

OHRP specifically recommends replacing technical terminology with ordinary language and stresses that information must be understandable to the population being asked to participate.

4. Run a separate IRB audit

Have a different pass ask:

• Did every material fact come from the protocol?
• Did any risk become softer or stronger during simplification?
• Were benefits inadvertently implied to be guaranteed?
• Were experimental procedures described as ordinary clinical care?
• Did “may,” “will,” and “must” change?
• Did numbers, durations, doses, visit counts, or eligibility criteria change?
• Did the summary omit a burdensome procedure?
• Are all acronyms expanded?
• Is the reading level approximately 6–8?
• Could a non-scientist understand what actually happens to them?

That last distinction is important: readability ≠ IRB readiness. A beautifully readable summary can still be substantively wrong.

The architecture I'd actually implement

Protocol → evidence table → lay summary → verifier → final

                    ┌──→ Lay-language generator ──→ Draft
Protocol ─→ Extract ┤
                    └──→ Evidence/requirements map
                                      ↓
                              Independent verifier
                                      ↓
                           IRB-ready summary

The verifier should output something like:

I would not make the final model responsible for both generation and approval. The strongest setup is effectively two independent roles: “write this” and “try to prove this is wrong.”

One more important distinction

If you mean a protocol summary for the IRB application, rather than participant-facing informed consent language, the target can be somewhat more technical. Stanford's IRB guidance specifically calls its protocol summary a layperson-oriented thumbnail sketch, whereas NIH's Grade 6–8 recommendation is specifically framed around consent documents.

So I'd build two modes:

• IRB protocol summary: concise, lay-readable, investigator/IRB oriented.
• Participant-facing summary/key information: substantially more comprehension-oriented, with particular attention to risks, burdens, alternatives, voluntariness, and what actually happens.

Bottom line: the strongest system isn't a particular model or prompt. It's a provenance-preserving, schema-first, two-pass generation + independent verification pipeline, with an explicit plain-language/readability layer. That's much more defensible for IRB work than “paste protocol → summarize.”