← RegLegBrief Findings by regulator All AI Labs whitepapers Hallucination Register Methodology Right of reply
Model said
19
surcharge-paying countries
IMF PR 24/376 says
20
surcharge-paying countries
AI Hallucination Research › Regulators › IMF › White paper

Off by one — and the one that's wrong is the baseline

Both Claude Opus 4.7 and Claude Sonnet 4.6, each with web search active, cited 19 surcharge-paying countries as the pre-reform baseline when describing the IMF's October 2024 surcharge reform. IMF Press Release 24/376 records 20. The resulting 20-to-13 trajectory the regulator documents becomes a 19-to-13 trajectory in both models' outputs, a fabricated baseline that misrepresents a politically prominent reform.

📅 Published 7 Jun 2026 ⚙️ Methodology v2.3 📋 IMF PR 24/376 · Oct 2024 reform
📰Read the public briefing for this regulation→
The reform and its baseline

October 2024: the surcharge reform and the number that matters

The IMF's October 2024 surcharge reform cut borrowing costs for the Fund's most indebted members. It lowered the threshold at which surcharges kick in, reduced the rates, and reworked the time-based schedule. When practitioners summarise the reform, the number they reach for is the surcharge-payer count: where it was before the change, where it's headed after.

IMF Press Release 24/376 — verbatim
The number of countries subject to surcharges in fiscal year 2026 is expected to fall from 20 to 13. The approved measures will lower IMF borrowing costs by about US$1.2 billion annually, or reduce payments on the margin of the rate of charge as well as surcharges on average by 36 percent.
Both models said
19
Pre-reform baseline
countries with surcharges
IMF PR 24/376 records
20
Pre-reform baseline
countries with surcharges
Expected post-reform (FY2026)
13
Countries subject
to surcharges
Failure analysis

Same wrong number, same two models, same web search active

The convergence is what makes this finding material for AI lab teams. Both Claude Opus 4.7 and Claude Sonnet 4.6, the two principal frontier configurations tested, produced the same wrong baseline independently. Both had web search active. The single-digit off-by-one error (19 instead of 20) is consistent across model generations, and it compounds through the reform narrative: every policy brief, background note, or research summary that describes the reform's country-impact trajectory using the model's output carries a fabricated baseline.

It is not random. Both models under-counted by exactly one, in the same direction. That kind of consistent drift will clear most QC checks unless someone independently verifies against Press Release 24/376.

SURCHARGE-PAYING COUNTRIES — TRAJECTORY IMF PR 24/376 (correct) 20 → Oct 2024 reform 13 (FY2026 expected) Model output (both models) 19 13 (matched) −1 ✗
Both models produced a 19-to-13 trajectory. IMF PR 24/376 records a 20-to-13 trajectory. The endpoint matched; the baseline didn't.
Reform parameters — what's correct

The actual reform figures for AI lab reference

US$1.2bn
Annual reduction in IMF borrowing costs (Press Release 24/376)
36%
Average reduction in rate-of-charge and surcharge payments
20 → 13
Surcharge-paying countries: pre-reform to FY2026 expected (correct trajectory)
Oct 2024
Reform date · IMF Press Release 24/376

The RLB Specialist Panel offers IMF and any named entity a permanent right of reply. Full regulation hub: IMF-CHARGES-SURCHARGE-REFORM-2024 →

Hallucination Register: reglegbrief.com/hallucination-register/