Ask one AI a hard question and the answer arrives polished and confident. Nothing inside that answer shows which parts are wrong, or what it quietly left out.
HAIA-CAIPR, Cross AI Platform Review, is the governance framework for sending the same question to several AI platforms at once and reading the comparison instead of trusting any single reply. It is pronounced kay-per, and the resemblance to caper is intentional, because the point is to walk away with intelligence no one platform could produce.
What CAIPR Is
CAIPR sends one task to every platform in a chosen group, collects what each returns, and puts the whole comparison in front of a named human who decides. One AI then compiles those returns into a single view. That compilation is treated as the least trustworthy artifact in the room rather than the most, because it sits furthest from the raw evidence and closest to the conclusion.
The framework came out of practice rather than design. Multi-AI work was producing published output from 2023, the method was first disclosed in February 2024, it ran as a five-platform routine from September 2025, and it received its name in March 2026.
The Problem It Solves
A single platform cannot tell a reader which of its own claims it invented. In one documented session, a platform returned a European Commission Delegated Act that does not exist, at 92 percent declared confidence, shaped almost perfectly to look like validation of work already on file. Nine other platforms in the same dispatch did not repeat it, and that silence is what raised the flag.
Adding platforms is not a fix on its own, and the research now says so plainly. A panel of nine frontier models drawn from seven families carries roughly two independent votes’ worth of information, because the models tend to make the same mistakes on the same items (Kohli, 2026). Three independently developed models share a mean error correlation of 0.77 (Spiro, 2026). Agreement across AI platforms is a question worth asking, not an answer worth trusting.
CAIPR treats it that way. The count is diagnostic and never a vote, and unanimous agreement on a claim that matters triggers verification outside the pool rather than confidence inside it.
How It Works
- Dispatch. The same task and the same collection instructions go to every platform in the group. Three is the floor, and the ceiling is however many separately operated platforms the work can reach.
- Isolation. No platform sees another’s answer before producing its own, so convergence means something and is not an echo.
- Read. The human reads every raw return before any of it reaches a synthesizer. This is the step that cannot be delegated without losing the audit.
- Synthesize and audit. One AI, the Navigator, compiles the return set. Its output carries the highest scrutiny, preserves the minority positions rather than smoothing them away, and approves nothing.
- Decide. A named human decides, reading both the returns and the compilation, with accountability attached to the decision.
The CAIPR session chain. Source authority is classified at the moment each input arrives, and the count never establishes truth.
Nine conditions hold across every session, running from parallel isolated generation through minority preservation to human decision authority. Five further choices vary by session, including how many platforms run, how the returns are structured, and how deeply the human reads, and each of those choices names what it gives up. The full set is specified in the paper.
What the Comparison Produces
Subtraction: what is wrong gets exposed
Fabrication, stale retrieval, and citations that format correctly while supporting nothing become visible when other platforms fail to repeat them.
Addition: what is missing gets supplied
One platform surfaces the source none of the others found, or frames the problem in a way none of them reached. In a single ten-platform session, one platform supplied eleven of sixteen unique contributions, which is why the solitary voice is checked rather than discarded.
Warning: what is learned gets published
Every session produces dated evidence about how specific platforms behaved. Keeping that private would make the framework a competitive advantage. Publishing it makes the comparison a public good, and that is the reason the paper exists in the form it does.
When It Is Worth the Cost
CAIPR costs money and attention, and it earns that cost only where a missed error or a missed contribution cannot be recovered afterward. Four kinds of work meet that bar. Publication-bound work carries an author’s name beyond their control, so a fabricated citation outlives the correction. Regulatory and legal claims turn a wrong figure into a position of record. Decisions that bind third parties give the people affected no chance to check the reasoning. And framework work sends a definition into documents nobody will revisit.
It is the wrong tool for routine work, where role-assigned HAIA-RECCLIN dispatch does the job at a fraction of the effort. It is also wrong for early ideation, because critique constrains itself around an existing draft instead of generating alternatives to it.
Where CAIPR Sits
CAIPR is one framework inside the HAIA ecosystem. Factics supplies the evidence discipline at every checkpoint, pairing a fact with a tactic and a measurable outcome. Checkpoint-Based Governance supplies the authority, since every checkpoint in a CAIPR session is a CBG checkpoint and CAIPR creates no authority of its own. HAIA-RECCLIN supplies the ten-field format the returns arrive in, which is what makes a minority of one assessable rather than merely present.
Two record-keeping protocols catch what a session produces. HAIA-SCOPE holds custody of each citation, and HAIA-CARCS holds the record of the session itself.
Read the Full Work
HAIA-CAIPR: Cross AI Platform Review, Fourth Edition
The complete framework: nine invariants, five configuration variables, the eight core operations, nine documented synthesizer failure modes, the regulatory positioning against Article 14, eight open questions, and five outcomes that would falsify it.
Download the paper as a PDF
Fourth Edition, version 6, September 2026. Published under Creative Commons Attribution 4.0 International.
The specifications on GitHub
Open source, free to use, free to argue with.
Common Questions
What does HAIA-CAIPR actually stand for?
Cross AI Platform Review. HAIA is the ecosystem it belongs to, the Human Artificial Intelligence Assistant. The acronym was chosen by a human after eleven platforms produced forty-seven candidates and none of them survived review.
How many AI platforms does it take?
Three is the working minimum and a floor rather than a guarantee. Sessions commonly run three, five, seven, nine, eleven, or thirteen, and the number is set against what a mistake would cost rather than against what is convenient.
If the platforms all agree, is the answer right?
No. Agreement can mean shared training data, shared search results, or a blind spot all of them inherited. Under CAIPR, agreement with no dissent anywhere on a claim that matters is a flag that sends the question outside the AI pool for checking.
Does this make AI trustworthy?
It does not, and it does not try to. CAIPR makes a set of fallible sources collectively useful by exposing what one invents, recovering what the rest missed, and leaving the decision with a person who can explain the reasoning and answer for the result.
Is any of this automated?
Not today. Every session in the published record ran by hand, with the human dispatching, collecting, and routing directly. The agent layer and the enforcement layer that would carry the mechanics are specified and built, and neither is deployed.
Basil C. Puglisi, MPA
A Human-AI Collaboration
#AIassisted using HAIA Ecosystem