Research
What we measure, and why we build the way we do.
Two lines of work, both published so you can check them. The first maps what research on compliance AI actually evaluates, and sets out the evaluation set we use on our own system. The second asks why capable teams refuse tools that work, and what that means for how a tool should be built.
Evaluation
What the field measures, and what compliance teams actually need measured.
We map the published literature on retrieval-grounded language models for regulated finance, and we publish the evaluation set we use on our own system. Both are here so you can check them, including the parts that do not flatter us.
Paper 01 · September 2026
Retrieval-grounded, adapter-tuned language models for regulated finance: a map of the 2021–2026 literature
8,532 papers clustered into twenty themes, and the most-cited 456 finance and legal RAG papers labelled by what they measure. End-task accuracy appears in 71% of those; refusal and abstention in 2%. The specific US consumer-finance rules are almost absent.
Read the map →
Evaluation · v0.1 draft
CFR-QA: an evaluation set for consumer-finance regulation, including refusal
77 items across Regulations Z, E, CC and B. Answers are scored on facts, false premises have to be corrected, and refusal is scored in both directions. We are publishing the design before any score.
See the design →
Adoption and design
Why people refuse tools that work
A working paper on moralized resistance to generative AI, and five technical notes extending it. The design argument in Section 8 is the reason this company builds the way it does.
Most accounts of technology adoption treat refusal as a verdict on a tool. The working paper below argues that a significant share of resistance to generative AI is not a verdict on the tool at all — it is a defense of the self, transmitted socially and expressed in moral language.
That argument has a design consequence, and it is not a comfortable one for the way most of this industry builds. If the barrier is identity rather than usability, then making a system easier to use does not address the objection. Making the user’s own judgment visible in the output does.
The five notes are arguments, not studies. They extend the paper’s propositions into the domain we work in and cite published work by others. None of them report experiments run by JDM Inc, and each says so at the top.
Working paper
Refusal as Identity Work: Toward a Theory of Moralized Resistance to Generative AI
Jesse McCleary
Three literatures — the moralization of effort, identity-protective cognition, and complex contagion — combine into an account of why capable people form confident positions on generative AI before encountering it, and why reversing that position is not evidence that the original concern was answered. Five propositions are offered for empirical work.
Read the paper →
Technical notes
Five short pieces extending the paper’s propositions. Each argues a position and cites published work; none report original findings.
01
Capability and Resistance
Why the best objection is not the most informative one
Kahan’s climate finding — that greater numeracy produced more polarization, not less — applied to high-expertise domains. If ability upgrades the defense rather than neutralizing it, then sophistication of argument and depth of engagement come apart, and only one of them is a signal.
02
Costless Positions in Procurement
How refusal moves through a peer network
Adopting a tool is a complex contagion: it needs repeated reinforcement, a real trial, and tolerance of visible early incompetence. Declining it is a simple contagion that costs one sentence and confers standing. That asymmetry shapes regulated-industry buying more than the product does.
03
Legibility by Design
Building for an objection that is not about usability
The paper’s design implication, taken seriously. Legibility is not explainability: one is a property of the machine offered to an auditor, the other is whether the user can say what they contributed and defend it. This is the note that grounds the product.
04
Conversion and Measurement
Why asking people about AI recovers the wrong variable
If refusal was socially produced, the reversal probably was too — which makes the convert the least reliable witness on the question they were originally raising. What that breaks in survey methodology, and what a behavioral instrument would have to do instead.
05
Offloading without Atrophy
The question the moral framing was standing in front of
Strip the moral vocabulary and a well-posed empirical question remains. The Bastani field experiment contains both the harm and the mitigation in the same study, which turns a moral dispute into an engineering specification with one known-good configuration.
A note on what this is
The paper is a conceptual argument, presently unfalsified rather than supported, and it carries a risk its author names directly in Section 9: a framework explaining resistance as identity-protective can be turned into a device for dismissing every objection to AI as psychology. That would be an abuse of it. The claim is that the strength and social distribution of objections are partly identity-driven. It is not a claim that the objections are false — and on the one substantive question the paper examines, it concludes the central worry is well founded.
Point it at your hardest question.
If you run compliance at a bank, credit union, lender or RIA and you want to see the evaluation set, or to put your own questions into it, write to us. We will tell you plainly whether the corpus covers them today.
research@jmvent.com
We reply to everyone.