Agentic Arithmetic Validation for Financial Transaction Documents
Why OCR and rule-based checks miss arithmetic errors in financial transaction documents, and how agentic AI recalculates, reconciles and flags every mismatch without manual review.

Abstract
The core of a financial transactional document is a set of numerical claims that must agree with one another. Invoices, credit notes and account statements include line amounts, subtotals, taxes, discounts, charges and balances, and each number on that document is implicitly assumed to follow from the others. Document extraction successfully tags the information in the document. But, does not establish whether they hold. One needs to explicitly observe and run a check to ensure the math adds up – i.e. someone needs to validate the correctness and the completeness of the information in the system.
Here, we describe agentic arithmetic validation: where every extracted document as an overdetermined system of arithmetic relations, independently reconstructs each relation from its constituent parts, localizes any disagreement to the field most likely responsible, attributes the residual to a probable cause, and records the evidence behind every conclusion.
We trace the lineage of arithmetic checking, formalize the invariant families of invoice-like and statement-like documents, and state design principles that distinguish an agentic validator from a rule list. We then evaluate seven frontier models from three vendors as stand-alone validators on a curated document set. The strongest models computed accurately but misread representation, applied inconsistent policy and produced false evidence; others missed or invented errors. These results indicate that a purpose-built validator remains necessary.
Introduction: Extraction is not Validation
An extracted invoice is a set of assertions, not a set of facts. When a supplier states that ten units at 50.00 cost 500.00, that the lines sum to a subtotal, and that tax and freight bring the payable to a given total, each figure is a claim about the others. Modern document AI recovers those claims with high fidelity. Faithful capture of a wrong total, however, is still a wrong total.
A document goes through multiple functions and hence the propensity for incorrect information increases. The vendor may err: a mistyped price, a transposed subtotal, a tax applied twice. The channel may corrupt: scanning, OCR and layout parsing can misread digits, drop decimal points or attach a value to the wrong field. And the document may be correct by its own arithmetic but wrong in context: a self-consistent tax at a rate that does not apply to the ship-to jurisdiction.
Arithmetic validation addresses the first two directly and creates the preconditions for the third. Its premise is simple: a figure is accepted only after it has been independently recomputed from the other figures that determine it, within a stated tolerance, with the working retained.
This report makes three contributions:
A formalization of the arithmetic invariants of transaction documents and the tolerance they require (Section 4).
Design principles for an agentic validator, illustrated on worked examples (Sections 5 and 6).
An empirical study of seven frontier models used as stand-alone validators, with a released test kit (Section 7).
Why Arithmetic Validation is Critical
Arithmetic validation is a financial control. Its value comes after capture and before posting or matching, the last point at which an error is still not too expensive to correct. Five properties make it critical.
Dimension | Failure without the control | Where does recomputation (arithmetic check) help |
Financial leakage | A subtotal error propagates into tax and payable; overpayment is hard to recover | The residual is quantified before payment. |
Tax and compliance | A tax can be arithmetically consistent yet levied on the wrong base or rate | Consistency is established first, so rate correctness can be tested against a reference |
Downstream integrity | Wrong amounts enter the ledger, distort accruals, and may surface only at reconciliation or close | Errors are contained at the document and not trickle down to the account books |
Operational cost | Reviewers re-read whole documents, or learn to ignore noisy flags | Findings point to specific fields with the discrepancy and a likely cause |
Auditability | "The system accepted it" is not evidence | Audit trails and reasoning is available for every validation check outcome. |
Note that errors compound: a wrong figure at the line level can lead to an error across every figure computed from it, so the document can remain internally consistent at the header while being wrong throughout (Example A, Section 6). Second, error is asymmetric in cost: a false negative moves money, while a false positive costs reviewer attention. A validator must therefore be judged on both detection and restraint; one that flags too freely is ignored, and an ignored control is no control.
