AI Hallucination ResearchRegulatorsGlobal standard-settersINTBIS-CPMICPMI-API-HARMONISATION-CROSS-BORDER-2024White paperDetail › Finding
AI Labs · updated 2026-06-04 · methodology v2.1

Finding on 'Q001 Probe' for Claude Sonnet 4.6 with web search ON

RLB Citation ID: RLB-H-INT-BIS-CPMI-API-HARMONISATION-CROSS-BORDER-2024-Q001-Sonnet46
What the RLB Specialist Panel found

Finding on 'Q001 Probe' for Claude Sonnet 4.6 with web search ON

  • Question (paraphrased to protect IP): A payments counsel prepared a client memo on the CPMI's October 2024 API harmonisation report (d224). The memo correctly identified the four recommendation categories but fabricated specific sub-requirements for each category — presented as approximate guidance for what each individual recommendation requires — when that content is inaccessible in any public source.
  • AI's response: "Category 1 recommendations (approximately 2–3 recommendations) broadly require: (i) establishing or designating a governance body to maintain harmonised API standards on an ongoing basis; (ii) creating a process to evaluate and endorse industry API proposals"
  • Regulator's text: Individual recommendation text is not accessible in any public source. The landing page identifies four recommendation categories and states the report contains ten recommendations in total.
  • Why the AI went wrong: The model used the category label — "harmonisation governance and processes" — as a generative prompt, producing plausible institutional sub-requirements by analogy with how standards governance typically works. The output is structurally consistent with regulatory governance norms but has no basis in the CPMI text. The "approximately 2–3 recommendations" qualifier signals the model knew it was constructing rather than retrieving, yet it presented the content as categorical requirement guidance.
  • Cited source(s):
Impact for this audience

The 'approximately 2–3 recommendations' qualifier embedded in the fabricated output is a leaked inference signal: the model knew it was constructing an answer, not retrieving one. The calibration gap is that this internal signal did not propagate to a refusal or a clearly-marked 'this is my inference' framing — it was suppressed in favour of producing a structurally complete response. Post-training reward for producing complete answers may be overriding the calibration signal for retrieval confidence.

← Previous finding Finding on 'Q010 Probe' for Claude Opus 4.7 with web search ON Next finding → Finding on 'Q005 Probe' for Claude Sonnet 4.6 with web search ON
Cite this finding

Each finding has a stable Citation ID (RLB-F-… for aggregated case-study findings, RLB-H-… for raw per-model hallucinations) — like a DOI, the ID always resolves to the canonical finding even if URLs change.

RLB Citation ID: RLB-H-INT-BIS-CPMI-API-HARMONISATION-CROSS-BORDER-2024-Q001-Sonnet46
Plain text Download
RegLeg Specialist Panel (2026). "Finding on 'Q001 Probe' for Claude Sonnet 4.6 with web search ON — AI Labs." Citation ID: RLB-H-INT-BIS-CPMI-API-HARMONISATION-CROSS-BORDER-2024-Q001-Sonnet46. RegLegBrief AI Hallucination Research, published 2026-06-04. https://reglegbrief.com/regulators/j1/int/bis-cpmi/cpmi-api-harmonisation-cross-border-2024/whitepaper/finding/INT-BIS-CPMI-INT-001-CPMI-API-HARMONISATION-CROSS-BORDER-2024-v1-001--sonnet-46-websearch/
APA 7th edition Download
RegLeg Specialist Panel. (2026). Finding on 'Q001 Probe' for Claude Sonnet 4.6 with web search ON [Hallucination finding RLB-H-INT-BIS-CPMI-API-HARMONISATION-CROSS-BORDER-2024-Q001-Sonnet46]. RegLegBrief AI Hallucination Research. https://reglegbrief.com/regulators/j1/int/bis-cpmi/cpmi-api-harmonisation-cross-border-2024/whitepaper/finding/INT-BIS-CPMI-INT-001-CPMI-API-HARMONISATION-CROSS-BORDER-2024-v1-001--sonnet-46-websearch/
Bluebook / OSCOLA (US + UK legal) Download
RegLeg Specialist Panel, Finding on 'Q001 Probe' for Claude Sonnet 4.6 with web search ON [RLB-H-INT-BIS-CPMI-API-HARMONISATION-CROSS-BORDER-2024-Q001-Sonnet46], RegLegBrief AI Hallucination Research (June 04, 2026), https://reglegbrief.com/regulators/j1/int/bis-cpmi/cpmi-api-harmonisation-cross-border-2024/whitepaper/finding/INT-BIS-CPMI-INT-001-CPMI-API-HARMONISATION-CROSS-BORDER-2024-v1-001--sonnet-46-websearch/.
BibTeX Download
@misc{reglegbrief_RLB_H_INT_BIS_CPMI_API_HARMONISATION_CROSS_BORDER_2024_Q001_Sonnet46,
  author    = {RegLeg Specialist Panel},
  title     = {Finding on 'Q001 Probe' for Claude Sonnet 4.6 with web search ON},
  year      = {2026},
  publisher = {RegLegBrief AI Hallucination Research},
  note      = {Hallucination finding Citation ID: RLB-H-INT-BIS-CPMI-API-HARMONISATION-CROSS-BORDER-2024-Q001-Sonnet46},
  url       = {https://reglegbrief.com/regulators/j1/int/bis-cpmi/cpmi-api-harmonisation-cross-border-2024/whitepaper/finding/INT-BIS-CPMI-INT-001-CPMI-API-HARMONISATION-CROSS-BORDER-2024-v1-001--sonnet-46-websearch/}
}
← Back to case study summary Case study detail →