Paper deep dive
CLAIR-Fin: An Adversarial Multi-Agent Framework for Claim-Level Verification and Adaptive Debate in Cross-Modal Financial QA
Fatema Tuj Johora Faria, Mukaffi Bin Moin, Jubayer Al Mahmud, M. F. Mridha, Md. Alam Hossain
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/17/2026, 4:07:00 AM
Summary
The paper introduces CLAIR-Fin, a nine-agent framework designed for claim-level verification and adaptive debate in cross-modal financial question answering. It addresses hallucination in retrieval-augmented generation by decomposing questions into atomic claims stored in a Financial Claim Ledger. The system employs Asymmetric Evidence Authority to weight evidence based on claim type and modality, Chain-of-Custody Verification to check grounding at hand-offs, and an Adaptive Rebuttal Cycle for adversarial debate. Evaluated on the BB-FinQA-X dataset, CLAIR-Fin significantly improves faithfulness compared to baselines like HyDE and Graph-RAG, while also providing a mechanism for selective abstention when evidence is insufficient.
Entities (12)
Relation Signals (9)
Fatema Tuj Johora Faria â affiliatedwith â Ahsanullah University of Science and Technology
confidence 95% ¡ Fatema Tuj Johora Faria 1 ... 1 Ahsanullah University of Science and Technology, Bangladesh
CLAIR-Fin â evaluatedon â BB-FinQA-X
confidence 95% ¡ We evaluate CLAIR-Fin on BB-FinQA-X, a 500-question cross-modal financial evaluation set
CLAIR-Fin â uses â Asymmetric Evidence Authority
confidence 95% ¡ Each claim is resolved through Asymmetric Evidence Authority, which conditions evidence trust on claim type
CLAIR-Fin â uses â Adaptive Rebuttal Cycle
confidence 95% ¡ an Adaptive Rebuttal Cycle, which routes contested claims through adversarial debate
CLAIR-Fin â uses â Financial Claim Ledger
confidence 95% ¡ CLAIR-Fin, a nine-agent framework that decomposes each question into atomic claims maintained in a typed Financial Claim Ledger.
CLAIR-Fin â uses â Chain-of-Custody Verification
confidence 95% ¡ Chain-of-Custody Verification, which checks grounding at the hand-off between drafting and adversarial review
BB-FinQA-X â derivedfrom â Bangladesh Bank Annual Report
confidence 90% ¡ BB-FinQA-X, a 500-question cross-modal financial evaluation set built from Bangladesh Bank Annual Report material
CLAIR-Fin â â
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Existing defenses against hallucination in retrieval-augmented and multi-agent pipelines remain partial: evidence is trusted despite modality disagreement, debate verifies an aggregate report rather than individual claims, and such verification occurs only after drafting, leaving inter-agent errors undetected until the final text. To close this gap, we present CLAIR-Fin, a nine-agent framework that decomposes each question into atomic claims maintained in a typed Financial Claim Ledger. Each claim is resolved through Asymmetric Evidence Authority, which conditions evidence trust on claim type rather than treating all modalities as equally reliable; Chain-of-Custody Verification, which checks grounding at the hand-off between drafting and adversarial review rather than only at the pipeline's exit; an Adaptive Rebuttal Cycle, which routes contested claims through adversarial debate whose depth scales with what that debate finds; and a terminal entailment audit paired with a continuous Hallucination Risk Index that distinguishes claims that passed scrutiny from claims never contested. We evaluate CLAIR-Fin on BB-FinQA-X, a 500-question cross-modal financial evaluation set built from Bangladesh Bank Annual Report material, stratified by query type, format, and difficulty. Relative to a single-pass retrieval-augmented generation baseline, it raises faithfulness ($0.780 \rightarrow 0.889$) while abstaining on 5.4% of questions when evidence is insufficient rather than forcing an unsupported response, and it exceeds stronger retrieval-strategy baselines such as HyDE and Graph-RAG on faithfulness ($\leq 0.874$).
Tags
Links
- Source: https://arxiv.org/abs/2608.13706v1
- Canonical: https://arxiv.org/abs/2608.13706v1
Trouble viewing inline? Open PDF directly â
Full Text
110,527 characters extracted from source content.
Expand or collapse full text
CLAIR-Fin: An Adversarial Multi-Agent Framework for Claim-Level Verification and Adaptive Debate in Cross-Modal Financial QA Fatema Tuj Johora Faria 1 , Mukaffi Bin Moin 1 , Jubayer Al Mahmud 2 , M. F. Mridha 3 , Md. Alam Hossain 2 1 Ahsanullah University of Science and Technology, Bangladesh 2 Jashore University of Science and Technology, Bangladesh 3 American International University - Bangladesh Correspondence: mukaffi28@gmail.com, fatema.faria142@gmail.com Abstract Existing defenses against hallucination in retrieval-augmented and multi-agent pipelines remain partial: evidence is trusted despite modality disagreement, debate verifies an ag- gregate report rather than individual claims, and such verification occurs only after draft- ing, leaving inter-agent errors undetected un- til the final text.To close this gap, we presentCLAIR-Fin, a nine-agent framework that decomposes each question into atomic claims maintained in a typed Financial Claim Ledger. Each claim is resolved through Asym- metric Evidence Authority, which conditions evidence trust on claim type rather than treat- ing all modalities as equally reliable; Chain-of- Custody Verification, which checks grounding at the hand-off between drafting and adversar- ial review rather than only at the pipelineâs exit; an Adaptive Rebuttal Cycle, which routes con- tested claims through adversarial debate whose depth scales with what that debate finds; and a terminal entailment audit paired with a con- tinuous Hallucination Risk Index that distin- guishes claims that passed scrutiny from claims never contested. We evaluateCLAIR-Finon B-FinQA-X, a 500-question cross-modal fi- nancial evaluation set built from Bangladesh Bank Annual Report material, stratified by query type, format, and difficulty. Relative to a single-pass retrieval-augmented generation baseline, it raises faithfulness (0.780â 0.889) while abstaining on 5.4% of questions when evidence is insufficient rather than forcing an unsupported response, and it exceeds stronger retrieval-strategy baselines such as HyDE and Graph-RAG on faithfulness (⤠0.874). 1 Introduction Financial institutions increasingly rely on large language models to answer questions over long, multimodal reports, where the same fact may ap- pear as prose, a table, and a chart within one docu- ment exceeding several hundred pages. A misread Analyst Query What was the overall balance of payments outcome in FY25, and how did it contrast with the FY24 outcome? Generated Response The overall balance of payments outcome in FY25 showed an improvement compared to FY24, as evidenced by a narrowing of the current account balance deficit and an increase in workers' remittance inflows. Specifically, the current account balance deficit decreased to USD 139 million in FY25. CLAIR-FinFinancial Analyst Submit Query to CLAIR-Fin Returns Audited Response From Analyst Query to Audited Answer Analyst Query Annual Report (Text ¡ Tables ¡ Charts) The service sector remained the largest contributor to GDP, with a 51.62% share in FY25. (1) Narrative Evidence Agent (2) Tabular Evidence Agent (3) Visual Evidence Agent Evidence Collection (1) Affirmative Counsel (2) Adversarial Counsel Adversarial Debate CitationâEntailment Audit Figure 1: An analystâs question enters CLAIR-Fin, where specialized agents extract evidence for each claim. Each claim undergoes EvidenceâDebateâAudit, with contested claims adversarially reviewed and cita- tion entailment verified before synthesis into the final response. fiscal-year label or an approximated chart value reported as exact can materially change a conclu- sion; even state-of-the-art vision-language mod- els hallucinate on chart-reading tasks (Wang et al., 2025). Cross-modal benchmarks building on ear- lier tabular-textual QA work (Chen et al., 2021; Zhu et al., 2021), including FinanceBench (Islam et al., 2023), FAMMA (Xue et al., 2025), and XFin- Bench (Zhang et al., 2025), find current LLMs fail many realistic financial questions, worsening as context grows and evidence sits away from a docu- mentâs start (Ji et al., 2026), precisely the regime 1 arXiv:2608.13706v1 [cs.CL] 13 Aug 2026 long reports fall into. Such systems must retrieve relevant content, reconcile conflicting cross-modal evidence, and answer faithfully or decline when evidence is insufficient. Several lines of prior work address pieces of this problem, but each leaves a gap.Multi- agent frameworks such as MDocAgent (Han et al., 2025) coordinate specialized agents over long doc- uments, and multimodal retrieval frameworks such as MultiFinRAG (Gondhalekar et al., 2025) and FinRAGBench-V (Zhao et al., 2025) jointly in- dex tables, figures, and text, yet none conditions evidence trust on claim type. Financial debate frameworks such as FinDebate (Cai et al., 2025) and Structured Adversarial Synthesis (Sadhu et al., 2025) apply adversarial roles to report-level anal- ysis (Du et al., 2024; Chan et al., 2023), but run a fixed round count over a holistic thesis regardless of how contested a claim is. Verification approaches such as Chain-of-Verification (Dhuliawala et al., 2023), FinGround (Guo et al., 2026), and FRED (Tan et al., 2025) instead check atomic claims, and self-consistency sampling offers a claim-agnostic alternative (Manakul et al., 2023); both act only after drafting and pair no binary outcome with a continuous risk estimate (Section 2.3). Bench- marks such as FinBen (Xie et al., 2024) also skip non-calendar fiscal years and the dense statistical tables typical of South Asian central-bank report- ing. Across this literature, evidence authority is uniform, verification is late if applied at all, and abstention is rarely measurable (Wen et al., 2025; Kirichenko et al., 2025). We introduce CLAIR-Fin (Claim-Ledger Adversarial Inference and Retrieval for Financial document understanding), a nine-agent framework for reliable question answering over multimodal financial documents, assessed on central-bank an- nual reports (Appendix A), whose fiscal-year def- initions, currency denominations, and reporting practices differ from the corporate filings predomi- nantly represented in prior evaluation suites (Sec- tion 2). A Planner-Orchestrator decomposes each question into a Financial Claim Ledger (FCL) containing atomic, typed assertions, which are pop- ulated by three evidence agents and reconciled by a Ledger Guardian through Asymmetric Evidence Authority (AEA), a claim-type-aware weighting scheme that prioritizes tabular evidence for pre- cise numerical values while favoring narrative text for causal attribution. Before adversarial review, Chain-of-Custody Verification (CoCV) validates and restores evidence grounding, preventing attri- bution drift from propagating into subsequent rea- soning. Disputed claims are then escalated to an Affirmative and an Adversarial Counsel through an Adaptive Rebuttal Cycle (ARC), with deliberation depth expanding according to unresolved findings rather than following a predetermined budget. A Judge-Auditor subsequently enforces a terminal en- tailment gate and records each AEA determination alongside its equal-weight counterfactual; a contin- uous Hallucination Risk Index (HRI) differentiates assertions that endured substantive scrutiny from those that were never challenged. Finally, a Brief Synthesizer formulates the response and abstains whenever the available evidence cannot adequately support an answer. We structure our evaluation of this framework around the following research questions: ⢠RQ1. To what extent does modality-aware ev- idence prioritization improve faithfulness and correctness over modality-agnostic retrieval? â˘RQ2. How reliably does hand-off-level verifi- cation reduce the propagation of unsupported claims? â˘RQ3. Does routing adversarial debate to con- tested claims improve factual grounding over skipping debate entirely? â˘RQ4. Does continuous risk estimation pro- vide a more informative reliability signal than binary verification alone? ⢠RQ5. How does the framework perform across single-modal and multimodal presenta- tion formats, and which configurations remain most challenging? 2 Related Work 2.1 Multimodal Financial Document Understanding and Retrieval-Augmented Generation Extending retrieval-augmented generation to mul- timodal financial documents is closest to our set- ting. General-purpose multimodal agents coordi- nate text and image agents over long documents (Han et al., 2025), while financially specialized re- trieval frameworks batch table and figure images through a lightweight multimodal model, escalat- ing to text+table+image context only when needed (Gondhalekar et al., 2025), within a RAG paradigm 2 (Gupta et al., 2024) assessed via a reference-free protocol scoring faithfulness, relevancy, and con- text precision/recall (Es et al., 2024). This im- proves what evidence is retrieved but treats re- trieval as complete once cited: a table cell and an approximately-read chart value for the same quan- tity are cited as interchangeable, with no mecha- nism to adjudicate which to trust when they dis- agree, a gap faithfulness scores compound since they measure entailment by some evidence, not the correct one. 2.2 Multi-Agent Debate and Financial Multi-Agent Systems A second line of work uses multi-agent de- bate, LLM instances critiquing and revising out- puts, to improve factuality and reasoning (Du et al., 2024). Financial applications add domain- specific structure: specialist-role frameworks as- sign earnings, market, sentiment, valuation, and risk personas with a trust/skeptic/leader safety layer (Cai et al., 2025), dialectical frameworks stage bull/bear/devilâs-advocate roles over earnings-call transcripts (Sadhu et al., 2025), and recent work asks whether added coordination improves out- comes relative to cost (Nguyen and Pham, 2026). The common assumption is that debate is a prop- erty of the report, not the claim: disagreement goes unflagged and debate depth stays fixed regardless of contestedness, leaving claim-level, difficulty- adaptive verification unaddressed. 