SINGAPORE, June 10, 2026. The RegLeg Brief Specialist Panel released today findings showing that Two frontier AI models, each running with web search enabled, produced confidently wrong reconstructions of two of the most operationally consequential mechanics in the International Monetary Fund's 2024 Guidance Note on Financing Assurances and Sovereign Arrears. The first failure is on the entry conditions for the Lending Into Official Arrears Strand 4 pathway, where both models invented tests the Guidance Note does not contain.

The second failure is on the creditor-coverage threshold for the "sufficient set" in pre-emptive restructurings, where Opus 4.7 imported a majority-of-financing test from a different part of the framework into a question where the Guidance Note specifies no numerical threshold.

The regulation under review is the 2024 IMF Guidance Note on financing assurances and the Fund's policy on lending into official arrears, the operational framework staff and the Executive Board apply when a member country seeks Fund financing while one or more official bilateral creditors are unwilling or unable to commit to a restructuring on terms consistent with the program parameters.

The Guidance Note distinguishes a sequence of strands, of which Strand 1 (representative standing forum agreement, with a majority-of-financing test for adequacy) and Strand 4 (Fund-determined additional safeguards where the upstream strands do not yield a sufficient creditor outcome) are the strands that the findings touch. The Note also distinguishes post-default cases from pre-emptive cases, and in pre-emptive cases requires financing assurances from a "sufficient set" of creditors, without specifying a numerical coverage threshold for that set.

What the models said

Claude Sonnet 4.6 was asked when the Strand 4 pathway is activated, and specifically whether a bilateral creditor's failure to respond to a restructuring consent request within four weeks satisfies the entry conditions, or whether an affirmative refusal to restructure is required. The model answered that Strand 4 is not available simply because one creditor is slow or silent, that there must be an affirmative signal of unwillingness to engage, and that the country should document this for Fund staff.

The Guidance Note states the Fund shall seek Strand 4 safeguards where an adequately representative agreement has not been reached through a representative standing forum, and where consent is not forthcoming. The four-week consent window is the trigger the Note specifies; the affirmative-refusal test is not. The model substituted a conduct-based reading for a structural reading.

Claude Opus 4.7 was asked the same activation question through a sovereign debt management brief frame. The model described an obligation of good-faith engagement with all official bilateral creditors on terms consistent with program parameters and inter-creditor equity, a test that the non-participating creditor's stance is the binding obstacle, and a criterion that activation advances orderly resolution. None of those three elements appears in the Strand 4 entry conditions. The Note specifies a three-part structural gate: unavailability of a Strand 1 representative-forum agreement, absence of creditor consent within four weeks of request, and inability to satisfy the Strand 3 criteria.

The model produced a framework that reads as plausible policy reasoning and that is not the framework the Guidance Note contains.

Claude Opus 4.7 was also asked, in a Finance Minister briefing frame, what creditor coverage satisfies IMF financing assurance requirements in a pre-emptive debt restructuring, and how the "deemed away" mechanism works for creditors who do not commit. The model answered that a "sufficient set" must account for more than 50 percent of the total financing contributions required from official bilateral creditors over the program period, plus any applicable standing creditor forum, plus any creditor with significant influence over the debtor.

The Guidance Note specifies that in pre-emptive cases, financing assurances would only be sought from a sufficient set of creditors, and that if a sufficient set commits, creditor coordination has de facto been achieved and other creditors' arrears would be deemed away for the purposes of Fund arrears policy. The Note specifies no numerical coverage threshold for the sufficient set in pre-emptive cases. The 50-percent figure the model produced is the majority-of-financing test from the Strand 1 adequately-representative-Paris-Club-agreement context, where it does appear, transposed into a different part of the framework where it does not.

A fourth output, from Opus 4.7 on a G20 roundtable presenter frame, asked the same pre-emptive coverage question and produced the same majority-of-financing transposition. The convergence inside Opus 4.7 across two differently framed questions about pre-emptive coverage is part of the finding.

Why the framework matters

The Strand 4 entry conditions are the gate to the Fund's most consequential financing assurance pathway. Activating Strand 4 changes the burden on the borrower country, on the Fund's staff and Board, and on the non-participating creditor whose conduct is at issue. A debt management team that advises the country to wait for an affirmative refusal before triggering Strand 4 leaves the country slower to access Fund financing than the Guidance Note allows.

