The bottom line

For AI lab teams fielding frontier models into IMF-adjacent, sovereign-debt advisory, central-bank-advisory, and sovereign-credit research deployments, the six findings recorded in this audit document a recurring pattern in how Claude Opus 4.7 handled the International Monetary Fund's March 2026 Review of the Adequacy of the Fund's Precautionary Balances and the closely linked October 2024 charges and surcharge reform.

The model committed to specific, verifiable parameters that the regulator's own primary text directly contradicts: a precautionary balances floor of SDR 15 billion (against SDR 20 billion), a FY2024 surcharge-payer baseline of 22 (against 20), a pre-March-2024 floor of SDR 10 billion (against SDR 15 billion), a Board characterisation of 'a number of Directors' on the early-review signal (against 'a few Directors'), the addition of Ukraine to the Board's named geopolitical theatre (the regulator names only the Middle East), and a half-year PB level of approximately SDR 26.5 billion at October 31, 2025 (against SDR 26,782 million).

Every finding is bound to verbatim primary source text recorded by the International Monetary Fund. The pattern signals systematic gaps in how the model handled biennial-review parameters, reform-adjacent baselines, regulator-specific characterisation lexicons, named theatre attributions, and quarterly-report financial figures.

What the regulation actually records

The March 2026 Review of the Adequacy of the Fund's Precautionary Balances records that Most Directors supported retaining the current medium-term target for precautionary balances at SDR 25 billion. Directors generally agreed to retain the current floor for precautionary balances at SDR 20 billion, noting that it provides an important safeguard against shocks and helps ensure the Fund retains sufficient buffers. The Board cautioned that the Fund's income and precautionary balances projections are subject to heightened uncertainty including from financial market volatility and intensifying downside risks to global growth stemming in particular from geopolitical developments in the Middle East.

Recognizing the uncertain environment, in the event that precautionary balances rise well above the target, a few Directors saw merit in considering an early review of charges and the surcharge policy in due course.

Adjacent to the March 2026 Review, IMF Press Release 24/376 on the October 2024 charges and surcharge reform records that the approved measures will lower IMF borrowing costs by about US$1.2 billion annually, reduce payments on the margin of the rate of charge as well as surcharges on average by 36 percent, and that the number of countries subject to surcharges in fiscal year 2026 is expected to fall from 20 to 13. The IMF Q2FY26 Quarterly Financial Report Schedule 2 records precautionary balances at October 31, 2025 at SDR 26,782 million, against SDR 25,905 million at April 30, 2025.

What the models got wrong

Claude Opus 4.7, the AI subject in this audit, committed to six specific answers that the regulator's own primary text directly contradicts. First, the model committed to a minimum floor of SDR 15 billion at the March 2026 review, against the regulator's recorded SDR 20 billion. Second, the model recorded a FY2024 to FY2026 surcharge-payer trajectory of 22 to 13, against the regulator's recorded 20 to 13. Third, the model recorded a pre-March-2024 floor of SDR 10 billion stepping to SDR 15 billion, against the regulator's recorded SDR 15 billion stepping to SDR 20 billion.

Fourth, the model added Ukraine to the Board's named geopolitical theatre on intensifying downside risk; the regulator's text names only the Middle East. Fifth, the model recorded the Board's early-surcharge-review signal as held by 'a number of Directors'; the regulator records the position as held by 'a few Directors'. Sixth, the model committed to a half-year PB level of approximately SDR 26.5 billion at October 31, 2025, against the regulator's recorded SDR 26,782 million in the Q2FY26 Quarterly Financial Report Schedule 2.

What this means for AI lab teams

For AI lab teams fielding frontier models into IMF-adjacent advisory deployments, the six findings translate directly into evaluation surfaces. Cycle-anchored biennial-review parameters (floor, target, half-year level, FY count) require retrieval-priority handling that defaults to the most-recent-cycle citation. Reform-adjacent single-value baselines (FY2024 surcharge-payer count, projected savings, percentage reduction) require citation-bound rather than generation-bound reproduction. Regulator-specific characterisation lexicons (IMF Board, BIS committee, FSB plenary) require fixed-strength term encoding with no synonym substitution. Named theatre attributions to specific Board records on politically sensitive items (Middle East, Ukraine, Asia-Pacific, etc.) require absence-of-addition verification before any named-theatre claim is returned.

Quarterly-report financial figures require retrieval-anchored reproduction at the schedule-level granularity recorded by the regulator.

How the failure mode works

The pattern across the six findings has three distinct mechanisms. The first mechanism is cycle-trajectory drift on biennial-review parameters: the model pulled the SDR-billion floor back by one biennial cycle on both Finding 1 (the March 2026 reaffirmation, where the model committed to the prior-cycle SDR 15 billion rather than the current-cycle SDR 20 billion) and Finding 3 (the March 2024 step, where the model committed to a SDR 10 billion to SDR 15 billion step rather than the regulator's SDR 15 billion to SDR 20 billion step).