2.3 Claim-Level Verification, Extraction Reliability, and Faithfulness Evaluation Closer to claim-level verification, a distinct litera- ture verifies individual statements rather than whole reports. Surveys characterize the fieldâs dominant pipeline, retrieve, decompose, check entailment, applied once to an already-finished claim (Dmonte et al., 2025); graph-structured verification converts a claim into an entity-relationship graph, checking each triplet before a verdict (Jeon and Lee, 2025), closer in spirit to our claim ledger but applied to open-domain claims rather than cross-modal evi- dence. Self-consistency methods instead sample multiple outputs and aggregate via majority agree- ment, on the premise that reproducibility signals reliability (Wang et al., 2023), though open-book QA depends on trustworthy extraction (Islam et al., 2023) and LLM-generated summaries frequently omit source figures despite high surface scores (Yang et al., 2024). Verification thus checks a claim only once formed, and reproducibility does not guarantee correctness. 2.4 Research Gap and Positioning of CLAIR-Fin Across these three directions, each addresses one piece of the problem, evidence arbitration, adap- tive debate, or hand-off verification, in isolation, and none operates jointly. CLAIR-Fin closes this gap: it conditions evidence trust on claim type so cross-modal disagreement is resolved by a stated, auditable prior; routes claims to adversarial debate only when evidence coverage is insufficient, scal- ing depth to scrutiny rather than a fixed budget; and verifies grounding at the hand-off between drafting and adversarial review, treating even self-consistent evidence as suspect. 3 CLAIR-Fin Framework 3.1 Problem Definition We address faithful QA over long, multimodal fi- nancial documents, where evidence spans prose, tables, and charts. Task. Given a questionqover corpusD, the framework produces answerawith citations, or abstains when evidence is insufficient. Claim decomposition.qis decomposed into ordered atomic claimsC = c 1 ,...,c n ,n ⤠8. Each c i = (id i , text i ,Ď i ) has a claim type Ď i âT =FACT_NUMERIC, FACT_TREND, CAUSE_ATTRIBUTION, RATIO_IDENTITY (1) fixing how its evidence is weighted (Section 3.4); claims resolve sequentially and independently be- fore synthesis. Evidence. Evidence spans four modalities, M =text, table, chart, tool_derived,(2) wheretool_derivedis a deterministically com- puted quantity (e.g., year-over-year growth) rather than directly read.Each iteme = (m, src, page, content, conf)carries modalitymâ M, provenance, and confidence confâ [0, 1]. Financial Claim Ledger (FCL). Claims and evidence accumulate in a typed, directed multigraphG = (V,E)with six node types (claim, text span, table cell, derived metric, chart region, constraint check) and edgesr â SUPPORTS, SAME_AS, VIOLATES_CONSTRAINT. Gpersists for the questionâs lifetime as the sole verification, audit, and citation artifact. We write 3 Phase I: Document Ingestion Phase I: Claim DecompositionPhase V: Adversarial Verification & Adaptive Rebuttal Phase I: Multimodal Evidence Retrieval Phase IV: Evidence Fusion & Authority- Weighted Escalation Phase VII: Authority-Weighted Terminal Audit Phase VIII: Answer Synthesis & Selective Abstention Phase VI: Chain-of-Custody Verification 1. Page Quality Check For each page, determine whether: The text layer is clean, and There is no chart or image signal. If both conditions are satisfied: Native Extraction extracts the text layer and rule- based tables without a model call. Otherwise: A single vision-model call extracts prose, tables, and charts together. A stronger vision model is used when chart/image signals are detected; otherwise, a cheaper model is used. 2. Table Verification Tables extracted by the vision model undergo up to three independent extraction passes. If 2 out of 3 passes achieve cell agreement ⼠0.8, the table is tagged high-confidence. Otherwise, the first-pass result is retained and tagged low-confidence. A duplicated section-total consistency check is also performed. 3. Sentence-Aware Chunking The extracted content is chunked using: Approximately 1,000 characters per chunk 2-sentence overlap No mid-sentence cuts The resulting units are organized as: 1 document / chunk 1 document / table, plus 1 additional unit per table row 1 document / chart Each unit is tagged with its: Source, Page, Modality 4. Embedding and Indexing All chunks are embedded and upserted into the vector store. Deterministic IDs are used. Re-ingestion therefore updates existing entries rather than creating duplicates. This produces the Indexed Corpus D. Given the indexed corpus D and question q, the Planner-Orchestrator decomposes the question into a set of claims: C=c1,...,cn, nâ¤8 Each claim is assigned one of four types: FACT_NUMERIC FACT_TREND CAUSE_ATTRIBUTION RATIO_IDENTITY The claims are processed sequentially. For each claim ci, three evidence agents operate in parallel. 1. Narrative Evidence Agent Retrieves up to 8 textual evidence units. Applies hedge, fact, and attribution tags. 2. Tabular Evidence Agent Retrieves up to 6 relevant table entries. Grounds evidence using: Row, Column, Value Uses a calculator when required. Calculator-derived results are tagged tool_derived. 3. Visual Evidence Agent Retrieves up to 5 relevant chart/visual findings. Produces hedged approximate findings where exact visual values cannot be established. The Ledger Guardian combines the retrieved evidence into an evidence ledger G. It also establishes SAME_AS links between equivalent evidence across different modalities. Claim coverage is calculated as: A(ci) =âw(Ďi,m) where w(Ďi,m) represents the authority weight of modality m for claim type Ďi. Escalation Decision ---> If: A(ci)<0.75 the claim is escalated to adversarial verification. Otherwise, the claim follows the fast path directly to the terminal audit. For escalated claims, the system performs adversarial verification. Affirmative CounselAdversarial Counsel The Affirmative Counsel drafts a brief using the supporting evidence: supp(ci) The Adversarial Counsel attempts to identify weaknesses in the draft, including: Numeric errors Scope mismatches Fiscal-year/temporal inconsistencies Causal overclaims Citation gaps Visual over-precision Each attack receives a severity: Ďâlow,medium,high The system checks whether the evidence entails the generated draft: â=entailsâ§ÎłâĽ0.5 If the draft is not grounded: Repair unused: return to Affirmative Counsel for revision. Repair exhausted: proceed to the terminal audit with the available evidence. If the draft is grounded: The claim proceeds to Adversarial Counsel. The adversarial loop terminates when: Abstention is recommended, or No high-severity attack remains, or The maximum number of rounds is reached: Ďmax=2 If a high-severity attack remains and rounds are available, the claim returns to Affirmative Counsel for another rebuttal. The Judge-Auditor determines the most authoritative modality for the claim: m*(ci)=argmax w(Ďi,m).confm(ci) The system maintains an Authority Docket that compares: Asymmetric winner: authority determined according to claim type and modality-specific weights. Uniform winner: modalities treated with equal authority. The final draft must pass an entailment gate: ÎłâĽ0.5 Claim Verdicts Each claim receives one of three verdicts: InsufficientEvidence Supported StronglySupported The system also calculates the Hallucination Risk Index (HRI): HRI=0.40(1âÎł)+0.25(1âs)+0.20(repairs)+0.15(highAttacks) After all claims have been processed, the Brief Synthesizer generates the final response. The synthesizer follows these rules: Produce one concise line per claim/verdict. Include citations for supported claims. Explicitly abstain from claims that lack sufficient evidence. Allow rephrasing only when citations are preserved. Abstain from the overall answer only if no claim passes the verification process. Figure 2: End-to-end CLAIR-Fin framework (Phases IâVIII, Section 3). A question is decomposed into atomic claims (Phases IâI), for which modality-specific agents retrieve evidence (Phase I) that is fused and authority- weighted to decide fast-path or escalation (Phase IV). Escalated claims pass through adversarial debate and a hand-off grounding check (Phases VâVI) before a terminal, authority-weighted audit (Phase VII). Once every claim is resolved, the framework synthesizes a cited answer or abstains (Phase VIII). score(u,v)for aSUPPORTSedgeâs confidence and supp(c) = v : (v,c, SUPPORTS) â Eforcâs supporting evidence. 3.2 Phase I: Document Ingestion Every source PDF is processed once, offline, into corpusD, which Phase I (Section 3.4) retrieves against. Each page is handled independently: native ex- traction (text layer plus rule-based table detection) is trusted only when the text layer is not garbled and the page has no chart/image signal; otherwise a structured vision-model call extracts prose, ta- bles, and charts from a rendered image, a stronger model when a chart or image is present, a cheaper one otherwise. Table extraction is least reliable: repeated vi- sion calls at temperature 0 can disagree. Each ta- ble is extracted up to three times and checked for pairwise agreement (Equation(4)); two-of-three agreement is tagged high-confidence, else the first extraction is kept, tagged low-confidence, and re- tained in the corpus for provenance, though its cells are excluded rather than downweighted when the Tabular Evidence Agent retrieves evidence (Sec- tion 3.4), so the table is kept as a record without ever being usable evidence. A second heuristic flags a table low-confidence if a row duplicates its own section-total row. Prose is split into sentence-aware chunks (1,000- character target, two-sentence overlap, sentence- boundary splits only). Each table yields one whole- table document plus one per labeled row; each chart yields one document per page. Documents are tagged with source, page, and modality, embed- ded, and upserted under a deterministic id, forming corpusD. 3.3 Phase I: Claim Decomposition Claim decomposition convertsqinto atomic, in- dependently verifiable claims: verifying a whole answer as one unit lets an unsupported figure hide inside a well-supported narrative, whereas decom- posing first lets every stage check grounding ex- actly. The Planner-Orchestrator retrieves a top-12, modality-agnostic preview (300 characters each) so claim text reuses the documentâs own terminol- ogy, then a structured LLM call (temperature 0) proposes 1â8 claims with text and typeĎ i âT; on failure it falls back to one claim restatingq, typed CAUSE_ATTRIBUTION, chosen because it commits to no specific number or trend that could later be flatly contradicted; as Section 3.5 discusses, this 4 typing also means the fallback claim is the one most likely to fast-path past debate rather than face additional scrutiny, so âdefaultâ should not be read as âmost rigorously checked.â Claims have no de- pendency structure, so each oneâs evidence, debate, and audit are fully independent, enabling per-claim auditability (Section 3.8). 3.4 Phase I: Multimodal Evidence Retrieval For claimc i , retrieval draws modality-specific evi- dence: a table figure, its prose interpretation, and an approximate chart restatement may all differ, so a dedicated agent per modality grounds evidence before it reaches the ledger. Given(text i ,Ď i )andD, retrieval blends dense and lexical signal: a candidate pool of4k(four times the requestedk) from vector search is reranked by score(x,d) = 0.65 1 1 + dist L2 (x,d) + 0.35 Ă |terms(x)⊠terms(d)| |terms(x)| . (3) and the topkreturned. Narrative retrievesk=8 text passages, taggedfact/hedge/attribution, extracting any(metric, period, value, unit)quadru- ple astool_derived. Tabular retrievesk=6ta- ble passages, grounding(row, column, value, unit) cells; cells failing consensus are dropped, not down- weighted. When two grounded cells share a row across columns, a deterministic calculator derives â % = (v cur â v prev )/v prev orâ p = v cur â v prev astool_derived. Visual retrievesk=5chart pas- sages, one finding per chart, kept as hedged prose so authority weighting governs its influence. Table extraction agreement (Section 3.2) uses: agree(T a ,T b ) ââ shape(T a ) = shape(T b ) ⧠|(i,j) : T a [i,j] = T b [i,j]| |T a | ⼠0.8. (4) This mirrors the Phase I agreement rule (Sec- tion 3.2), since agreement measures reproducibil- ity, not correctness.The phase passes per- modality evidence listse text ,e table ,e chart to the Ledger Guardian, each carrying only a (modality,confidence)pair, reusable unmodi- fied by later stages. 3.5 Phase IV: Evidence Fusion and Authority-Weighted Escalation Evidence fusion merges the three lists intoG, links same-metric evidence across modalities, and de- cides whether debate is needed: strong, agreeing evidence skips debate; weak or conflicting evidence does not, and âenoughâ evidence differs by claim type (a numeric claim needs a table cell, an attribu- tion claim needs prose). Each item becomes a typed node linked to the claim viaSUPPORTS(score = retrieval confidence); the Ledger Guardian addsSAME_ASedges when a table cellâs row label (âĽ4 characters) appears in a chart/text description, a lexical heuristic (Section 7). ForRATIO_IDENTITYclaims, a constraint node records VIOLATES_CONSTRAINT if uncomputed. AEA first acts here, in coverage form: each type Ďhas fixed weightsw(Ď,m)over modalities (Ta- ble 1, sums to 1 per row), giving coverage A(c i ) = X mâM i w(Ď i ,m),(5) whereM i âMis the set of modalities for whichc i has at least one grounded supporting evidence item inG(i.e.,M i = m : âv â supp(c i ), mod(v) = m ). The claim escalates iffA(c i ) < 0.75; else it proceeds to audit (Section 3.8). Three low- authority chart mentions forFACT_NUMERICcon- tribute less than one table cell. One consequence of this rule is worth flagging explicitly: since w(CAUSE_ATTRIBUTION, text) = 0.75(Table 1), a causal claim supported by a single grounded prose passage reachesA(c i ) = 0.75exactly, which fails the strict escalation inequality and fast-paths directly to audit without debate. This is deliberate: prose is the sole authoritative source for causal at- tribution here, so a single well-grounded passage is treated as sufficient coverage rather than auto- matically contested, and the terminal audit (Sec- tion 3.8) still independently checks the drafted sen- tence against that same evidence. It does mean causal overclaim, one of six adversarial attack cate- gories (Section 3.6), is structurally the attack type least likely to reach the Adversarial Counsel, a limi- tation of the coverage rule (Section 7), not evidence causal attributions are adequately scrutinized. Table 1: Asymmetric Evidence Authority weights w(Ď,m), assigning trust to each evidence modality per claim type; weights sum to 1 within each row and are a fixed design prior rather than a learned or human- validated distribution (Section 7). Ďtabletool_derivedtextchart FACT_NUMERIC0.550.300.100.05 FACT_TREND0.45â0.100.45 CAUSE_ATTRIBUTION0.25â0.75â RATIO_IDENTITY0.400.60â 5 The phase leaves an updatedG,escalate(c i ), and A(c i )in shared state, conditioning modality trust on claim type rather than a global weight: an ap- proximate chart reading and an audited table cell are interchangeable evidence for a causal attribu- tion but not for an exact figure. 