A team that advises the country to argue good-faith engagement, holdout-as-binding-obstacle, and orderly-resolution as the entry criteria builds the activation case on grounds the Note does not recognise. Either path raises the cost and time of the activation negotiation with Fund staff, and risks an activation that the staff paper, when written, cannot anchor on the Guidance Note.

The "sufficient set" threshold matters at the design stage of the restructuring perimeter. A finance ministry desk officer or sovereign debt advisor designing the creditor outreach for a pre-emptive case, working off a 50-percent coverage assumption, scopes the perimeter to a shortlist that the Guidance Note does not require. If actual coverage exceeds 50 percent but is not what staff would treat as a sufficient set in the case-specific judgement the Note actually applies, the country may proceed under a false assurance and discover at the Board paper stage that staff are not prepared to deem the remainder away.

If actual coverage is below an unspecified internal staff threshold but above 50 percent, the same problem.

The reform is also operative in current cases. The financing assurances framework is the framework that the Fund applies to the Common Framework cases, to non-Common-Framework restructurings, and to the pre-emptive cases that creditor countries and debtor countries are increasingly framing pre-default. A wrong reading of either the Strand 4 entry conditions or the sufficient-set coverage requirement carries directly into live case work.

What we tested

The Specialist Panel test on this regulation focused on the activation conditions for Strand 4 and on the creditor-coverage rules for pre-emptive financing assurances, because those are the questions where the Guidance Note text is most operationally consequential and where divergence is least recoverable downstream. Both models were given the same effective date for the 2024 Guidance Note, the same regulator (IMF Strategy, Policy and Review Department), and the same substrate (the Guidance Note covering Q1, Q3 and Q6 of the financing assurances framework). Both were run with web search enabled.

Both were asked the activation and coverage questions in operational terms, framed for the readers who use these answers.

The Panel did not test the post-default arrears policy, the Strand 2 or Strand 3 entry conditions in isolation, or the broader review of the lending-into-arrears framework that the Guidance Note sits within. Those tests were out of scope for this finding set. The finding is narrow and specific: on the Strand 4 entry conditions and the pre-emptive sufficient-set coverage rule, two frontier models with web search enabled produced reconstructions that do not match the Guidance Note text.

Failure classification

All four findings are classified as inference drift. The Sonnet 4.6 Strand 4 answer reconstructs the entry test from a policy-reasoning prior (affirmative refusal as the appropriate trigger) rather than from the Guidance Note's structural language (consent not forthcoming within four weeks). The Opus 4.7 Strand 4 answer reconstructs the entry test from three plausible policy criteria (good-faith engagement, holdout-as-obstacle, orderly resolution) none of which the Guidance Note contains. The two Opus 4.7 sufficient-set answers reconstruct the coverage threshold by transposing the Strand 1 majority-of-financing test into the pre-emptive context where no numerical threshold is specified.

The shared failure mode, across both AIs and across both questions, is that web search did not anchor the answers on the Guidance Note's operative language. Both models had retrieval enabled. Neither model surfaced the four-week consent window as the Strand 4 trigger. Neither model surfaced the absence of a numerical threshold for the pre-emptive sufficient set.

The convergence between Sonnet 4.6 and Opus 4.7 on the Strand 4 question, through different policy-reasoning paths to the same wrong destination, is the signal that the Guidance Note's structural reading is under-indexed in the content both models pull from, relative to the policy-reasoning reconstructions.

Who is exposed

A sovereign debt management team advising a country on whether to trigger Strand 4 against a non-responsive bilateral creditor, working off Sonnet 4.6's affirmative-refusal test, waits for a signal the Guidance Note does not require and delays the country's access to Fund financing. The same team working off Opus 4.7's good-faith-and-orderly-resolution test argues the activation on grounds Fund staff cannot anchor on the Note.

A finance ministry desk officer designing the creditor perimeter for a pre-emptive restructuring, working off Opus 4.7's 50-percent coverage threshold, scopes the outreach to a shortlist the Guidance Note does not require and risks a staff paper that does not deem the remainder away.

