Fragility is not falsehood.
That's the harder problem.

We've written recently about fabricated citations and amplification cascades — the ways evidence quality breaks down at the level of sources. But there's a subtler and in some ways more important failure mode: claims that are true, cited correctly, and accepted as established, but whose evidential foundations are thinner than the weight being placed on them.

A position paper published in May 2026 by researchers at Lexsi Labs introduces a precise term for this: fragile assurance. A claim exhibits fragile assurance when the evidential structure does not support the strength of the assertion being made. The claim may be correct. The citation may be real. The methodology may be sound. And none of that tells you whether the evidence is strong enough to bear the load being placed on it.

The paper focuses on AI governance, but the argument generalises to any domain where evidence has to do work — policy research, regulatory compliance, investment due diligence, clinical guidelines. Anywhere a claim is used to justify a decision, the question of whether the evidence is strong enough to support that use is distinct from the question of whether the evidence exists.

The audit gap

The paper's central argument is about a structural mismatch the authors call the audit gap. AI governance frameworks enacted between 2019 and 2026 — the EU AI Act, California SB-53, Singapore's AI Verify, South Korea's AI Basic Act, India's AI Governance Guidelines, the Council of Europe AI Convention — all require reviewable evidence of specific properties. The problem is that current assurance methods are not built to produce that evidence.

Behavioural evaluations and red-teaming, the standard tools for AI safety assurance, can characterise observable model outputs. They cannot verify latent representations or long-horizon behaviours. A model that hides a problematic property during evaluation generates the same behavioural record as a model that never had it. The evidence produced by the evaluation is real. The conclusion drawn from it may still be fragile.

Thirteen jurisdictions enacted or strengthened seventeen governance instruments in this pattern between 2019 and early 2026. All of them presume that documented risk management, traceable testing, and conformity assessment constitute evidence of the properties they are meant to regulate. The paper argues this presumption is structurally unsound for a specific class of high-consequence claims — and that the gap is widening rather than closing as governance language sharpens.

Where the problem appears in research

The governance context is specific, but the underlying dynamic appears in any field where evidence is produced under pressure and used to justify decisions. It shows up in a few recognisable patterns.

The single-study problem

A finding from one study, in one context, gets cited as establishing a general principle. The study is real. The citation is accurate. But the evidence for the generalisation is thin — a finding from thirty organisations in one country over six months does not establish a universal pattern. The fragility is in the gap between what the evidence actually shows and how it is used.

The proxy-evidence problem

Process compliance gets treated as evidence of outcome quality. A framework is followed, boxes are ticked, documentation is produced — and the documentation is then treated as evidence that the underlying thing is safe, reliable, or effective. The documentation is real. The inference from documentation to quality is fragile.

The recency problem

Evidence from 2019 on a rapidly-evolving question is treated as current. The study exists. The citation is correct. But the applicability of that evidence to conditions in 2026 is a separate question from whether the study was sound when it was conducted. Evidence ages, and the rate at which it ages varies enormously by domain.

Why governance frameworks are particularly exposed

The Lexsi Labs paper focuses on AI governance because that's where the stakes are currently highest and the audit gap is most visible. But the dynamic it describes — governance instruments presupposing evidence that assurance practice cannot produce — is not unique to AI.

Regulatory compliance in any fast-moving domain faces the same structural problem. The requirements are written based on what sound practice looks like in principle. The evidence produced to demonstrate compliance is whatever current methodology can actually generate. The gap between the two grows when the domain moves faster than the methodology.

For policy researchers and compliance professionals, this matters in a practical way. When a governance framework requires documented evidence of a property, the question of whether the documentation actually constitutes evidence of that property is a separate question — one that the framework itself often doesn't help you answer. A system card is not evidence that a model is safe. A conformity assessment is not evidence that a system is compliant with the outcome the regulation is trying to achieve. These documents are evidence that a process was followed. Whether the process reliably produces the outcome is a different claim, and often a fragile one.

The connection to fabricated citations

In our previous piece on fabricated citations, we argued that citation presence is no longer a meaningful evidence guarantee. Fragile assurance is the next level of that problem. Even when citations are real, even when sources exist and resolve and contain the finding attributed to them, the evidence can still be insufficient for the use being made of it.

The three-part distinction from that piece holds here too, extended by one step:

Most research workflows operate somewhere between steps one and two. The fabrication crisis makes step one newly unreliable. Fragile assurance describes what happens when evidence passes all four steps individually but the aggregate is still insufficient — when each source is real, each finding is correctly cited, independent corroboration exists, and the claim is still being used to carry more weight than the evidence as a whole can bear.

What this means for how research is communicated

The practical implication is about how confidence is communicated alongside claims. A brief that presents a finding as established when the evidence is actually thin at the edges is not being dishonest — the finding may well be correct. But it is creating fragility in the decision-making process that depends on the brief.

The appropriate response is not more hedging language. "May suggest" and "appears to indicate" do not communicate evidential fragility — they communicate authorial uncertainty, which is different. What's needed is a way to distinguish between a claim that is well-established, a claim that is directionally supported but not robustly evidenced, and a claim that is plausible but rests on thin evidence. These are different epistemic statuses, and conflating them is where fragile assurance enters research outputs.

For the governance context specifically, the paper recommends protocols that generate auditable verification artifacts — evidence that can be independently reproduced and checked — rather than documentation that a process was followed. The same principle applies to research more broadly: what makes evidence non-fragile is not that it exists, but that it can be independently reproduced, checked, and challenged.

A true claim with fragile evidence is not a safe claim to build a decision on. Recognising that distinction is the start of actually improving evidence quality, rather than just expanding evidence volume.