Historical Context: From Checking to Computing
Every era of commercial arithmetic has paired faster calculation with a way to check it. References below show that redundancy, independent re-computation and a durable record have been explored for several problems across several decades:
Period | Work | Contribution to arithmetic assurance |
1202 | Fibonacci, Liber Abaci | Popularised casting out nines in Europe: a cheap residue check on a hand calculation. A digit transposition leaves a difference divisible by nine, a property bookkeepers still use. |
1494 | Pacioli, Summa de arithmetica | Codified double-entry bookkeeping. Every amount is recorded twice, so a trial balance that fails to agree is evidence of error: redundancy as a control. |
1642 | Pascal's arithmetic machine | Built to relieve his father's tax-administration arithmetic in Rouen; mechanized carrying for fiscal sums. |
1670s | Leibniz's stepped reckoner | Extended mechanical calculation to multiplication; Leibniz argued that skilled people should not lose hours to routine calculation. |
1822 | Babbage's Difference Engine | Motivated by errors in printed tables, many introduced in transcription and typesetting. Babbage proposed computing and printing mechanically to remove the human copy step, the precursor of today's capture-error problem. |
1879–1888 | Ritty's cash register; Burroughs' printing adding machine | Registers recorded each sale to constrain error and theft; Burroughs, a former bank clerk, added a printed tape. Calculation acquired an audit trail. |
1950–1960 | Hamming codes; Luhn's check digit | Redundant digits detect, and sometimes correct, corruption in stored and transmitted numbers. Luhn's scheme still guards card and account numbers. |
1985 | IEEE 754 floating point | Standardised binary arithmetic, and with it the recognised hazards of representing decimal currency in binary; Goldberg (1991) became the standard account. |
1992–2013 | COSO framework; Sarbanes-Oxley §404 | Made documented, testable controls over financial reporting an obligation, not a preference. |
1998–2013 | Panko on spreadsheet error; JPMorgan CIO task force; Herndon, Ash and Pollin | Showed that calculation errors in trusted tools are common, persistent and consequential, including in risk models and published economics. |
2021–2023 | Verifiers, PAL, ReAct, Toolformer | Showed that language models make arithmetic slips, and that delegating computation to deterministic tools and verifying outputs materially improves reliability. |
To summarize, redundancy detects error: a document that states its own subtotal and total is a ledger with built-in cross-footing. The copy step is the weak link: Babbage's transcription problem is today's extraction problem. Computation should be delegated: from the Burroughs tape to tool-using models, trust follows from a check that can be replicated.
Problem Formulation: The Document as a System of Claims
We model a validated document as a set of fields and a set of arithmetic relations among them. A relation holds when its two sides agree within a tolerance; validation is the evaluation of every relation, followed by the inference of which fields explain any that fail.
For line i, let q, p, d and t denote quantity, unit price, line discount and line tax, and n the stated line amount. At the header, let N be the stated subtotal, T the total tax, C additional charges, D the header discount and G the stated total. A relation is satisfied when |lhs − rhs| ≤ τ, where τ is the tolerance that applies to that relation.
1 Invariant families
Family | Relation (schematic) | What a failure suggests |
Line reconstruction | n ≈ q × p − d, or the tax-inclusive form n ≈ (q × p − d)(1 + r) | Mis-keyed price or quantity, or a misread line amount |
Aggregation (cross-footing) | Σ n ≈ N | Omitted, duplicated or misread line; transposed subtotal |
Tax by amount | Σ tax components ≈ T | Missing or double-counted tax component |
Tax by rate | rate × taxable base ≈ tax amount | Wrong rate, wrong base, or rounding outside convention |
Discount | rate × base ≈ discount amount; sign constraints | Discount applied to the wrong base, or a sign error |
Header closure | N + T + C − D ≈ G | Charges or tax counted twice or omitted |
Plausibility | G ≠ N whenever T > 0 | A total that cannot be correct |
Representation | amount in words ≡ amount in figures; converted ≡ original × document-date rate | Capture error, or a conversion at the wrong date |
Balance continuity (statements) | opening + Σ credits − Σ debits ≈ closing; each running balance follows from its predecessor; a period's opening equals the prior closing | Carry-forward error, missed transaction, or broken period chain |
The system is overdetermined: most fields appear in more than one relation. This redundancy is what allows localization. A field that breaks every relation it participates in, while its neighbours satisfy theirs, is the likely culprit.
2 Tolerance and representation
Tolerance is a first-class design parameter. It must reflect the currency's minor unit, the issuer's rounding convention (per line or on the total; half-up or half-even), and the number of rounded terms being combined, since rounding residue grows with term count. Too tight, and legitimate rounding produces false positives; too loose, and small material errors pass. Tolerances should be configurable per client and per relation, and the applied tolerance should be recorded with each finding.