3.6 Phase V: Adversarial Verification and Adaptive Rebuttal For escalated claims, the Affirmative Counsel and Adversarial Counsel draft and stress-test a claim before the audit gate, since a single-pass draft-then- check misses flaws only visible under scrutiny (e.g., a conflated fiscal-year label), while fixed-round debate wastes budget on easy claims. Givensupp(c i )and prior findings, the Affir- mative Counsel drafts (temperature 0.2); the Ad- versarial Counsel returns a scorecard of zero or more attacks across six types (numeric, scope, fiscal-year/temporal, causal overclaim, citation gap, visual over-precision) with severityĎ â low, medium, high , plus an abstain flag. This is ARC: rebuttal triggers, incrementingĎ, iffĎ=high exists andĎ < Ď max =2; otherwise the claim pro- ceeds to audit (including on an abstain recommen- dation), so debate depth tracks scrutiny, not a fixed quota. The phase yields a (possibly revised) brief, ad- versarial findings, and an abstain flag; cost scales with claim difficulty, not claim count, operational- izing claim-level debate that fixed-round methods (Section 2.2) do not. 3.7 Phase VI: Chain-of-Custody Verification CoCV checks that the affirmative draft stays grounded at the hand-off to adversarial review, not only at the pipelineâs end, catching drift before the Adversarial Counsel reasons over ungrounded content. Given the draft andsupp(c i ), CoCV reuses the Judge-Auditorâs entailment call (Section 3.8) at this earlier hand-off: an LLM-as-judge call returns â â entails, neutral, contradictsand con- fidence Îł, with grounding intact iff â = entails ⧠γ ⼠0.5.(6) On failure, one bounded repair is attempted; if still ungrounded, custody is marked bro- ken and the claim routes directly to audit as InsufficientEvidence, otherwise it proceeds to review grounded. Every check is logged. The in- ternal hand-off placement, plus the bounded single repair, guarantees fixed, predictable cost rather than an open-ended correction loop. 3.8 Phase VII: Authority-Weighted Terminal Audit The terminal audit (Judge-Auditor) reaches a final, citable verdict, resolving cross-modal disagreement via claim-type-conditioned authority rather than the drafting modelâs own judgment. Ifcustodywasbrokenornoevi- denceremains,theclaimresolvesto InsufficientEvidence(confidence0). Otherwise,per-modalityconfidenceis conf m (c i ) = max vâsupp(c i ), mod(v)=m score(v,c i ), and the winning modality follows Section 3.5âs weighting, now as argmax: m â (c i ) = arg max mâM [w(Ď i ,m)¡ conf m (c i )], s(c i ) = w(Ď i ,m â )¡ conf m â (c i ). (7) An equal-weight counterfactual (w unif (Ď i ,m) = 1/|M i | ) is also scored, yieldingËm(c i ); both are logged to an Authority Docket recording whether the asymmetric prior changed the winning modal- ity. Evidence is presented to drafting ordered by descending authority. A grounded sentence is drafted (temperature 0) and passed through the en- tailment gate (Equation(6)); failure resolves to InsufficientEvidence, and passing claims re- ceiveStronglySupportedifs(c i ) ⼠0.5, else Supported. For passing claims, HRI is: HRI(c i ) = 0.40(1â Îł) + 0.25(1â s(c i )) + 0.20 Ď repair Ď max repair + 0.15 min n high 3 , 1 (8) clipped to[0, 1], whereÎłis entailment confidence, Ď repair â 0, 1is the number of custody re- pairs (Ď max repair = 1: CoCV attempts at most one repair per claim, Section 3.7), andn high is the count of high-severity adversarial findings; weights (0.40/0.25/0.20/0.15) are a fixed design choice, like the AEA table (Section 7). HRI distinguishes claims that survived genuine scrutiny from those fast-pathed and never contested. 3.9 Phase VIII: Answer Synthesis and Selective Abstention The Brief Synthesizer composes one final answer from resolved claims, citing evidence and abstain- ing on failed claims; naive concatenation risks bury- ing an abstention or dropping citations. 6 Each claim contributes one line: its answer with citations if it passed audit, or an insufficient- grounding sentence if not. One LLM pass (tem- perature 0.2) rephrases and reorders these lines without adding content, falling back to raw concate- nation on a dropped citation label. The answer is abstained iff no claim passed audit; one supported claim among several unsupported ones still yields a substantive, partial answer. This leaves the final answer, citations, an abstention flag, and the com- plete ledgerG, persisted for analysis, composed only from already-verified fragments so nothing upstream can be undone by an unverified step. 4 Dataset Construction We construct B-FinQA-X, a 500-question mul- timodal financial QA dataset grounded in the Bangladesh Bank Annual Report. Detailed dataset construction, annotation, validation, and distribu- tion statistics are provided in Appendix A. 5 Experimental Configuration Detailed implementation settings, experimental configurations, and evaluation protocols are pro- vided in Appendix B. 6 Results and Discussion We evaluate CLAIR-Fin on B-FinQA-X across three axes: automatic retrieval and generation met- rics by query type and format (Tables 8, 10, 11), framework-specific metrics (Table 7), and two- annotator human evaluation (Table 9), alongside ablations against the full system and a single-pass RAG baseline (Tables 2, 3); detailed analysis is in Appendix D. 6.1 RQ1: Modality-Aware Evidence Prioritization Conditioning evidence trust on claim type is the most conservative of CLAIR-Finâs four mecha- nisms, and the ablation confirms this. Removing AEA produces the smallest degradation (â) of any ablated mechanism, yet consistently: faithfulness 0.889â 0.883, context recall0.897â 0.893, ex- act correctness0.592â 0.585(Tables 2â3). 1 The Authority Docket reports an AEA impact rate of 0.515 (Table 7): the asymmetric prior changes the 1 Throughout this section,X â Ydenotes a metricâs change from the full systemâs scoreXto the ablated (or con- trasted) scoreY;âandâmark whether the change is an improvement or a degradation. Table 2: RAGAS retrieval and generation metrics for CLAIR-Fin, four single-mechanism ablations, and four retrieval-strategy baselines on B-FinQA-X (n = 500). Ans. Rel.: answer relevancy. Ctx. Prec./Ctx. Recall: context precision/recall. Each ablated row disables ex- actly one mechanism (Appendix A.1), holding the rest of the system fixed; w/o Term. Audit removes the ter- minal entailment audit; Vanilla RAG is a single-pass retrieve-then-generate baseline with none of the four mechanisms. ConfigurationFaith.âAns. Rel.âCtx. Prec.âCtx. Recallâ CLAIR-Fin0.8890.6960.8160.897 w/o Term. Audit0.8450.6870.8030.886 w/o ARC0.7700.6800.7810.862 w/o AEA0.8830.6920.8120.893 w/o CoCV0.8570.6890.8070.881 Vanilla RAG0.7800.6800.7000.840 HyDE RAG0.8740.6910.8010.885 Hierarchical RAG0.8650.6880.7520.831 Graph-RAG0.8320.6940.7290.889 Table 3: CLAIR-Fin-specific metrics under the same configurations as Table 2 (n = 500). Faith. Rate: share of published claims passing citation-entailment verification. Exact Corr.: exact correct answer rate. Cov.: answer coverage, the share of questions receiving a non-abstained answer. Deb. Util.: debate utilization rate. AEA Imp.: AEA impact rate. A dash (â) marks a metric undefined for that configuration. ConfigurationFaith. RateâExact Corr.âCov.âDeb. Util.AEA Imp. CLAIR-Fin0.7830.5920.9460.6460.515 w/o Term. Audit0.7410.5610.9350.6390.509 w/o ARC0.6820.5240.896â0.501 w/o AEA0.7760.5850.9400.644â w/o CoCV0.7530.5480.9220.6410.508 Note: Faithfulness (Table 2) is RAGASâs semantic faithfulness score; Faithfulness Rate (here) is the share of published claims that pass citation-entailment verification. winning modality in roughly half of contested de- cisions, a substantial share, not a handful of edge cases. 6.2 RQ2: Verification at the Drafting-to-Review Hand-off Verifying claim grounding at the drafting-to-review hand-off, alongside a final pre-publication au- dit, catches unsupported claims beyond what ei- ther check alone would, and the two checks carry unequal weight. Removing the terminal entailment audit drops (â) faithfulness further (0.889â 0.845) than removing CoCV (0.889â 0.857), faithfulness rate showing the same ordering (0.783 â 0.741vs.0.783 â 0.753; Tables 2â3). The terminal gate thus carries more of the faithful- ness guarantee than any single upstream check, yet CoCVâs non-trivial residual cost shows hand-off checking still contributes independently rather than being redundant: the two checks are complemen- 7 tary, not substitutable. 6.3 RQ3: Adaptive Allocation of Adversarial Debate Routing contested claims through adaptive adver- sarial debate, rather than skipping it entirely, is where the frameworkâs gains concentrate most heavily. Removing ARC causes the largest degra- dation (â) of any ablated mechanism: faithfulness 0.889â 0.770, exact correctness0.592â 0.524, answer coverage0.946â 0.896(Tables 2â3), with a debate utilization rate of 0.646 (Table 7): nearly two-thirds of claims are routed through it. Since debate is reserved for claims below the fast-path coverage threshold, removing it eliminates the sole verification opportunity for exactly the claims most likely to be wrong. Adaptive debate is thus the single most consequential mechanism evaluated, precisely because it is targeted rather than indis- criminate. 6.4 RQ4: Continuous Risk Estimation versus Binary Gating A continuous risk score is only worth reporting alongside a binary pass/fail outcome if it carries information the gate does not already capture, and HRI clears that bar. HRI correlates negatively with correctness (r = â0.072, Table 7), the theoreti- cally expected direction, while human-rated absten- tion appropriateness (4.06/3.95,Îş = 0.84, Table 9) independently corroborates that risk-sensitive ab- stention aligns with human judgment. The mod- est correlation magnitude is consistent with HRI adding information at the margin rather than du- plicating the binary gate, a directionally correct, non-redundant, human-corroborated signal, though not yet a formally calibrated probability. 6.5RQ5: Sensitivity to Evidence Presentation Format Performance is not uniform across presentation formats, and the gap between easiest and hard- est points to where cross-modal understanding still struggles. Text + Table achieves the highest faithfulness (0.915,â˛) and Chart Only the low- est (0.850,âź; Table 10), while Evidence Retrieval (0.839) and Multi-hop Reasoning (0.840) are the lowest-scoring query types, essentially tied (Ta- ble 11). Matched-pair comparisons confirm the effect is attributable to evidence format itself: Ta- ble OnlyâťText Only by+0.030and Text+Chart âťChart Only by+0.025despite identical content. Difficulty is thus concentrated in chart-dependent evidence and in query types demanding evidence synthesis or grounding, while every combined for- mat outperforms its weakest constituent modal- ity, indicating the frameworkâs cross-modal fusion adds real value. 