The short version
The IMF's March 2026 Executive Board review set a medium-term target of SDR 25 billion, a minimum floor of SDR 20 billion, and named the Middle East as the specific geopolitical theatre driving downside risk. The Board recorded the number of countries subject to surcharges in FY2026 falling from 20 to 13, a figure anchored in IMF Press Release 24/376. And a Q2FY26 Quarterly Financial Report placed precautionary balances at SDR 26,782 million at 31 October 2025.
Claude Opus 4.7 got none of these right, not across a single one of the six questions the RLB Specialist Panel put to it. The failures weren't random noise. They fell into a clear pattern: numbers that lag one biennial cycle behind, baselines inflated under deliverable pressure, Board lexicon softened in ways that change what a practitioner reads from it, and geographic attributions that expanded beyond what the regulator actually said.
The SDR 5 billion floor problem
The IMF Board's March 2026 review is unambiguous: Directors generally agreed to retain the current floor at SDR 20 billion. The model committed to SDR 15 billion, a full biennial cycle behind, including in a board briefing memo drafted for an emerging-market finance ministry client.
That five-billion-dollar gap is exactly what supervisors, counterparties, and internal QC reviewers check against the source first. A model that embeds a confident wrong number into a board deliverable doesn't create a minor inaccuracy. It creates a falsifiable error in a document that may reach a minister's desk.
The surcharge count inflation
IMF Press Release 24/376 records the FY2024 surcharge-payer baseline as 20 countries, with FY2026 expected to fall to 13. The model's policy-brief draft inflated the starting number to 22, producing a 22-to-13 trajectory. That looks like a steeper reform win than the regulator recorded, which matters for any sovereign-debt practitioner using the trajectory to plan multi-year debt-service scenarios.
A separate finding on the same reform: the model upgraded the Board's characterisation of the early-surcharge-review signal from "a few Directors", the regulator's precise term, to "a number of Directors." In IMF Board lexicon, these are not interchangeable. "A few" signals a minority view. "A number" can imply broader traction. A practitioner reading the upgraded phrasing in an advisory would draw a different signal on review timing.
Adding Ukraine to a sentence the Board didn't
The Board's March 2026 review names one specific geopolitical theatre in its downside-risk language: the Middle East. The model added Ukraine. This isn't a policy disagreement, it's a factual attribution error. A central-bank backgrounder or legal-and-policy advisory that attributes both named theatres to the Board record is contradicted by the face of the regulator's text.
For AI lab teams: this failure class, where a model expands a specific named list under generation pressure, is distinct from pure hallucination. The model has plausible reasons to associate Ukraine with IMF downside risk language in 2024–2026. The problem is that it substituted its own contextual knowledge for the regulator's specific textual record.
Three failure shapes, one audit
When these failures surface in the wild
These aren't edge-case prompts. The RLB Specialist Panel designed questions to mirror how practitioners actually use AI on this regulation: drafting board memoranda for EM finance ministry clients on Fund near-term lending capacity, drafting policy briefs on the October 2024 surcharge reform, preparing backgrounders for central-bank bilateral meetings with the IMF Managing Director, and drafting desk notes for sovereign-credit research teams on the half-year PB trajectory.
Any of these deliverables hitting the wrong floor figure, the wrong surcharge baseline, or the wrong Board lexicon is a falsifiable error in a client-facing document. Retrieval-anchored verification against the current biennial review text is the minimum mitigation. General-purpose prompting does not resolve this, the failures survive web-search-enabled configurations.
How the audit was run
The RLB Specialist Panel authenticated the primary source, the IMF's March 2026 Review of the Adequacy of the Fund's Precautionary Balances, IMF Press Release 24/376, and the Q2FY26 Quarterly Financial Report, as substrate v1 prior to testing. Each question was designed to mirror a real practitioner workflow: board memos, policy briefs, campaign reports, bilateral meeting backgrounders, and desk notes. Claude Opus 4.7 was tested with web search active.
Findings are immutably recorded under RLB Citation IDs. The IMF and any named entity holds a permanent right of reply on every finding. Full detail is in the Hallucination Register.
The RLB Panel's standing offer
The RLB Specialist Panel offers the International Monetary Fund and any other named entity a permanent right of reply on every finding published here. Corrections are applied with the same immutable citation discipline as the original findings. Contact: Right of Reply form.