All monetary arithmetic should be performed in exact decimal representation. Binary floating point cannot represent most decimal fractions exactly, and comparisons made on derived binary ratios can fail or pass for reasons unrelated to the document.
Agentic Arithmetic Validation
An agentic validator differs from a rule list in that it reasons over the whole document: it chooses among interpretations, decides which field to doubt, derives what can be derived, and knows when to stop and ask a human. Six principles define it.

agentic arithmetic validation · read, compute, decide, localize, record
Only a relation that no allowed reading can satisfy becomes a finding.
1. Separate reading from computing. Learned models are used where judgement is needed: identifying which span is a subtotal, whether a line is tax-inclusive, what a free-text term means. All arithmetic is then executed by a deterministic engine in exact decimal. A model that both reads and computes yields an answer that cannot be replayed; a model that only reads yields inputs whose consequences can be checked. This mirrors the program-aided and tool-use findings in the language-model literature.
2. Evaluate competing hypotheses before flagging. Many apparent failures are representational. A line amount may be net or tax-inclusive; a discount may apply before or after tax; a total may be stated gross or net. The validator tests each admissible reading and flags a relation only when no reading satisfies it. This is the single most important control on false positives: a supplier who invoices tax-inclusive would otherwise trigger a flag on every line, and reviewers would learn to ignore them.
3. Localize through the constraint structure. Because fields participate in several relations, the pattern of passes and failures identifies the probable source. If every line reconstructs, and the header closes over the stated subtotal, but the lines do not sum to that subtotal, the subtotal is the suspect, not the lines or the total.
4. Quantify and attribute the residual. A finding states the discrepancy amount, not merely its existence. The residual's structure carries diagnostic information: a residual equal to one line amount suggests an omitted or duplicated line; one equal to a tax or charge component suggests double counting; one that is a power-of-ten multiple suggests a decimal shift; and a residual divisible by nine is the classical signature of a digit transposition. Different causes call for opposite corrections, so attribution is as useful to a reviewer as detection.
5. Act independently within agreed limits, and record the origin of every piece of information. The validator may derive a missing value that is fully determined by others, such as a unit price from amount and quantity or a total tax from its components. Every such value is marked as system-supplied, names its source and working, and leaves the original intact. Human corrections should be treated as final.
6. Know the limits of arithmetic. Arithmetic proves consistency, not correctness. Once a tax is shown to be consistent with its base, its rate is compared against an authoritative jurisdiction reference; converted amounts are checked against the rate in force on the document date. Where the reference is unavailable, the finding says unverifiable , and completeness is judged only after every derivation has run.
Examples
All data below is synthetic; amounts are in a generic currency with two decimal places.
Example A: a header that closes over a wrong subtotal
Line | Description | Qty | Unit Price | Stated Amount | Recomputed |
|---|---|---|---|---|---|
1 | Industrial bearings | 120 | 14.75 | 1,770.00 | 1,770.00 |
2 | Hydraulic seals | 40 | 22.40 | 896.00 | 896.00 |
3 | Calibration service (hours) | 6 | 85.00 | 510.00 | 510.00 |
Subtotal | 3,716.00 | 3,176.00 | |||
Tax at 8% | 297.28 | 254.08 | |||
Freight | 135.00 | 135.00 | |||
Total Payable | 4,148.28 | 3,565.08 | |||
Line level validation adds up: 3,716.00 + 297.28 + 135.00 = 4,148.28. A check that tests only header closure passes this invoice. The lines sum to 3,176.00, a residual of 540.00. Because the lines and the header both hold, the constraint structure isolates the subtotal (P3). The residual is divisible by nine, consistent with the transposition of 3,176 into 3,716 (P4). The tax is consistent with the stated subtotal, so it inherits the error. Total exposure is 583.20: 540.00 of goods never supplied plus 43.20 of tax on them.
Example B: consistent versus correct tax
B1, inconsistent. A line states a taxable base of 12,500.00, a rate of 18% and a tax of 2,125.00. At 18% the tax is 2,250.00; the stated amount implies 17%. The rate relation fails with a residual of 125.00, and the finding reports both the stated and implied rates.