A G20 roundtable presenter or multilateral development bank policy team explaining how the 2024 framework works for pre-emptive cases, working off the same 50-percent reconstruction, propagates a numerical threshold into the policy discussion that the Guidance Note does not contain. A sovereign debt research analyst at a rating agency or sell-side macro desk, modelling the financing assurance perimeter for current cases, prices the perimeter off the wrong threshold.

The exposure is at the operative mechanics, not at the description of the framework. The Strand 4 entry conditions and the pre-emptive sufficient-set rule are the conditions that determine whether the framework yields a financing assurance outcome in a live case. Substituting an affirmative-refusal test, a good-faith-and-orderly-resolution test, or a 50-percent coverage threshold for what the Guidance Note actually requires is not a paraphrase; it is substitution at the operative gate.

Right of reply

The IMF Strategy, Policy and Review Department, which drafted the Guidance Note, and the IMF Communications Department were notified of the Specialist Panel findings on 4 June 2026 with a deadline for comment of 9 June 2026; no response was received by the time of release. Anthropic, which produces Claude Opus 4.7 and Claude Sonnet 4.6, was notified of the model-level findings on 4 June 2026 with the same deadline; no response was received by the time of release. The right of reply remains open and any response received will be appended to the public record.

Methodology note

The Specialist Panel runs each model under the same prompt configuration, captures the response verbatim, and compares the response against the regulator's published text. Findings are published only where the model commits to a specific operative claim that the regulator's published text contradicts. Paraphrase, qualitative characterisation of the framework, and well-calibrated refusals are not published. The substrate for this finding set is the 2024 IMF Guidance Note on financing assurances and the Fund's policy on lending into official arrears, covering Q1, Q3 and Q6 of the framework.

The substrate is held in the Panel's audit record and is identifiable by the citation IDs RLB-H-INT-IMF-IMF-GUIDANCE-FINANCING-ASSURANCES-SOVEREIGN-ARREARS-2024-Q001-Sonnet46, RLB-H-INT-IMF-IMF-GUIDANCE-FINANCING-ASSURANCES-SOVEREIGN-ARREARS-2024-Q001-Opus47, RLB-H-INT-IMF-IMF-GUIDANCE-FINANCING-ASSURANCES-SOVEREIGN-ARREARS-2024-Q003-Opus47, and RLB-H-INT-IMF-IMF-GUIDANCE-FINANCING-ASSURANCES-SOVEREIGN-ARREARS-2024-Q006-Opus47.

Right of Reply

These findings and associated work have been put up in public with a view of the greater good for the development of a safer AI ecosystem. Any party reading this or any finding on reglegbrief.com may contact us and have an unconditional right of reply; the Specialist Panel will publish any factual correction or contextual response alongside the original finding, with no editorial gatekeeping. Researchers, regulators, and compliance teams with questions on methodology or specific findings can reach the Specialist Panel via the same channel.

Source & Methodology Standards

RegLeg Brief is operated by Verdus Technologies Pte. Ltd. (UEN 201616982R), incorporated in Singapore. The RLB Specialist Panel, with an aggregate of over 60 years of public-policy and industry experience, documents only confirmed hallucination findings, under a methodology that requires a verbatim regulator excerpt for every documented claim. All findings, citation IDs, model outputs, regulator excerpts, and methodology notes are open-access.

Primary source verified: IMF Guidance Note on Financing Assurances in the Context of Sovereign Arrears (2024) · Substrate documents: R2-REGULATION-Q1_Q3_Q6_Guidance_Note_Sovereign_Arrears.pdf · IMF portal: imf.org

Citation IDs referenced:

• RLB-H-INT-IMF-IMF-GUIDANCE-FINANCING-ASSURANCES-SOVEREIGN-ARREARS-2024-Q001-Opus47
• RLB-H-INT-IMF-IMF-GUIDANCE-FINANCING-ASSURANCES-SOVEREIGN-ARREARS-2024-Q001-Sonnet46
• RLB-H-INT-IMF-IMF-GUIDANCE-FINANCING-ASSURANCES-SOVEREIGN-ARREARS-2024-Q003-Opus47
• RLB-H-INT-IMF-IMF-GUIDANCE-FINANCING-ASSURANCES-SOVEREIGN-ARREARS-2024-Q006-Opus47