7 Conclusion Faithful question answering over long financial documents requires reconciling evidence across text, tables, and charts that do not always agree, a gap that prior multimodal retrieval, multi-agent debate, and claim-level verification methods leave unresolved. CLAIR-Fin closes this gap through a nine-agent framework built around a typed Finan- cial Claim Ledger, in which (1) evidence trust is conditioned on claim type rather than treated uni- formly (Asymmetric Evidence Authority, AEA); (2) grounding is checked at the hand-off between drafting and adversarial review rather than only at the pipelineâs exit (Chain-of-Custody Verification, CoCV); (3) debate is allocated adaptively to con- tested claims rather than run for every claim re- gardless of difficulty (the Adaptive Rebuttal Cycle, ARC); and (4) a terminal audit is paired with a continuous Hallucination Risk Index rather than a binary verdict alone. Empirically, on B-FinQA- X, CLAIR-Fin exceeds Vanilla RAG on faith- fulness (0.780 â 0.889; Table 2), and ablation confirms all four mechanisms are non-redundant, with removing ARC producing the largest drop (0.889â 0.770); these results suggest claim-type- conditioned evidence weighting and hand-off-level verification are properties other multi-agent sys- tems could adopt. Future work will replace the substring-matching heuristic linking cross-modal evidence with a semantic matcher; test generaliza- tion across institutions, languages, and models; and extend evaluation to multi-turn settings reflecting financial analyst use. Limitations CLAIR-Fin is designed for a specific problem set- ting, faithful, citation-grounded question answering over long, multimodal financial documents where narrative text, tables, and charts must be reconciled under strict correctness constraints. The following limitations define the scope of our claims rather than qualify the contributions above. Dataset Scope. B-FinQA-X is constructed from the Bangladesh Bank source report and cov- 8 ers a single institution, a single language (English), and a single central-bank reporting convention. The 500 questions are stratified by query type, format, and difficulty, but not by document diversity: every question is grounded in the same source corpus, so findings about cross-modal conflict rates, ex- traction reliability, and abstention behavior reflect this specific document family and should not be as- sumed to transfer to other central banks, fiscal-year conventions, or languages without further valida- tion. Methodological Constraints. Several compo- nents rest on fixed, hand-specified design choices rather than learned or calibrated parameters: the AEA weight table, HRIâs component weights, and the escalation and entailment-pass thresholds are all set by design, so their absolute values should be read as one reasonable operating point, not an optimum. Cross-modal linking relies on lexical substring overlap between a table cellâs row label and other nodesâ text, a coarse heuristic that misses semantically equivalent references and can occa- sionally over-link on coincidental matches. The framework also depends throughout on an LLM-as- judge for entailment checking, used identically at the hand-off and terminal-audit stages, so a system- atic bias in that judge is not independently caught by having two checkpoints. Finally, adversarial de- bate is capped at two rebuttal rounds and grounding repair at one attempt; a claim requiring more con- testation than this budget allows is resolved with whatever confidence the framework reaches, not necessarily full resolution of the disagreement. Evaluation Scope. Our automatic metrics, in- cluding RAGAS-style faithfulness and relevancy scores, are themselves computed by an LLM judge, sharing methodology with the frameworkâs own en- tailment checks; human evaluation addresses this but covers two annotators over the full set rather than a larger pool with formal inter-rater sampling. The ablation study isolates each mechanismâs con- tribution individually but not every combination, and does not measure wall-clock latency or the ad- ditional LLM calls each mechanism introduces, so the accuracy-cost trade-off is not directly quantified here. In particular, the w/o ARC ablation (Tables 2â 3) contrasts adaptive, claim-targeted debate against no debate at all; we do not additionally evaluate a fixed nonzero round budget applied uniformly to ev- ery escalated claim (e.g., always runningĎ max = 2 rounds regardless of what the debate finds), so the reported gain reflects debateâs presence versus its absence, not adaptive allocation versus a fixed al- ternative allocation. We also do not evaluate robust- ness to adversarial or out-of-distribution questions, or to noisy or corrupted source PDFs beyond the extraction-reliability findings from the evaluation corpus itself. Generalizability and Future Extensions. The claim-type taxonomy, evidence modalities, and au- thority weight table are specific to financial statisti- cal reporting; they were not designed with transfer to other high-stakes domains (legal contracts or clinical reports, for instance) in mind. The un- derlying architecture is more general, however: claim-level decomposition, modality-conditioned authority, hand-off verification, and continuous risk scoring do not depend on financial content, mak- ing them a natural direction for future adaptation. The framework has also been evaluated only in En- glish and against one model family; multilingual extension and validating whether the same weights and thresholds hold across models are open ques- tions. Finally, this work evaluates the framework offline, per-question, rather than in a deployed, multi-turn setting, so calibration drift, user trust over repeated interactions, and integration with an- alyst workflows remain future work. Ethics Statement Purpose and Intended Use. This work targets faithful question answering over long, multimodal financial documents, using the Bangladesh Bank Annual Report as an evaluation setting, intended to assist analysts and other domain-literate users in locating and verifying facts within reports that in- terleave narrative text, tables, and charts. CLAIR- Fin is a decision-support and information-retrieval aid, not a substitute for expert financial or regu- latory judgment; outputs should be reviewed by a domain-literate user, particularly claims marked low-confidence or abstained. Its scope is limited to the single-class central-bank statistical reporting setting evaluated here (Section 7). Data Sources and Privacy.B-FinQA-X is con- structed entirely from the Bangladesh Bank An- nual Report (Bangladesh Bank, 2025), a publicly available statistical and policy publication, used under fair, non-commercial academic research use. No personally identifiable information is involved: the report consists of aggregate macroeconomic, sectoral, and price statistics, and the constructed questionâanswer pairs (Appendix A) similarly con- 9 cern aggregate indicators rather than individuals. No preprocessing beyond the extraction pipeline (Appendix B.2) was applied prior to annotation. We did not seek formal institutional ethics approval, as this work involves no human subjects, no per- sonal data, and no data collection beyond manual annotation of a public government document. Annotator Compensation.Dataset construction and its independent review (Stages 1â2, Ap- pendix A.4) were carried out by the paper authors. Domain validation (Stage 3) and the blind human evaluation of system outputs (Table 9) were car- ried out by two external banking-sector domain experts, professionals with relevant working knowl- edge of this document type who are not employed by Bangladesh Bank. Their participation was vol- untary and uncompensated. Fairness and Bias. Several sources of potential bias are made explicit here. First, B-FinQA-X is drawn from a single institution, language (English), and reporting convention (Section 7); findings should not be assumed to generalize without further evaluation. Second, Asymmetric Evidence Author- ity (Section 3.5) encodes a fixed, hand-specified prior about which modality to trust per claim type, a design choice, not learned or validated against hu- man judgment, and thus a potential source of bias if wrong for a given type; every authority decision is logged against an equal-weight counterfactual in an Authority Docket (Section 3.8), making its actual influence auditable rather than silently ex- ercised. Third, the frameworkâs language model components inherit whatever biases are present in their training and alignment; we do not evaluate these independently. Fourth, retrieval and entail- ment judgments are themselves LLM-mediated, so bias in what a model considers relevant or en- tailed can propagate into which evidence a claim is judged supported by. Risks and Potential Misuse.CLAIR-Finâs out- puts may be inaccurate or hallucinated despite the verification mechanisms in Section 3; the Halluci- nation Risk Index and audit verdicts are calibration signals, not correctness guarantees. Use outside the evaluated domain carries unquantified risk of de- graded faithfulness and abstention behavior, since thresholds and authority weights were designed and evaluated in this single setting (Section 7). Over- reliance, treating a passing verdict or low HRI as a substitute for independently checking cited ev- idence, is a realistic risk, since the audit verdict signals grounding, not external factual correctness. We are not aware of a misuse vector unique to this framework beyond these general LLM-system risks. Societal Impact. Potential benefits. A system that verifies claim-level grounding across modal- ities and abstains when evidence is insufficient could support more reliable access to information in long, statistically dense public documents. Its emphasis on auditable evidence arbitration (Sec- tion 3.8) and selective abstention (Section 3.9) makes automated financial QA more transparent than a system that always answers without indicat- ing confidence. Potential risks.Automation bias, trusting confident-sounding output without independent ver- ification, remains a risk regardless of these safe- guards, particularly for users without domain liter- acy. Performance and abstention behavior are, by construction, unequal across settings beyond the one evaluated (Section 7). Mitigation Strategies. The framework incorpo- rates safeguards directly, not as external add-ons: (1) every published claim carries an explicit citation (Sections 3.1, 3.9); (2) grounding is verified at the hand-off between drafting and adversarial review via Chain-of-Custody Verification, not only at final generation (Section 3.7); (3) a terminal entailment audit gates publication of any unentailed claim (Section 3.8); (4) selective abstention is a first-class, frequently-exercised outcome rather than a fallback (Section 3.9); and (5) the Authority Docket (Sec- tion 3.8) makes the AEA bias risk noted above auditable rather than hidden. Blind human evalua- tion by two external domain experts, independent of dataset construction (Appendix A.1), provides an additional, independent check on outputs. Future Ethical Considerations. Future work should validate the frameworkâs evidence-authority weights and thresholds against human-labeled ground truth rather than treating them as a fixed prior, allowing AEAâs fairness properties to be assessed empirically rather than only made au- ditable. Extending evaluation to other institutions, languages, and reporting conventions (Section 7) is a prerequisite for responsible deployment beyond the evaluated setting. Broader human-centered evaluation involving domain experts and end users, and continued attention to automation bias in real 10 analyst workflows, are natural next steps. References Bangladesh Bank. 2025.Annual report 2024â 2025.https://w.b.org.bd/pub/annual/ anreport/ar2024-2025.pdf. Tianshi Cai, Guanxu Li, Nijia Han, Ce Huang, Zimu Wang, Changyu Zeng, Yuqi Wang, Jingshi Zhou, Haiyang Zhang, Qi Chen, Yushan Pan, Shuihua Wang, and Wei Wang. 2025. FinDebate: Multi-agent collaborative intelligence for financial analysis. In Proceedings of The 10th Workshop on Financial Tech- nology and Natural Language Processing, pages 268â 282, Suzhou, China. Association for Computational Linguistics. Chi-Min Chan, Weize Chen, Yusheng Su, Jianxuan Yu, Wei Xue, Shanghang Zhang, Jie Fu, and Zhiyuan Liu. 2023.Chateval: Towards better llm-based evaluators through multi-agent debate. Preprint, arXiv:2308.07201. Harrison Chase. 2022. Langchain.https://github. com/langchain-ai/langchain. Framework for de- veloping applications powered by large language models. Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan Routledge, and William Yang Wang. 2021. FinQA: A dataset of nu- merical reasoning over financial data. In Proceedings of the 2021 Conference on Empirical Methods in Nat- ural Language Processing, pages 3697â3711, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. 2023. Chain-of-verification reduces hallucination in large language models. Preprint, arXiv:2309.11495. Alphaeus Dmonte, Roland Oruche, Marcos Zampieri, Prasad Calyam, and Isabelle Augenstein. 2025. Claim verification in the age of large language mod- els: A survey. Preprint, arXiv:2408.14317. Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch. 2024. Improving factuality and reasoning in language models through multiagent debate. In Proceedings of the 41st Inter- national Conference on Machine Learning, ICMLâ24. JMLR.org. Shahul Es, Jithin James, Luis Espinosa Anke, and Steven Schockaert. 2024. RAGAs: Automated evalu- ation of retrieval augmented generation. In Proceed- ings of the 18th Conference of the European Chap- ter of the Association for Computational Linguistics: System Demonstrations, pages 150â158, St. Julians, Malta. Association for Computational Linguistics. Chinmay Gondhalekar, Urjitkumar Patel, and Fang- Chun Yeh. 2025.MultiFinRAG: An Optimized Multimodal Retrieval-Augmented Generation Frame- work for Financial Question Answering . In 2025 IEEE International Conference on Big Data (Big- Data), pages 7163â7172, Los Alamitos, CA, USA. IEEE Computer Society. Dongxin Guo, Jikun Wu, and Siu Ming Yiu. 2026. Fin- ground: Detecting and grounding financial hallu- cinations via atomic claim verification. Preprint, arXiv:2604.23588. Shailja Gupta, Rajesh Ranjan, and Surya Narayan Singh. 2024. A comprehensive survey of retrieval- augmented generation (rag):Evolution, cur- rent landscape and future directions.Preprint, arXiv:2410.12837. Siwei Han, Peng Xia, Ruiyi Zhang, Tong Sun, Yun Li, Hongtu Zhu, and Huaxiu Yao. 2025. Mdocagent: A multi-modal multi-agent framework for document understanding. Preprint, arXiv:2503.13964. LangChain Inc. 2024. Langgraph: Build stateful, multi- actor applications with llms.https://github.com/ langchain-ai/langgraph. Software framework for building stateful and multi-agent LLM applica- tions. Pranab Islam, Anand Kannappan, Douwe Kiela, Re- becca Qian, Nino Scherrer, and Bertie Vidgen. 2023. Financebench: A new benchmark for financial ques- tion answering. Preprint, arXiv:2311.11944. Hyewon Jeon and Jay-Yoon Lee. 2025. GraphCheck: Multipath fact-checking with entity-relationship graphs. In Findings of the Association for Compu- tational Linguistics: EMNLP 2025, pages 24728â 24745, Suzhou, China. Association for Computa- tional Linguistics. Lanlan Ji, Dominic Seyler, Gunkirat Kaur, Manjunath Hegde, Koustuv Dasgupta, and Bing Xiang. 2026. PHANTOM: A benchmark for hallucination detec- tion in financial long-context QA. In The Thirty-ninth Annual Conference on Neural Information Process- ing Systems Datasets and Benchmarks Track. Ankur Joshi, Saket Kale, Satish Chandel, and Dinesh Pal. 2015. Likert scale: Explored and explained. British Journal of Applied Science & Technology, 7:396â403. Polina Kirichenko, Mark Ibrahim, Kamalika Chaudhuri, and Samuel J. Bell. 2025. Abstentionbench: Reason- ing llms fail on unanswerable questions. Preprint, arXiv:2506.09038. Potsawee Manakul, Adian Liusie, and Mark Gales. 2023. SelfCheckGPT: Zero-resource black-box hallucina- tion detection for generative large language models. In Proceedings of the 2023 Conference on Empiri- cal Methods in Natural Language Processing, pages 9004â9017, Singapore. Association for Computa- tional Linguistics. 