B2, consistent but wrong. A line states a base of 2,400.00 and tax of 204.00 at 8.5%. The arithmetic is exact. The authoritative jurisdiction rate for the ship-to address is 7.25%, giving 174.00. Arithmetic alone cannot find this 30.00 over-collection; it is caught only because consistency is established first and the rate is then tested against a reference (P6).
Example C: a false positive avoided
A line shows 10 units at 50.00 with a line amount of 590.00. Under the net reading, q × p = 500.00 and the line fails. Under the tax-inclusive reading, 500.00 × 1.18 = 590.00 and the line holds. The validator records the line as tax-inclusive and raises no flag (P2).
Example D: an account statement with a carry-forward error
Date | Transaction | Credit | Debit | Stated Balance | Recomputed |
|---|---|---|---|---|---|
01 Oct | Opening balance | 48,250.00 | 48,250.00 | ||
03 Oct | Customer receipt | 12,400.00 | 60,650.00 | 60,650.00 | |
07 Oct | Supplier Payment | 3,815.50 | 56,834.50 | 56,834.50 | |
12 Oct | Payroll transfer | 9,200.00 | 47,364.50 | 47,634.50 | |
18 Oct | Customer Receipt | 5,600.00 | 52,964.50 | 53,234.50 | |
25 Oct | Bank charges | 1,147.25 | 51,817.25 | 52,087.25 | |
31 Oct | Closing balance | 51,817.25 | 52,087.25 |
The summary relation fails: 48,250.00 + 18,000.00 − 14,162.75 = 52,087.25, not 51,817.25. The summary check alone says only that the statement is 270.00 short. Row-by-row continuity localizes it: only the 12 October row fails against its predecessor, and every row after that is consistent with the wrong figure it inherited. The error is a single carry-forward, and its residual of 270.00 is again divisible by nine, the signature of 47,634.50 transposed to 47,364.50. A period-continuity check against the prior statement's closing balance would additionally confirm the opening figure.
Empirical Study: Frontier Models as Validators
If frontier models now compute well, is a purpose-built validator still needed? We test whether a general model, prompted with the checks of the invariant families in Section 4, can serve as the control itself.
1 Setup
We used three suites of documents, 29 in total: 19 with one or two errors and 10 clean. They range from 14-line invoices to 250-line invoices and a 150-row statement, and exact decimal answer keys were generated with the documents. Two clean controls (D04 and S11) state VAT-inclusive line amounts in an unlabeled column; a third (S08) labels the column.
Four Claude models (claude-opus-5-5, claude-sonnet-5-5, claude-haiku-5-5 and claude-fable-5-1, extended reasoning on) ran through the API: 36 runs, of which 32 were unaided and 4 had a code tool. ChatGPT (Think mode) and Gemini 3.1 Pro and 3.6 Flash (thinking High) ran once per suite in their consumer apps in October 2026, with identical prompts.
The design follows controlled perturbation (Section 9): known errors are injected in verified documents, with rounding and representation variants as negative controls. We score detection, root-cause localization, the stated residual, and false alarms on clean documents.
2 Results
Note that we define a financial control to be right on every document, apply the same policy every time, and support each finding with true evidence. No model met all three.
Model | Runs | Errors detected | Root cause located | Residual correct | D04 flagged | S11 flagged | S08 flagged |
Claude (4 models, API) | 36 | 252 of 252 | 252 | 252 | 12 of 12 | 7 of 16 | 0 of 16 |
ChatGPT, Think (app) | 3 | 19 of 19 | 19 | 19 | 1 of 1 | 1 of 1 | 0 of 1 |
Gemini 3.1 Pro, High (app) | 3 | 19 of 19 | 19 | 18 | 1 of 1 | 1 of 1 | 0 of 1 |
Gemini 3.6 Flash, High (app) | 3 | 18 of 19 | 18 | 17 | 1 of 1 | 1 of 1 | 0 of 1 |
Note that D04, S11 and S08 are clean documents, so any error on them is a false positive.