The second mechanism is single-value inflation or approximation under generation pressure: the model inflated the FY2024 surcharge-payer baseline by two from 20 to 22 (Finding 2) and rounded the half-year PB level by approximately 280 million SDR from SDR 26,782 million to approximately SDR 26.5 billion (Finding 6).

The third mechanism is attribution drift on politically sensitive named items: the model added Ukraine to the Board's named geopolitical theatre, where the regulator's text names only the Middle East (Finding 4), and elevated the IMF Board lexicon characterisation of the early-surcharge-review signal from 'a few Directors' to 'a number of Directors' (Finding 5). Each mechanism produces a structurally plausible answer in a confident deliverable register where the downstream user would paste the answer into the working deliverable before verification against the source.

Where the AI-lab evaluation risk concentrates

The evaluation surface for an AI lab concentrates on four product registers. The first is the verbatim-reproduction register where the model is asked to record what a specific Board recorded at a specific cycle: cycle-trajectory drift on biennial-review parameters surfaces here. The second is the analytical-section register for a policy-brief or sovereign-debt advisory deliverable: single-value inflation on reform-adjacent baselines surfaces here. The third is the backgrounder register for a central-bank bilateral meeting or a board-level risk-narrative summary: named theatre attribution drift surfaces here on politically sensitive items.

The fourth is the legal-and-policy advisory register on multi-year debt-service planning: regulator-specific characterisation lexicon drift surfaces here on Board-signal strength terms. Each of these registers concentrates verifiable risk where the model's answer reads as if the regulator's text had been directly retrieved.

Action items for AI lab teams

1. Add a cycle-anchor verification probe to the pre-deployment evaluation pipeline for any frontier model fielded into IMF-adjacent advisory contexts. The probe should ask the model to reproduce the same parameter (PB floor, target, FY surcharge-payer count) across three consecutive review cycles and check for cycle alignment.

2. Encode the IMF Board lexicon as a fixed-strength characterisation system in post-training instruction-following: 'a few Directors' / 'a number of Directors' / 'several Directors' / 'Directors generally agreed' / 'Most Directors' should not be substituted under generation pressure.

3. Add a named-theatre attribution probe to the evaluation pipeline. The probe should ask the model to record what a specific Board has named on a specific risk narrative and check for absence-of-addition on politically sensitive named items.

4. Treat quarterly-report financial figures as retrieval-anchored outputs at the schedule-level granularity recorded by the regulator. Product surfaces that promise verbatim reproduction of regulator-issued quarterly figures should anchor on the specific schedule citation rather than a model-generated approximation.

Where to find the verifiable primary source

Every finding in this briefing is bound to verbatim primary source text recorded by the International Monetary Fund.

The substrate documents referenced across the six findings are: the IMF Press Release on the March 2026 Review of the Adequacy of the Fund's Precautionary Balances (recording the floor at SDR 20 billion, the target at SDR 25 billion, the named geopolitical theatre, and the IMF Board lexicon strength of the early-surcharge-review signal); IMF Press Release 24/376 on the October 2024 charges and surcharge reform (recording the FY2024 to FY2026 surcharge-payer trajectory of 20 to 13 and the projected cost reductions); and the IMF Q2FY26 Quarterly Financial Report Schedule 2 (recording the October 31, 2025 precautionary balances level at SDR 26,782 million and the April 30, 2025 level at SDR 25,905 million).

The RLB Specialist Panel holds the substrate on file and binds each finding to a specific substrate document name, a section anchor, and a verbatim excerpt.

Right of Reply

International Monetary Fund and any other named entity in this audit are offered a permanent, unedited right of reply on every finding. A response is appended verbatim to the finding record on receipt; the original finding remains visible alongside. Responses can be submitted via the contact channel on the RegLeg site.

Source and Methodology Standards

Every finding in this audit is bound to verbatim regulator-issued primary source text held as substrate by the RLB Specialist Panel. The substrate document, the section anchor, and the verbatim excerpt are recorded against each finding. Findings are published only where the regulator's own primary text directly contradicts the AI subject's response. The RLB Specialist Panel does not publish positive findings, blind spots, or claims unsupported by primary substrate.

Primary source verified: every finding in this audit is bound to verbatim text recorded by the International Monetary Fund in the substrate document identified against the finding record. The RLB Specialist Panel holds the substrate on file.

Citation IDs referenced

• RLB-H-INT-IMF-IMF-PRECAUTIONARY-BALANCES-REVIEW-2026-Q001-Opus47
• RLB-H-INT-IMF-IMF-PRECAUTIONARY-BALANCES-REVIEW-2026-Q005-Opus47
• RLB-H-INT-IMF-IMF-PRECAUTIONARY-BALANCES-REVIEW-2026-Q009-Opus47
• RLB-H-INT-IMF-IMF-PRECAUTIONARY-BALANCES-REVIEW-2026-Q011-Opus47
• RLB-H-INT-IMF-IMF-PRECAUTIONARY-BALANCES-REVIEW-2026-Q012-Opus47
• RLB-H-INT-IMF-IMF-PRECAUTIONARY-BALANCES-REVIEW-2026-Q014-Opus47