11 Phat Nguyen and Thang Pham. 2026. Toward reliable evaluation of llm-based financial multi-agent systems: Taxonomy, coordination primacy, and cost awareness. Preprint, arXiv:2603.27539. OpenAI. 2024a. Hello gpt-4o.https://openai.com/ index/hello-gpt-4o/. Accessed: 30 June 2026. OpenAI. 2024b.New embedding models and api updates.https://openai.com/index/ new-embedding-models-and-api-updates/. Accessed: 2026-08-09. OpenAI. 2025. Gpt-4.1 mini.https://openai.com/ index/gpt-4-1/. Accessed: 30 June 2026. Saisab Sadhu, Biswajit Patra, and Tanmay Basu. 2025. Structured adversarial synthesis: A multi-agent framework for generating persuasive financial analy- sis from earnings call transcripts. In Proceedings of The 10th Workshop on Financial Technology and Nat- ural Language Processing, pages 283â291, Suzhou, China. Association for Computational Linguistics. Likun Tan, Kuan-Wei Huang, and Kevin Wu. 2025. Fred: Financial retrieval-enhanced detection and edit- ing of hallucinations in language models. Preprint, arXiv:2507.20930. Jianguo Wang, Xiaomeng Yi, Rentong Guo, Hai Jin, Peng Xu, Shengjun Li, Xiangyu Wang, Xiangzhou Guo, Chengming Li, Xiaohai Xu, Kun Yu, Yux- ing Yuan, Yinghao Zou, Jiquan Long, Yudong Cai, Zhenxiang Li, Zhifeng Zhang, Yihua Mo, Jun Gu, and 3 others. 2021. Milvus: A purpose-built vec- tor data management system. In Proceedings of the 2021 International Conference on Management of Data, SIGMOD â21, page 2614â2627, New York, NY, USA. Association for Computing Machinery. Xingqi Wang, Yiming Cui, Xin Yao, Shijin Wang, Guoping Hu, and Xiaoyu Qin. 2025. Charthal: A fine-grained framework evaluating hallucination of large vision language models in chart understanding. Preprint, arXiv:2509.17481. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-consistency improves chain of thought reasoning in language models. Preprint, arXiv:2203.11171. Bingbing Wen, Jihan Yao, Shangbin Feng, Chenjun Xu, Yulia Tsvetkov, Bill Howe, and Lucy Lu Wang. 2025. Know your limits: A survey of abstention in large language models. Transactions of the Association for Computational Linguistics, 13:529â556. Qianqian Xie, Weiguang Han, Zhengyu Chen, Ruoyu Xiang, Xiao Zhang, Yueru He, Mengxi Xiao, Dong Li, Yongfu Dai, Duanyu Feng, Yijing Xu, Haoqiang Kang, Ziyan Kuang, Chenhan Yuan, Kailai Yang, Zheheng Luo, Tianlin Zhang, Zhiwei Liu, Guojun Xiong, and 15 others. 2024. Finben: a holistic fi- nancial benchmark for large language models. In Proceedings of the 38th International Conference on Neural Information Processing Systems, NIPS â24, Red Hook, NY, USA. Curran Associates Inc. Siqiao Xue, Xiaojing Li, Fan Zhou, Qingyang Dai, Zhixuan Chu, and Hongyuan Mei. 2025. Famma: A benchmark for financial domain multi- lingual multimodal question answering. Preprint, arXiv:2410.04526. Xinqi Yang, Scott Zang, Yong Ren, Dingjie Peng, and Zheng Wen. 2024. Evaluating large language mod- els on financial report summarization: An empirical study. Preprint, arXiv:2411.06852. Zhihan Zhang, Yixin Cao, and Lizi Liao. 2025. Xfinbench: Benchmarking llms in complex finan- cial problem solving and reasoning.Preprint, arXiv:2508.15861. Suifeng Zhao, Zhuoran Jin, Sujian Li, and Jun Gao. 2025. Finragbench-v: A benchmark for multimodal rag with visual citation in the financial domain. Preprint, arXiv:2505.17471. Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and Tat- Seng Chua. 2021. TAT-QA: A question answering benchmark on a hybrid of tabular and textual con- tent in finance. In Proceedings of the 59th Annual Meeting of the Association for Computational Lin- guistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3277â3287, Online. Association for Computational Linguistics. Appendix A B-FinQA-X: Dataset Construction and Validation A.1 Evaluation Protocol We evaluate CLAIR-Fin on B-FinQA-X (Ap- pendix A), 500 questions stratified by query type, format, and difficulty. The audit verdict (Sec- tion 3.8) is an internal, per-claim signal; a separate harness maps each runâs answer and abstention flag to a four-way outcome (Correct, Partial, Incorrect, Abstained) against gold references, the outcome space we report. Automatic evaluation uses RAGAS (Es et al., 2024) (context precision, recall, faithfulness, an- swer relevancy) via a GPT-4.1 mini judge (OpenAI, 2025), distinct from the GPT-4o backbone (Ope- nAI, 2024a), which also assigns the four-way label. We additionally report exact correctness, answer coverage, the Authority Docketâs changed-outcome rate (Equation(7)vs. uniform counterfactual), de- bate utilization rate, and HRI calibration (correla- tion with gold-label correctness). 12 Human evaluation: the same two external banking-sector domain experts who performed Stage 3 dataset validation (Appendix A.4), and who are independent of the two paper authors who constructed the dataset and wrote its gold answers, rate all 500 answers on a five-point scale across six dimensions (correctness, faithfulness, citation quality, clarity, abstention appropriateness, overall quality), blind to confidence/HRI, with agreement reported as quadratic weighted CohenâsÎş. Because these evaluators did not author the gold answers or the dataset itself, their ratings are not confounded by familiarity with items they personally wrote. Ablation. We disable one mechanism at a time, AEA (uniform weighting), CoCV, ARC (routing every escalated claim to audit after one draft), and the terminal entailment audit (w/o Term. Audit), plus a single-pass RAG baseline, isolating each componentâs marginal contribution. A.2 Data Sources and Selection Criteria The source document is the Bangladesh Bank An- nual Report (Bangladesh Bank, 2025), a publicly available statistical and policy publication issued by Bangladeshâs central bank, used here under fair, non-commercial academic research use; no propri- etary or personally identifiable data is involved. We drew questions from material in the report se- lected against two criteria: (1) high density of narra- tive, tables, and charts describing the same underly- ing economic indicators, the condition under which cross-modal evidence conflicts are most likely to arise; and (2) self-containment, so a grounded ques- tion does not require evidence from outside the selected material. Material without comparable multimodal density was excluded from question sourcing. No automated preprocessing was applied; annotators worked directly from the original report text, tables, and charts, so every question and an- swer traces to an unaltered source passage. A.3 Dataset Construction Pipeline Dataset construction proceeded in four chronologi- cal steps. Step 1, Source Selection. Source material (Ap- pendix A.2) was selected per the two inclusion cri- teria: cross-modal density around shared indicators, and self-containment. Step 2, Drafting. Questions and gold answers were written directly against the source by Author 1 (Appendix A.4, Stage 1). Each item records its supporting evidence, the specific text passage, table cell, or chart element it depends on, together with preliminary labels (query type, difficulty, format), drafted toward the balanced target distribution re- ported in Appendix A.7. Step 3, Validation. Every drafted item passed through the three-stage annotation process (Ap- pendix A.4): independent review by Author 2, domain validation by the two external banking- sector domain experts (Reviewers 3â4), and manda- tory joint consensus resolution of any flagged item. Items were revised or discarded at this step rather than included as originally drafted. Step 4, Release. A final quality-control pass (Appendix A.6) verified that the released 500 items satisfied the target balance across all three annota- tion dimensions (Appendix A.7) before the dataset was frozen. A.4 Manual Annotation Protocol B-FinQA-X was manually constructed and vali- dated through a three-stage process involving four annotators: two paper authors, who drafted and independently reviewed every item, and two exter- nal banking-sector domain experts, professionals with relevant working knowledge of this document type, independent of Bangladesh Bank itself, who jointly performed Stage 3 domain validation and, separately, the blind human evaluation of system outputs reported in Table 9 (Appendix A.1). Nei- ther domain expert was compensated for this work; their participation was voluntary. (i) Stage 1, Dataset Construction (Author 1). The first author read the source material and con- structed questionâanswer pairs covering a range of information needs, recording each itemâs support- ing evidence and preliminary labels for query type, presentation format, and difficulty, drafted toward the balanced target distribution (Appendix A.7). (i) Stage 2, Independent Review (Author 2). The second author independently reviewed every item, checking question clarity, answer correctness, evidence grounding, and label correctness against a shared annotation guideline (Appendix A.5). Dis- crepancies were recorded for later discussion rather than resolved unilaterally. (i) Stage 3, Domain Validation (Reviewers 3â4). Two independent banking-sector domain ex- perts, external to Bangladesh Bank, reviewed the full dataset for financial correctness, banking ter- minology, numerical accuracy, faithfulness to the source, practical relevance, and difficulty-label con- sistency, flagging items with ambiguous wording, 13 chapter: Chapter 3, chapter_title: Price and Inflation, question: Among SAARC countries, which country is projected to have the highest inflation rate in 2025 and what is that rate?, question_type: Fact Extraction, reasoning_skill: Retrieval, evidence_source: Table, difficulty: Easy, answer: Bangladesh is projected to have the highest inflation rate among SAARC countries at 10.0%. chapter: Chapter 3, chapter_title: Price and Inflation, question: Describe the path of point-to-point headline inflation in Bangladesh during FY25., question_type: Trend Analysis, reasoning_skill: Temporal Reasoning, evidence_source: Text + Chart, difficulty: Medium, answer: Point-to-point headline inflation started at a high level, peaked near the beginning of FY25, and then gradually declined to 8.48% in June 2025, compared with 9.72% in June 2024. The point-to-point non-food inflation increased from 9.15 percent in June 2024 to 9.37 percent in June 2025 with fluctuations throughout FY25. However, point-to-point non-food inflation remained above 9.0 percent during FY25 (Chart 3.04). Supply-side constraints and substantial rise in import costs continued to reflect non-food inflation. chapter: Chapter 3, chapter_title: Price and Inflation, question: What is the projected path of inflation in advanced economies from 2024 to 2026?, question_type: Fact Extraction, reasoning_skill: Retrieval, evidence_source: Text, difficulty: Easy, answer: Inflation in advanced economies is projected to decline from 2.6% in 2024 to 2.5% in 2025 and 2.2% in 2026. Inflation is projected to reach target levels in advanced economies from 2.6 percent in 2024 to 2.5 percent in 2025 and 2.2 percent in 2026, where strong policy credibility and softer demand momentum support the disinflationary process. At the same time, the decline is expected to be gradual rather than sharp, reflecting persisting upward pressures in some advanced economies due to tariffs, service sector costs, and regulated price changes. Nonetheless, the broader trend indicates a return toward pre-pandemic patterns, albeit with risks from trade tensions, financial volatility, and policy uncertainty. Inflation in emerging market and developing economies is projected to ease gradually, reaching around 5.3 percent in 2025 and 4.7 percent by 2026 (Table 3.07), though it will remain higher than in advanced economies. The overall trend reflects softer demand conditions in some economies, alongside tighter monetary policies, but the outlook is uneven across regions. Figure 3: Three B-FinQA-X examples from the Price and Inflation chapter, grounded in a table, a chart, and narrative text respectively, each showing the source excerpt, its recorded annotation fields (Table 4), and the gold answer, with the answerâs supporting figures highlighted in the source. Table 4: Fields recorded for each of the 500 items in B-FinQA-X, including the three controlled annotation dimensions, query type, difficulty, and presentation format, used to stratify the dataset and report disaggregated results in Section 6. FieldDescription QuestionThe natural-language question posed against the source report. AnswerThe gold answer, written to be verifiable directly against the recorded evidence. EvidenceThe specific source passage, table cell(s), or chart element(s) the answer depends on. Source page(s)The page or page range in the source report where the evidence appears. Query typeOne of six categories, defined in (a) below. DifficultyOne of three levels, defined in (b) below. Presentation formatOne of six modality categories, defined in (c) below. ChapterSource chapter (Macroeconomic Performance and Prospects; Real Sector Trends; Price and Inflation). incorrect terminology, or weak evidentiary support. Consensus Resolution. All flagged items were jointly discussed by all four annotators and re- solved against the shared guideline rather than by majority vote or a single adjudicator; the agreed label became the final gold-standard annotation. Because agreement was enforced procedurally through mandatory joint resolution of every flagged item rather than parallel independent labeling by all four annotators, we do not report a chance- corrected agreement statistic (for example, Fleissâ Îş); the protocol was designed to eliminate residual disagreement prior to release rather than to measure it post hoc. A.5 Annotation Schema and Guidelines Items are labeled along three dimensions: query type (Fact Extraction, Comparison, Trend Anal- ysis, Numerical Calculation, Multi-hop Reason- ing, Evidence Retrieval), difficulty (Easy, Medium, Hard), and presentation format (Text Only, Table Only, Chart Only, Text + Table, Text + Chart, Ta- ble + Chart). Evidence Retrieval requires identi- fying and grounding the relevant evidence itself rather than being handed a pre-identified passage or cell; Multi-hop