Failure mode | What we observed |
Claude: reasoning gaps | |
Misread representation | Correct VAT-inclusive invoices were rejected (D04 in 12 of 12 runs, S11 in 7 of 16; the labelled S08 in 0 of 16) |
Inconsistent policy | The same structure got different verdicts depending on model, document context and tool access |
False evidence | Correct verdicts were justified with figures that are not in the document |
Propagation noise | One error was reported as 11 findings; 10 of 32 unaided runs listed inherited fields as errors |
Contract drift | Up to 9 of 16 field names fell outside the predefined vocabulary |
Latency | 42 to 114 s per 150 to 250-line document, against milliseconds for a custom engine |
ChatGPT and Gemini: basic gaps |
|
Missed error | Gemini Flash passed D08, a statement with a 72.00 carry-forward error running across ten rows |
Invented errors | Gemini Pro flagged four correct balances and both totals in S03 |
Wrong amount | Gemini Flash gave a tax error as 62.58 instead of 69.06; Gemini Pro left one error unquantified |
Misread representation | All three rejected D04 and S11 |
3 Failure examples
The following table lists actual model outputs where errors were observed in documents with known answers.
Model | What the document says | What the model said | Why it matters |
Claude, ChatGPT, Gemini | D04: 4 days × 650.00 shown as 3,120.00 (VAT-inclusive); net 3,250.00 + VAT 650.00 = 3,900.00 reconciles | Lines overstated; invoice in error | A correct invoice is rejected because of an assumed reading |
Claude | D07: line 2 reads 102.00, where 12 × 0.85 should be 10.20 | "The line amounts 75.60 + 10.20 + 144.00 sum to 229.80, not the stated 321.60" | The finding cites 10.20 as if it were on the document: false evidence |
ChatGPT | D08: a single 72.00 transposition at row 7 of a statement | Eleven separate errors (rows 7 to 16, and the closing balance) | One correction becomes eleven review items |
Gemini 3.6 Flash | D08, the same statement | Clean | A real error passes into the ledger |
Gemini 3.1 Pro | S03, row 26: a credit of 3,556.73, added correctly | Wrong by 7,113.46, treated as a debit because the description says "Transfer out" | A correct row is reported as an error |
4 Failure analysis
Claude's gaps are in reasoning, not arithmetic. All 252 errors were caught and quantified exactly, including a 0.10 slip in a 7-digit multiplication and a 0.72 transposition at row 121 of 150. But correct invoices were rejected under an assumed reading of an unlabeled column, and identical structures drew different verdicts across models and tool access. A critical gap observed was false evidence: the finding on D07 presents the model's own correction as the document's content, misleading the reviewer.
The other models showed more basic failures. ChatGPT matched Claude on detection, but false positives were repeated as well. Gemini Flash passed a statement that contained a real error. Gemini Pro flagged the correct rows of S03. In 32 of that statement’s 60 rows the description contradicts its credit or debit column, which is a flaw in our generator, but the prompt said to follow the columns and the other models did.
Prior work agrees. Unaided arithmetic degrades with digit count and length across GPT, Gemini and Claude models (Bie et al., 2026; McKinney, 2025). The best model scored 63% on 500 real-world calculations (Herambourg et al., 2025). Nine of 14 model runs flagged 95 to 100% of clean financial statements (Panda, 2026), and symbolic re-checking removed the false accepts an LLM judge let through (Hall et al., 2026).
5 Implications for validator design
The gaps help identify the principles of the validation design:
An explicit representation policy removes the tax-inclusive false positive(P2).
Exact decimal computation cannot miss, invent or misstate a residual, and it runs in milliseconds (P1).
Constraint-based localization separates root causes from their consequences (P3).
Evidence generated from the computation cannot be misdescribed to the document (P4).
Versioned rules and a fixed output schema give the same verdict on every run
Evidence, Provenance and Auditability
Every validation outcome attributes the findings across 6 different dimensions: a severity, a stable reason code, natural language reason, the assertion with the actual values compared, the reasoning that led to the conclusion, and, where possible, the discrepancy and a suggested resolution. Example A's subtotal finding would read as follows.
Element | Content |
Field | Subtotal |
Severity | High: blocks posting until resolved |
Assertion | Sum of line amounts (3,176.00) ≠ stated subtotal (3,716.00); all lines reconstruct; header closes over the stated subtotal |
Residual | 540.00; divisible by nine, consistent with a digit transposition |
Effect | Tax and total inherit the error; payable overstated by 583.20 |
Suggested action | Verify the subtotal against the source image; correct and revalidate |
The outcomes are further quantified, Severity grades include correct, suspect or unverifiable, wrong, and system-supplied. AI and system generated values are explicitly marked, sourced and accompanied by the original. Human corrections are treated as ground truth with gold labels.
