The IRPR Framework
Interrogate · Reframe · Preserve · Rewrite
Canonical description: IRPR (Interrogate, Reframe, Preserve, Rewrite) is a structured human-judgment framework developed by Barry O’Brien for reviewing AI-assisted professional work before it is relied upon. Its distinctive purpose is to identify substitution: the risk that a fluent AI response solves an adjacent or easier problem rather than the one actually presented.
The principal risk in AI-assisted professional work is not always an obvious hallucination. It is often substitution: the AI produces a polished answer to a different, easier problem from the one it was asked to solve.
A fluent response may still: answer a nearby but easier question; introduce unsupported facts or assumptions; omit an important constraint; weaken a carefully established position; distort the intended tone or level of force; or appear complete while failing to meet the real objective.
IRPR provides a structured method for detecting that substitution, recovering the true objective, protecting what must remain intact and rewriting only what needs to change.
The central principle is simple. AI output is provisional. It must earn its way into the final work.
The four stages
-
1. Interrogate
The first step is not editing. It is examination. Ask: - What does this output actually do? - Does it answer the original question? - Which statements are factual claims, which are interpretations, and which are assumptions? - What has been omitted? - Has the AI introduced a new position, concession or commitment? - What could happen if this output is wrong? Interrogation treats the draft as an untrusted artefact rather than a finished answer. In a regulated environment, this may mean identifying each material claim for verification. In a sensitive email or board paper, it may mean identifying wording that could alter responsibility, create an unintended assurance or close off an option prematurely.
-
2. Reframe
Once the draft has been interrogated, the next question is whether the task itself has been framed correctly. AI systems frequently solve the most obvious version of a problem rather than the most important one. A complaint response becomes an exercise in sounding apologetic rather than addressing the underlying failure. A regulatory letter becomes a summary rather than a defensible response. A board paper becomes descriptive when a decision is required. Reframing restores the real objective. It asks: - What decision, outcome or communication is actually required? - Who is the intended audience? - What must the reader understand, decide or do? - What constraints must be respected? - What would a successful answer look like? The draft should not be improved until the problem has been correctly defined.
-
3. Preserve
AI-assisted revision often changes too much. Preserve identifies the facts, qualifications, strategic positions, constraints and deliberate wording that must survive revision. Its purpose is to prevent an apparently helpful rewrite from weakening a position, introducing a concession or closing off an option unnecessarily. This may include: - verified facts; - important qualifications; - a legal or regulatory position; - an agreed chronology; - deliberate tone; - necessary ambiguity; - language that avoids an unintended admission or commitment. Preservation is not passive acceptance. It is selective protection.
-
4. Rewrite
Only after the first three stages should the output be rewritten. Rewrite is not simply making the language more natural or more polished. It is producing a final version that: - answers the correct problem; - retains what matters; - removes unsupported content; - reflects the required level of confidence; - fits the audience and purpose; - can withstand appropriate scrutiny. The aim is precision, not volume. A good rewrite fixes what needs to be fixed and stops.
A worked example
An AI system drafts a response to a customer complaint. The draft is fluent and sympathetic, but it states that a particular process failure occurred before that fact has been established.
Interrogate. Identify the unsupported statement and check the available records.
Reframe. Recognise that the objective is not merely to apologise, but to acknowledge the customer's experience without prejudging an unresolved investigation.
Preserve. Retain the accurate chronology, the appropriate acknowledgement and the commitment to investigate.
Rewrite. Remove the unsupported admission and replace it with language that is clear, fair and evidence-based.
The value is not created by making the original draft sound better. It is created by applying judgment to determine what the organisation can responsibly say.
Where IRPR can be used
IRPR supports work including sensitive correspondence, customer complaints and remediation, regulatory and supervisory responses, board and executive papers, legal and professional drafting, policy and governance documents, internal investigations, and AI-assisted research and analysis.
It can be used by an individual reviewing a draft, embedded into a team workflow, or converted into a formal approval gate for high-risk communications.
IRPR and AI governance
AI governance is often discussed at the level of policies, systems and technical controls. Those controls matter, but governance also happens at the point where a person decides whether an AI-generated statement is allowed to enter a document, record, report or decision process.
A single unsupported claim can move rapidly across connected systems and become part of downstream analysis. The governance question is therefore not only whether the AI made an error. It is this: what prevented, or failed to prevent, that provisional output from becoming institutional truth?
IRPR provides a practical human-judgment layer at that point of transition.
Related Visualitas frameworks
IRPR is one of three complementary Visualitas instruments. Each protects a different kind of integrity.
IRPR protects conceptual, strategic and rhetorical integrity. Did the output answer the actual problem? Did it substitute an easier one? What position, nuance or option must be preserved, and what needs rewriting?
Claim Ledger protects factual and evidential integrity. What claims were made? What source supports each one? What is verified, what is inferred and what is unsupported? Where an incorrect claim has entered connected work, Claim Ledger identifies what remains reliable, isolates what has been contaminated and traces where the unsupported claim has travelled.
GATE provides the audit and release discipline: Ground, Assess, Test and Exit. IRPR stands on its own as a judgement method, while also serving as the judgement layer at the Test stage of the Visualitas Self-Audit instrument. In that role, it examines whether an output has correctly understood, bounded and reconstructed the task. Factual claims are tested separately against a declared evidence boundary.
The audit concludes in a bounded release decision: Pass within scope, Conditional Pass within scope, or Hold. It does not declare that content is safe or true. Any pass is limited to the declared evidence, audit coverage, scope and intended use. A structured self-audit is recorded as such and is not presented as independent verification. A named human validation owner remains accountable for consequential release.
IRPR is not intended to replace source verification, professional expertise or formal organisational controls. It works alongside Claim Ledger and GATE by providing a disciplined method for testing the problem being answered, protecting what must remain intact and determining what should be rewritten.
IRPR improves the judgement applied to an output. Verification tests whether its claims are adequately supported. Human oversight determines whether it may be relied upon for its intended use.
Origin and provenance
Barry O'Brien developed the IRPR Framework at Visualitas. Its earliest documented public application currently identified was on 12 March 2026, and the completed four-stage framework was first stated publicly on 16 March 2026. This was followed by the IRPR Framework Working Paper in March 2026. The framework was applied publicly to a well-known editorial AI failure of 19 March 2026. Visualitas tracks this provenance through dated source files, public posts, and archived copies.
Preferred citation
O'Brien, B. (2026) The IRPR Framework: Interrogate, Reframe, Preserve, Rewrite. Visualitas. Available at: visualitas.ie/IRPR
Origin document: O'Brien, B. (March 2026) The IRPR Framework: Interrogate, Reframe, Preserve, Rewrite. Unpublished working paper, Visualitas.
First published: 18 July 2026
Version: 1.1
Last updated: 8 August 2026
About the author
Barry O'Brien is the founder of Visualitas. His work focuses on AI governance, EU AI Act readiness, Microsoft 365 Copilot, verification workflows and AI-assisted decision design. He also lectures on artificial intelligence with UCD Professional Academy.