Reasoning requires combining evidence from more than one location or modality. Figure 3 illustrates this schema with three worked examples, one per evidence format, drawn from the Price and Inflation chapter. A.6 Quality Control and Validation Every item was checked against a fixed criterion set during review (Appendix A.4): (1) answer correct- ness; (2) evidence correctness; (3) absence of am- biguity; (4) absence of duplicate questions; (5) nu- merical accuracy; (6) query-type label correctness; (7) difficulty label correctness; and (8) presentation- format label correctness. Items failing any criterion were revised and re-checked rather than included 14 as originally drafted; items that could not be re- vised to satisfy all criteria were discarded. A final quality-control pass over the complete, revised set confirmed that the released dataset satisfied the tar- get distribution across all three annotation dimen- sions (Appendix A.7) and that annotations were consistent across all 500 items before the dataset was frozen. Table 5 gives the query-type-by-difficulty dis- tribution; Table 6 below gives the corresponding breakdown by presentation format. Table 5: Distribution of the 500 B-FinQA-X items by query type and difficulty level. Query TypeEasyMediumHardTotal Fact Extraction706515150 Comparison357525135 Trend Analysis20351065 Numerical Calculation15351060 Multi-hop Reasoning10301050 Evidence Retrieval2510540 Total17525075500 Table 6: Distribution of the 500 B-FinQA-X items by presentation format and difficulty level. Presentation FormatEasyMediumHardTotal Text Only355015100 Table Only355015100 Chart Only1825750 Text + Table527523150 Text + Chart1825750 Table + Chart1725850 Total17525075500 A.7 Dataset Statistics B-FinQA-X contains 500 questionâanswer pairs drawn from the Bangladesh Bank Annual Report, in English, spanning all six query types, three dif- ficulty levels, and six presentation formats (Ap- pendix A.5). Key balance properties: (i)Difficulty balance. The dataset is weighted toward Easy and Medium items (175 and 250 of 500), with Hard items deliberately kept to a smaller share (75), reserved for multi-hop, cross-modal, or computation-heavy items rather than an equal split. (i) Query type coverage. Coverage ranges from 150 items (Fact Extraction, the largest cate- gory) to 40 items (Evidence Retrieval, the smallest), reflecting each reasoning typeâs rel- ative prevalence in the source rather than an artificially uniform split. (i)Presentation format coverage. Coverage ranges from 150 items (Text + Table, the most common evidence combination in the source) to 50 items each for Chart Only, Text + Chart, and Table + Chart. (iv)Completeness. Every combination of query type, presentation format, and difficulty de- fined in Appendix A.5 that occurs in the source report is represented by at least one item. Table 12 (Appendix B.2) gives the full query- type-by-presentation-format cross-tabulation, with the Easy/Medium/Hard split for every cell. B Experimental Setup We describe the implementation stack, ingestion pipeline, and inference configuration used to pro- duce the results in Section 6, before presenting the methodology in Section 3. This distinguishes what each mechanism computes from which library or model instantiates it, while specifying the imple- mentation details underlying our experiments. The complete codebase, prompts, and inference con- figurations will be released to support transparent evaluation and further research. B.1 Implementation Framework and Multi-Agent Orchestration CLAIR-Fin is implemented in Python and orches- trated as a LangGraphStateGraph(Inc., 2024). Each of the nine agents (Planner-Orchestrator; Narrative, Tabular, and Visual Evidence; Ledger Guardian; Affirmative and Adversarial Counsel; Judge-Auditor; and Brief Synthesizer) is a node over one shared, typed graph state, with condi- tional edges implementing the coverage-based es- calation route (Section 3.5) and the severity-gated rebuttal loop (Section 3.6). A Chain-of-Custody Verification checkpoint (Section 3.7) sits between drafting and adversarial review, reusing the Judge- Auditorâs entailment call rather than a tenth agent, with conditional routing keeping this non-linear flow explicit. LangChain (Chase, 2022) supplies the integration layer:Documentis the shared evi- dence unit;ChatOpenAI/OpenAIEmbeddingswrap chat/embedding calls;Milvuswraps the vector store (Appendix B.3). All typed-output calls use 15 Table 7: Framework-specific metrics for CLAIR-Finâs full system on B-FinQA-X (n = 500), covering answer outcomes, claim-level faithfulness and correctness, HRI calibration, and how often the debate (ARC) and authority- weighting (AEA) mechanisms are exercised. #MetricDescriptionOverall (n = 500) 1Answer Outcome DistributionPercentage of answers classified as Correct / Partial / Incorrect / Abstained 59.2% / 23.6% / 11.8% / 5.4% (296 / 118 / 59 / 27) 2Faithfulness RateâFraction of published claims passing citation-entailment verification 0.783 3Exact Correct Answer Rateâ Fraction of questions answered completely correctly 0.592 4Answer CoverageâFraction of questions receiving a non- abstained answer 0.946 5HRI CalibrationCorrelation between Hallucination Risk In- dex and answer correctness; more negative is better calibrated â0.072 6Debate Utilization Rate Fraction of instances routed through the de- bate stage 0.646 7AEA Impact RateFraction of decisions altered by asymmetric evidence aggregation 0.515 Table 8: Retrieval- and generation-quality metrics for CLAIR-Fin on B-FinQA-X (n = 500), combining RAGAS- style scores (context precision/recall, faithfulness, answer relevancy) with standard retrieval-ranking metrics (Hit Rate@8, MRR, Recall@8). #MetricCategoryWhat It MeasuresOverall (n = 500) 1Context PrecisionâRetrievalFraction of retrieved chunks that are relevant0.816 2Context RecallâRetrievalFraction of required evidence successfully retrieved0.897 3FaithfulnessâFaithfulnessDegree to which generated answers are supported by retrieved evidence0.889 4Answer RelevancyâGenerationDegree to which the generated answer addresses the user query0.696 5Context RelevancyâRetrievalTopical relevance of retrieved evidence0.639 6Hit Rate@8 / MRRâRetrievalPresence and ranking of relevant evidence within the top-8 retrieved results0.950 / 0.822 7Recall@8âRetrievalFraction of gold evidence retrieved within the top-8 results0.903 structured-output mode against Pydantic schemas, with a safe-default fallback (Section 3.3) on parse failure. B.2 Document Ingestion and Chunking Source PDFs are parsed page-by-page with PyMuPDF (fitz). Since native text extraction is unreliable for dense financial tables and charts, ev- ery page is also rendered to an image and passed to a vision-capable LLM that extracts tables and de- scribes charts directly (Section 3.4). Narrative text is chunked with a sentence-aware splitter: bound- aries are detected with a regex protecting common abbreviations (âDec.â, âapprox.â, âe.g.â), chunks grow to a 1,000-character target, and each car- ries the last two sentences of the previous chunk as overlap, guaranteeing every boundary falls on a sentence end, which matters for embedding quality and the entailment gate (Section 3.8). Tables and chart descriptions are embedded as their own docu- ments, separate from narrative chunks, so retrieval- time modality filtering (Appendix B.4) can address them independently. B.3 Embedding Model and Vector Database Allchunks,tablecells,andchartde- scriptionsareembeddedwithOpenAIâs text-embedding-3-large(3,072-dimensional, default dimensionality, no truncation) (OpenAI, 2024b); queries use the same model, sharing one vector space. Vectors are stored in Milvus Lite (Wang et al., 2021), an embedded, serverless, single-file mode requiring no external database infrastructure. Each entry holds its embedding plus a JSONmetadatafield with modality (text/table/chart), source document, page num- ber, and modality-specific attributes, letting one collection serve all three modalities and retrieval filter bymetadata[âmodalityâ](Appendix B.4). Similarity search uses Milvusâs default L2 distance index. B.4 Retrieval Configuration Retrieval blends dense and lexical signal rather than vector similarity alone (Equation(3), Section 3.4): a candidate pool four times the requested size (4k) is drawn by vector search and modality filter, then reranked by a weighted sum of normalized vec- tor similarity (1/(1 + L2 distance), weight 0.65) and stopword-filtered lexical term overlap (weight 0.35); the topkpassages are returned. Per-agent width is fixed: Narrativek=8, Tabulark=6, Vi- sualk=5(Section 3.4). The lexical component was added after observing lexically similar but 16 According to the IMF's World Economic Outlook (October 2025), what is the projected growth in global output for 2025 and 2026? chapter: Chapter 1 question_type: Fact Extraction reasoning_skill: Retrieval evidence_source: Text + Table difficulty: Easy Retrieved Contexts (1) Table 1.01 presents the IMF's World Economic Outlook projections for global output growth. It reports that world output grew by 3.5% in 2023 and 3.3% in 2024, and is projected to grow by 3.2% in 2025 and 3.1% in 2026. (2) According to the IMF's World Economic Outlook (WEO) published in October 2025, global output growth is expected to reach 3.2% in 2025 and 3.1% in 2026, following a growth rate of 3.3% in 2024, as shown in Table 1.01. Generated Response According to the IMF's World Economic Outlook (October 2025), the projected growth in global output is 3.2 percent for 2025 and 3.1 percent for 2026. Gold Answer 3.2 percent in 2025 and 3.1 percent in 2026 How did private investment as a percentage of GDP change in FY25, and what reasons are cited? chapter: Chapter 2 question_type: Comparison reasoning_skill: Comparative evidence_source: Text difficulty: Medium Retrieved Contexts (1) Private investment to GDP ratio further declined to 22.48 percent in FY25 from 23.96 percent in FY24, due mainly to a lack of confidence among the investors because of social and political unrest, and the interim administration. (2) Private investment to GDP ratio decreased slightly to 22.48 percent from 23.96 percent over the same period. Generated Response In FY25, private investment as a percentage of GDP decreased to 22.48 percent, down from 23.96 percent in FY24. This decline is primarily attributed to a lack of confidence among investors, which stems from social and political unrest, as well as the interim administration. Gold Answer Private investment declined from 23.96% of GDP in FY24 to 22.48% in FY25, mainly due to weakened investor confidence caused by social and political unrest and the transition to the interim administration. How did Bangladesh's real GDP growth rate change from FY24 to FY25 according to BBS estimates? chapter: Chapter 1 question_type: Trend Analysis reasoning_skill: Temporal evidence_source: Chart + Text + Table difficulty: Easy Retrieved Contexts (1) According to the Bangladesh Bureau of Statistics (BBS), the real GDP growth rate for FY25 is estimated to be 3.97 percent, down from 4.22 percent in FY24. (2) Chart 2.02, Trends in Bangladesh Real GDP Growth, shows that GDP growth was approximately 4.5 percent in FY24 and approximately 4.0 percent in FY25. (3) Table 1.02 reports the GDP (at constant market price) growth rates as 4.22 percent in FY24 and 3.97 percent in FY25. Generated Response According to estimates from the Bangladesh Bureau of Statistics (BBS), Bangladesh's real GDP growth rate decreased from 4.22 percent in FY24 to 3.97 percent in FY25 Gold Answer declined from 4.22 percent in FY24 to 3.97 percent in FY25. Figure 4: Three representative B-FinQA-X cases spanning different query types, evidence modalities, and difficulty levels: a Fact Extraction query grounded in text and a table (Easy), a Comparison query grounded in text alone (Medium), and a Trend Analysis query grounded in a chart, text, and a table (Hard). For each case we show the retrieved evidence, CLAIR-Finâs generated response, and the gold answer, with the cited figures highlighted across all three. In the third case, the chartâs approximate readings (4.5% and 4.0%) diverge from the tableâs exact figures (4.22% and 3.97%); the generated response follows the table and text rather than the chart, consistent with the Asymmetric Evidence Authority weighting described in Section 3.5. distinct aggregates (e.g., âoverall balance of pay- mentsâ vs. âcurrent account balanceâ) could sit near-equally close in embedding space; blending in lexical overlap corrects this without sacrificing dense retrievalâs recall. B.5 Language Models and Inference Settings All nine agents share GPT-4o as the chat back- bone, so ablation differences (Appendix A.1) re- flect mechanism design, not model change; the Chain-of-Custody checkpoint reuses this backbone via the Judge-Auditorâs entailment call. Vision- based table/chart extraction (Appendix B.2) also uses GPT-4o, called once per page at ingestion. Temperature is fixed per call site: 0 for claim de- composition and the terminal auditâs draft; 0.2 for affirmative drafting/rebuttal (Section 3.6) and an- swer composition (Section 3.9). Output-token bud- gets are capped per agent (e.g., 700 for affirma- tive/adversarial calls, 400 for the audit, 300 for entailment checks). B.6 Evaluation Tooling and Runtime Environment For each run, automatic retrieval and generation metrics are computed from the cited evidence and generated answer. The evaluation covers context precision, context recall, faithfulness, and answer relevancy. Framework-specific measures are de- rived from the Financial Claim Ledger, custody log, and Authority Docket persisted during each run (Section 3.9), with exact correctness, answer cover- age, AEA impact rate, debate utilization, and HRI calibration used for evaluation. A cost-tracking callback records token usage and cost for each API call. The pipeline is fully API-based, with Milvus Lite serving as the only local service. B.7 Configuration Management Every threshold and weight fixed in Section 3, the coverage-escalation cutoff (0.75), the Adaptive Re- buttal Cycleâs round cap (Ď max =2), the terminal entailment pass bar (0.5), the AEA weight table (Table 1), and the HRIâs term weights, is exter- nally configurable rather than hardcoded. A single pydantic-settingsobject is the sole source of truth for API keys, model names, and filesystem 17 Table 9: Human evaluation of CLAIR-Finâs 500 generated answers by two external, uncompensated banking-sector domain experts (Appendix A.4), independent of the paper authors who constructed the dataset, blind to each otherâs ratings and to the systemâs confidence scores and HRI. Each dimension is scored on a 1â5 scale (mean reported per evaluator) (Joshi et al., 2015), with inter-annotator agreement given as quadratic weighted Cohenâs Îş. DimensionDescriptionEvaluator 1Evaluator 2Îş CorrectnessâDoes the answer accurately reflect the source documents? 4.184.050.82 Human-rated FaithfulnessâAre all claims supported by the cited evi- dence? 4.344.210.85 Citation QualityâAre the cited sources appropriate and suf- ficient? 4.113.980.80 ClarityâIs the answer coherent, well-structured, and easy to understand? 4.394.270.87 Abstention Appropriatenessâ Was abstention the correct decision when used? 4.063.950.84 Overall QualityâOverall assessment considering correct- ness, faithfulness, citation quality, clarity, and abstention behavior 4.224.090.84 Table 10: RAGAS metrics on B-FinQA-X (n = 500), disaggregated by evidence presentation format. Overall scores are computed using sample-weighted aggregation across all presentation formats based on their respective sample counts. Presentation FormatSamplesFaithfulnessâAnswer RelevancyâContext PrecisionâContext Recallâ Text Only1000.8700.6750.7950.878 Table Only1000.9000.7050.8300.912 Chart Only500.8500.6650.7750.847 Text + Table1500.9150.7250.8450.925 Text + Chart500.8750.6850.8050.887 Table + Chart500.8800.6750.7950.881 Overall (n = 500)5000.8890.6960.8160.897 paths. Two YAML files carry mechanism-level val- ues: an agent-budgets file (thresholds, round/repair budgets, token caps, the Adversarial Counselâs at- tack taxonomy) and a separate AEA-weights file. Each agent reads these from the shared settings object rather than an inlined value, so a threshold change is a configuration edit, not a code change, making the ablation protocol (Appendix A.1) prac- tical: each ablated mechanism is realized by editing one value (or, for Chain-of-Custody Verification, removing one routing edge), keeping every run on the same code path except the value under test. C Supplementary Results This section reports the framework-specific metrics (Table 7), qualitative examples (Figure 4), general retrieval- and generation-quality metrics (Table 8), human evaluation (Table 9), and per-format and per- query-type breakdowns (Figures 5â7, Tables 10â 11) referenced from Section 6. 18 FaithfulnessAnswer RelevancyContext PrecisionContext Recall 0.0 0.2 0.4 0.6 0.8 1.0 Score 0.889 0.696 0.816 0.897 0.780 0.680 0.700 0.840 0.874 0.691 0.801 0.885 0.865 0.688 0.752 0.831 0.832 0.694 0.729 0.889 Comparison of CLAIR-Fin against retrieval baselines across faithfulness, relevancy, and context metrics CLAIR-FinVanilla RAGHyDE RAGHierarchical RAGGraph-RAG Figure 5: Performance comparison between CLAIR-Fin and four retrieval-strategy baselines on B-FinQA-X. Across the evaluated RAGAS metrics, CLAIR-Fin obtains the strongest overall performance, with a faithfulness score of 0.889 and the highest context precision and context recall. Answer relevancy remains broadly comparable across the evaluated systems (Table 2). Chart Only Text Only Text + Chart Table + Chart Table Only Text + Table 0.70 0.75 0.80 0.85 0.90 Score Faithfulness, relevancy, and context metrics across text, table, and chart formats FaithfulnessAnswer RelevancyContext PrecisionContext Recall Figure 6: Presentation format-wise evaluation of CLAIR-Fin across Text, Table, and Chart evidence formats on B-FinQA-X. Combined formats improve faithfulness, ranging from 0.850 for Chart Only to 0.915 for Text + Table (Table 10). 19 FaithfulnessAnswer RelevancyContext PrecisionContext Recall 0.0 0.2 0.4 0.6 0.8 1.0 Score 0.920 0.735 0.855 0.928 0.900 0.710 0.825 0.908 0.885 0.685 0.805 0.892 0.865 0.665 0.785 0.875 0.840 0.625 0.755 0.835 0.839 0.656 0.780 0.862 Retrieval and generation quality across query types Fact Extraction Comparison Trend Analysis Numerical Calculation Multi-hop Reasoning Evidence Retrieval Figure 7: Query-type analysis of CLAIR-Fin on B-FinQA-X across retrieval and generation metrics. Fact Extraction attains the highest faithfulness (0.920), followed by Multi-hop Reasoning (0.840) and Evidence Retrieval (0.839) (Table 11). Table 11: RAGAS metrics on B-FinQA-X (n = 500), disaggregated by query type. Results are reported across six query categories, with the overall score computed using sample-weighted aggregation based on the number of instances in each category. Query TypeSamplesFaithfulnessâAnswer RelevancyâContext PrecisionâContext Recallâ Fact Extraction1500.9200.7350.8550.928 Comparison1350.9000.7100.8250.908 Trend Analysis650.8850.6850.8050.892 Numerical Calculation600.8650.6650.7850.875 Multi-hop Reasoning500.8400.6250.7550.835 Evidence Retrieval400.8390.6560.7800.862 Overall (n = 500)5000.8890.6960.8160.897 Table 12: Distribution of the 500 B-FinQA-X items by query type and presentation format. Each cell shows Easy / Medium / Hard = Total. Text Only and Table Only share identical distributions by design (matched pairs), as do Chart Only and Text + Chart. Text + Table and Table + Chart were constructed independently. Single ModalityChartMulti-ModalityChart+ Query TypeText OnlyTable OnlyChart OnlyText+TableText+ChartTable+Chart Fact Extraction14/12/3=2914/12/3=296/7/2=1522/20/4=466/7/2=158/7/1=16 Comparison7/14/5=267/14/5=264/8/2=1410/23/8=414/8/2=143/8/3=14 Trend Analysis4/8/2=144/8/2=142/3/1=66/10/3=192/3/1=62/3/1=6 Numerical Calculation3/8/2=133/8/2=132/3/1=64/10/3=172/3/1=61/3/1=5 Multi-hop Reasoning2/6/2=102/6/2=101/3/1=53/9/3=151/3/1=51/3/1=5 Evidence Retrieval5/2/1=85/2/1=83/1/0=47/3/2=123/1/0=42/1/1=4 Format Total35/50/15=10035/50/15=10018/25/7=5052/75/23=15018/25/7=5017/25/8=50 Overall: 175 Easy + 250 Medium + 75 Hard= 500 Cell format: Easy / Medium / Hard = Total 20 D Detailed Analysis of Research Questions RQ1: Modality-Aware Evidence Prioritization In Table 2, disabling AEA lowers (â) faithfulness (0.889 â 0.883) and context recall (0.897 â 0.893); Table 3 shows exact correct answer rate falling (0.592â 0.585). Table 7âs Authority Docket reports an AEA impact rate of 0.515, meaning the asymmetric prior changes which modalityâs account is followed in roughly half of all scored decisions. The consistency of the drop across four metrics, rather than a large drop in one, suggests AEAâs effect is distributed evenly, correcting many small cross-modal disagreements rather than a few large ones, consistent with the mechanismâs design (Section 3.5) as a per-claim argmax over modality-weighted confidence rather than a global override. The 0.515 impact rate is the more decisive evidence: an equal-weight and an asymmetric-weight decision differ in outcome for roughly half of claims where multiple modalities compete, meaning the authority prior materially changes what the system reports as fact in a substantial fraction of contested cases. A natural objection is that a 0.006 faithfulness gap is modest. We read this as a floor rather than a ceiling: AEA only exercises influence when modalities actually compete, and the impact rate confirms this happens often enough (51.5% of scored decisions) that the aggregate effect is measurable and directionally consistent rather than noise. The improvement in context precision and recall alongside faithfulness is notable because AEA does not directly touch retrieval: its effect is mediated entirely through which evidence the drafting stage is shown and in what order, indicating authority weighting shapes not just what is stated but what is treated as relevant. Conditioning evidence trust on claim type therefore measurably and consistently improves faithfulness and correctness across a substantial share of genuinely contested cross-modal decisions rather than a small number of edge cases. RQ2: Verification at the Drafting-to-Review Hand-off Table 2 shows faithfulness falling (â) to0.889â 0.845without the terminal audit and0.889â 0.857 without CoCV, a drop ofâ0.044versusâ0.032. Table 3 shows the same ordering for faithfulness rate:0.783 â 0.741without the terminal audit versus0.783 â 0.753without CoCV; neither ablated configuration recovers full-system performance (0.889, Table 8). This ordering matches the two mechanismsâ architectural roles (Sections 3.7â3.8): the terminal audit is the single mandatory checkpoint every claim must pass before publication, whereas CoCV intervenes only at one hand-off, with a bounded single repair. Removing the terminal audit removes the only backstop guaranteed to catch a claimâs problems however they arose, including problems CoCV was never positioned to catch, since CoCV verifies grounding between drafting and adversarial review specifically, not the state of a claim after debate revises it. Removing CoCV instead still leaves the terminal audit in place to catch much of what CoCV would have caught mid-pipeline, consistent with its smaller cost. That CoCVâs removal still produces a non-trivial cost despite the terminal audit remaining active indicates the audit alone does not fully substitute for hand-off-level checking: an ungrounded draft CoCV would have repaired can still shape the adversarial findings it receives. This finding has a direct architectural implication: the point at which faithfulness is checked is not interchangeable. A mandatory, comprehensive final gate is the larger single contributor to faithfulness in this architecture, yet the residual gap between removing CoCV and matching full- system performance shows hand-off-level checking still contributes measurably on top of a strong terminal gate, rather than being made redundant by it. The two checks are therefore complementary rather than one subsuming the other. 21 RQ3: Adaptive Allocation of Adversarial Debate Table 2 reports faithfulness falling (0.889 â 0.770) without debate, a larger drop than removing AEA, CoCV, or the terminal audit individually. Table 3 shows exact correct answer rate falling (0.592â 0.524) and answer coverage falling (0.946â 0.896). Table 7 reports a debate utilization rate of 0.646, meaning nearly two-thirds of claims are routed through debate rather than fast-pathed. This matches how escalation is decided (Section 3.5): a claim only enters debate when its coverage score falls below the fast-path threshold, so debate is specifically reserved for claims where evidence is weakest or most contested. Removing debate therefore removes the only verification opportunity for exactly the claims most likely to be wrong, since those are, by construction, the claims coverage- based escalation identified as needing it; the effect is concentrated there rather than spread uniformly. The coverage drop from 0.946 to 0.896 is a secondary signal: without debate, more claims that would previously have been contested, revised, and supported are instead drafted once, fail the terminal gate on the first attempt, and are abstained rather than resolved. An important nuance is that this result should not be read as âmore debate is always betterâ: the framework does not debate every claim, and debate depth is designed to track difficulty (Section 3.6) rather than being applied indiscriminately. The magnitude of this ablationâs effect is evidence for the value of targeted debate on genuinely contested claims, not evidence that exhaustive debate on every claim would perform better; that is not tested here, and it would carry a proportional cost this experiment does not isolate. Adaptive adversarial debate is thus the single most consequential mechanism evaluated, precisely because it is allocated to the claims most likely to fail without it. RQ4: Continuous Risk Estimation versus Binary Gating Table 7 reports HRI calibration ofâ0.072alongside a faithfulness rate of 0.783, the binary measure HRI is intended to complement. Table 9 reports human-rated abstention appropriateness of 4.06 and 3.95 across the two evaluators, with quadratic weighted CohenâsÎşof 0.84, among the higher end of the six rated dimensions, tied with Overall Quality and behind only Clarity (0.87) and Human-rated Faithfulness (0.85). A negative HRI-correctness correlation is the theoretically expected direction: higher predicted risk should coincide with lower observed correctness, confirming HRI is not noise uncorrelated with claim quality. The modest magnitude is consistent with what HRI measures: Section 3.8 defines it as a combination of entailment confidence, authority score, custody repairs, and adversarial attack severity, several of which already gate the binary verdict (Equation (6)); HRI is correlated with, but deliberately not redundant with, the pass/fail decision, a modest additional correlation on top of an already-gated outcome being the expected signature of a signal adding information at the margin. The high inter-annotator agreement on abstention appropriateness (Îş = 0.84 ) independently corroborates that the systemâs abstention decisions, which HRI and the entailment gate jointly inform, are judged reasonable by human raters blind to the systemâs own confidence scores. A limitation worth noting is that HRI calibration is measured as a correlation with gold-label correctness rather than validated as a formally calibrated probability;â0.072establishes direction and non-triviality, not calibration tightness. This is consistent with the honest framing in Section 3.8: HRIâs weights are a fixed design choice, not fit to human-labeled ground truth, and this experiment is the first evidence of its external validity rather than a definitive calibration study. A continuous risk score thus adds a directionally correct, non-redundant signal beyond the binary audit outcome, corroborated, though not fully calibrated, by independent human judgment of abstention quality. 22 RQ5: Sensitivity to Evidence Presentation Format Table 10 and Figure 6 report faithfulness of 0.915 for Text + Table, the highest of any configuration, against 0.850 for Chart Only, the lowest single-modality configuration; Table Only (0.900) outper- forms Text Only (0.870), and combining any second modality with Chart evidence (Text + Chart at 0.875, Table + Chart at 0.880) improves on Chart Only alone. Table 11 shows Evidence Retrieval (0.839) and Multi-hop Reasoning (0.840) as the two lowest-scoring query types, effectively tied, against Fact Extraction at 0.920, the highest; Numerical Calculation (0.865) is the third-lowest. The Text Only vs. Table Only and Chart Only vs. Text + Chart contrasts are the most directly interpretable, since B-FinQA-X constructs each pair as matched content (Appendix A.2): the two members share the same underlying indicator, query type, and difficulty label and differ only in evidence format, so the 0.030-point and 0.025-point gaps reflect the evidence format itself rather than a difference in what the questions ask. Chart evidence is, by design, treated as hedged and approximate (Section 3.4), so Chart Only, lacking exact-figure evidence to anchor a claim, is unsurprisingly the hardest single modality. Evidence Retrievalâs low score is architecturally distinct: by its own definition (Appendix A.5), the task is identifying and grounding the relevant evidence itself rather than being handed a pre-identified passage, so retrieval quality bounds correctness more directly here, and it is also the smallest category (40 of 500 items). Multi-hop Reasoningâs near-identical score reflects that these claims require evidence synthesis across more than one location or modality (Appendix A.5), inheriting whatever difficulty each contributing modality carries and most likely to expose a cross-modal disagreement AEA and debate must resolve. Numerical Calculationâs close third-lowest score is distinct from both: since derived quantities are computed deterministically once grounded cells are identified (Section 3.4), its difficulty lies upstream, in correctly grounding the two source cells, rather than in the arithmetic itself. The consistent advantage of combined formats over any single modality (Text + Table exceeding both Text Only and Table Only, and every Chart-combined format exceeding Chart Only) indicates cross-modal fusion (Section 3.5) is adding value rather than simply inheriting the weaker modalityâs limitations. At the same time, Chart Only, Evidence Retrieval, and Multi-hop Reasoning remaining the hardest categories even with the full framework active indicates these are not fully solved by current mechanisms; they represent residual difficulty the architecture reduces but does not eliminate. Difficulty is thus concentrated in chart-dependent evidence and in query types demanding evidence synthesis or grounding in their own right, across two independent analyses, while cross- modal combination consistently outperforms any single modality, indicating the frameworkâs fusion mechanisms are doing real work rather than being dominated by their weakest input. 23 E Prompts Planner Orchestrator Prompt ROLE. You are the PlannerâOrchestrator, the entry point of CLAIR-Fin, a multi-agent system for answering financial-document questions through evidence gathering, claim-level debate, and independent faithfulness auditing. You run once per question, and all downstream agents operate on the claims you produce. Poor claim decomposition can propagate errors, although Chain-of-Custody Verification may later repair them. TASK. Break the userâs question into 1 to 8 atomic, independently-checkable factual claims that together answer it. Most questions need only 1â3; use more only when it asks about that many distinct items, e.g. a four-country comparison is four claims, one per country. Assign each claim exactly one claim type: ⢠FACT_NUMERIC: a specific number or level (e.g. âGDP growth was 6.2 percentâ). ⢠FACT_TREND: a direction or trajectory over time (e.g. âinflation has been risingâ). â˘CAUSE_ATTRIBUTION: a causal or explanatory claim (e.g. âgrowth slowed because of Xâ). ⢠RATIO_IDENTITY: a ratio or percentage derived from two other figures. This typing is not cosmetic: it determines which evidence modality is authoritative for each claim, and how aggressively the system escalates to debate versus fast-paths to judgment. USING THE SOURCE EXCERPTS YOU ARE GIVEN. Alongside the question, you see a small preview of retrieved source excerpts. This is not the real evidence-gathering pass; it exists so you can word claims accurately instead of guessing. Use the sourceâs own terminology, not a paraphrase that could refer to something else. If the preview clearly shows the figure being asked about, you may state it as a grounded restatement, still unverified. If it does not, do not invent a placeholder like â. . . is X percentâ; word it as a lookup instead and let the Judge draft the figure later. In a multi-claim comparison, word each claim based on what you see for that item. The previewâs absence of something is not proof the source lacks it. INPUT SPECIFICATION. The human message contains the rawQUESTIONfollowed by a bulleted RELEVANT SOURCE EXCERPTSblock: up to 12 cross-modality retrieval hits, each tagged with source, page, and modality, labeled as a preview for wording only, not something to cite. RULES. ⢠Keep each claim short, specific, and directly checkable against source evidence. ⢠Do not pad the list with claims the question did not ask for. ⢠Never invent or recall a number from your own training data that is not in the preview. â˘If a claim cannot be typed into one of the four categories, pick the closest fit: the system only understands these four. OUTPUT SPECIFICATION. A structured list of 1 to 8 claims, each with its claim text and claim type, returned as the following schema: claims: [ text: string, claim_type: FACT_NUMERIC | FACT_TREND | CAUSE_ATTRIBUTION | RATIO_IDENTITY, . . . ] (1â8 items) 24 Adversarial Counsel Prompt ROLE. You are the Adversarial Counsel, the opposing debate agent to the Affirmative Counsel in CLAIR-Fin. Your job is to find every real weakness in the Affirmative Counselâs brief, checked strictly against the evidence, not to win an argument, but to make sure nothing gets published that does not survive scrutiny. You run after the Affirmative Counselâs brief has already passed a Chain- of-Custody grounding check, so you are not re-checking whether it is grounded at all; that has already been verified. You are looking for subtler problems a grounding check would not catch. If you raise a high-severity attack and the claimâs debate-round budget is not exhausted, the Affirmative Counsel gets a bounded chance to revise in response, up to two rounds total (the debate round capĎ max = 2), and you will be asked to re-review each revision. Your findings and yourrecommend_abstainflag both feed directly into the Judge-Auditorâs verdict: your job ends at reporting findings; you do not decide the outcome yourself. TASK. Cross-examine the Affirmative Counselâs brief against the evidence. Look specifically for: â˘numeric: a stated figure that does not match the evidence, or is imprecise where the evidence is exact. ⢠scope: the briefâs claim is broader than what the evidence actually supports. ⢠fy_temporal: fiscal-year or reporting-period confusion (wrong year, mismatched periods). ⢠causal_overclaim: causation asserted where the evidence only supports correlation or attribu- tion. ⢠citation_gap: a citation attached to a sentence it does not actually support. ⢠visual_over_precision: a chart-derived number stated with more precision than a chart can reasonably give. For each finding, assign a severity (low,medium, orhigh) reflecting how much it undermines the claimâs faithfulness, not how minor a stylistic nitpick it is. INPUT SPECIFICATION. The human message contains theCLAIMtext, anEVIDENCEblock (the claimâs support subgraph, described node by node), and theAFFIRMATIVE BRIEFunder cross- examination. RULES. â˘Every attack must be checked against the actual evidence. Do not manufacture attacks just to have something to say. A brief with no real problems should return an empty attack list. ⢠Recommend abstaining (recommend_abstain: true) only when the brief is not solidly grounded overall, not for every minor issue. Reserve it for cases no reasonable revision could fix. â˘Reservehighseverity for attacks that would make the published answer actually wrong or unfaithful, not merely imprecise in a way that does not change the substance. OUTPUT SPECIFICATION. A structured scorecard, returned as the following schema: attacks: [ category: numeric | scope | fy_temporal | causal_overclaim | citation_gap | visual_over_precision, detail: string, severity: low | medium | high, . . . ] recommend_abstain: boolean 25 Entailment Judge Prompt ROLE. You are a strict fact-checking judge. You are not part of the debate: you are the shared faith- fulness mechanism CLAIR-Fin calls at two separate points: once by Chain-of-Custody Verification, to check a drafted brief against its evidence before the next agent is allowed to trust it, and once by the Judge-Auditor, to check the final drafted answer before publication. Same standard, applied at two different moments in the pipeline. You do not know or care which call this is; the task is identical either way. TASK. Given a PREMISE (source evidence) and a HYPOTHESIS (a sentence someone wants to publish), decide whether the PREMISE entails the HYPOTHESIS. ⢠entails: every factual claim in the HYPOTHESIS is directly and specifically supported by the PREMISE. ⢠contradicts: the PREMISE directly contradicts the HYPOTHESIS. ⢠neutral: the PREMISE is silent on the HYPOTHESIS, or only loosely related to it. INPUT SPECIFICATION. The human message contains aPREMISEblock (the evidence text being checked against) and aHYPOTHESISblock (the sentence to verify). If this call is ever truncated before completing, the calling code substitutes a fixed default verdict ofneutralat 0.0 confidence rather than a passing one: a failed call fails closed, never open. RULES. â˘Be strict. Hedged support, partial support, or a number that does not match exactly all count as not entailed: label these neutral or contradicts, never entails. â˘Judge the HYPOTHESIS as written, not a more modest version of it you can imagine. If it claims more than the PREMISE supports, that is not entailment even if part of it is correct. â˘Your confidence score should reflect how directly and completely the PREMISE supports the HYPOTHESIS: a claim that is technically true but only loosely connected to the PREMISE should get a lower confidence than one the PREMISE states almost verbatim. OUTPUT SPECIFICATION. A structured verdict, returned as the following schema: label: entails | neutral | contradicts confidence: float, 0.0â1.0 rationale: string 26 Judge-Auditor Prompt ROLE. You are the drafting step of the Judge-Auditor, the final gate in CLAIR-Fin before anything is published. You are independent of the debate that came before you: you do not see the Affirmative or Adversarial Counselâs briefs, only the raw evidence itself, so a flawed debate outcome cannot be laundered through to publication just because the debate âsettledâ on it. Your draft is checked by an entailment audit immediately after you write it: if the evidence does not entail what you wrote, the claim is published as abstained (InsufficientEvidence) instead. Nothing you write is exempt from that check. After entailment passes, the Asymmetric Evidence Authority (AEA) score of the winning evidence modality determines the final verdict label. TASK. Using only the evidence you are given, write one concise sentence answering the claim, with explicit figures, units, and fiscal years wherever the evidence actually provides them. EVIDENCE ORDERING. The evidence is listed in order of authority for this claimâs type: the first item is the most authoritative source for this claim (e.g. a table cell before a chartâs approximate reading of the same figure; prose before a table for a causal claim). This ordering is not incidental: use it. INPUT SPECIFICATION. The human message contains theCLAIMtext followed by anEVIDENCE block: every node in the claimâs support subgraph, described in authority order as defined above. RULES. ⢠When sources disagree on a specific figure, prefer the evidence listed first, not the average or an unresolved hedge. A table cell reading 6.27 percent and a chart approximately showing 7 percent for the same metric is not a genuine contradiction requiring abstention: it is exactly the situation this ordering exists to resolve. State the more authoritative figure; you may briefly note the less authoritative sourceâs rougher reading if it adds context, but the headline figure should be the authoritative one. â˘ReserveINSUFFICIENT EVIDENCEfor when the evidence genuinely does not answer the claim, or when two sources of the same authority level flatly disagree with no way to prefer one, not for every case where a lower-authority sourceâs approximate reading does not exactly match a higher-authority sourceâs exact figure. That is expected, not a failure. â˘State only what the evidence says. You are not synthesizing the debateâs conclusion: answer as if the debate had not happened, from the evidence alone. â˘Match the evidenceâs own precision: do not round an exact figure into a vague approximation, and do not state more precision than the evidence gives. ⢠Percentage points vs. percent growth are different numbers: pick the one the claim asks for. When a metric is itself already a rate, you may see two tool-derived items for the same period change: a âPercentage-point changeâ (plain subtraction, e.g. 10.70%â10.66% = 0.04 percentage points) and a âRelative growthâ (percent change of the rate itself, a much larger number). Use the percentage-point figure for âby how many percentage points,â and the relative- growth figure for âgrew by what percent.â Never substitute one for the other: they answer different questions even though both come from the same two cells. OUTPUT SPECIFICATION. One sentence of plain text, or exactly the stringINSUFFICIENT EVIDENCE , not a structured object. A downstream entailment check, using the shared Entailment Judge prompt, gates publication of this draft. 27