A reviewer goes to the flagged fields, and an auditor can replicate as is replay every conclusion from the values recorded beside it in the audit trails.
Evaluation Methodology
A validator should be judged on whether it catches material errors without false positives and on whether its findings let a reviewer act. We propose five measures for validator evaluation.
Measure | Definition | Why it matters |
Material detection rate | Share of documents with a discrepancy above a materiality threshold that are flagged (critical errors should always be flagged) | Errors that move money must not pass |
Clean-document flag rate | Share of verified-correct documents that receive any error-level flag | The main driver of reviewer trust and adoption |
Localization accuracy | Share of true findings attached to the field actually at fault | A correct alarm on the wrong field still costs an investigation |
Resolution without re-reading | Share of findings a reviewer resolves from the evidence alone | Measures whether the explanation is sufficient |
Run-to-run stability | Agreement of findings across repeated runs on the same input | Non-determinism is invisible unless deliberately measured |
Evaluation should combine labelled production samples with controlled perturbation: take verified-correct documents and inject known error types (digit transposition, decimal shift, omitted and duplicated lines, double-counted tax, a wrong but self-consistent rate) alongside rounding-only noise as negative controls. Detection should be reported per error type and weighted by monetary impact, since a validator that catches every one-cent rounding mismatch but misses a decimal shift is worse than useless.
Limitations and Threats to validity
Arithmetic validation is necessary, not sufficient. Its boundaries include the following:
Consistency is not legitimacy. A fraudulent invoice can be arithmetically perfect. Duplicate detection, vendor verification, forensic tests such as digit-distribution analysis (Nigrini, 1996) and approval workflows should always remain separate controls.
It cannot recover what was never captured. Where no other field or authoritative record determines a value, the validator flags it; it cannot read a figure back from an image the extractor missed.
Localization can be ambiguous. When two fields each break the same pair of relations, the structure may not single one out. The correct behavior is to name the candidates, not guess.
Reference data has its own error and latency. A jurisdiction rate or exchange rate is only as current as its source; findings that depend on it should say so.
Tolerance is a trade-off. No setting eliminates both missed errors and false alarms; tolerances must be set per client and revisited as document mix changes.
The validator produces evidence. Posting and payment decisions stay with the client's workflow and approvers.
Conclusion
Eight centuries of commercial arithmetic converge on one rule: a number is trusted after it has been checked by an independent route, and the check is recorded. Agentic arithmetic validation applies that rule to every extracted financial document. It treats the document as an overdetermined system, recomputes each relation exactly, and tests competing interpretations. It further localises and quantifies what fails, and leaves a record a reviewer can replicate. Frontier models compute well most of the time, but none we tested was dependable as a control on its own. The strongest frontier models misread correct invoices and narrated evidence the document did not contain, and others missed or invented errors. A purpose-built validator supplies the consistency and evidence a financial control requires. Extraction tells you what a document says. Validation tells you whether it is right, and shows its working.
References
Fibonacci, L. (1202). Liber Abaci. Pisa.
Pacioli, L. (1494). Summa de arithmetica, geometria, proportioni et proportionalità. Venice.
Pascal, B. (1645). Lettre dédicatoire à Monseigneur le Chancelier sur le sujet de la machine nouvellement inventée. Paris.
Leibniz, G. W. (1685). Machina arithmetica in qua non additio tantum et subtractio sed et multiplicatio nullo, divisio vero paene nullo animi labore peragantur. Manuscript.
Babbage, C. (1822). A Letter to Sir Humphry Davy on the Application of Machinery to the Purpose of Calculating and Printing Mathematical Tables. London.
Ritty, J., & Ritty, J. (1879). Cash register and indicator. US Patent 221,360.
Burroughs, W. S. (1888). Calculating-machine. US Patent 388,116.
Hamming, R. W. (1950). Error detecting and error correcting codes. Bell System Technical Journal, 29(2), 147–160.
Luhn, H. P. (1960). Computer for verifying numbers. US Patent 2,950,048.
IEEE (1985). IEEE Standard for Binary Floating-Point Arithmetic (IEEE 754-1985).
Goldberg, D. (1991). What every computer scientist should know about floating-point arithmetic. ACM Computing Surveys, 23(1), 5–48.
COSO (2013). Internal Control — Integrated Framework. Committee of Sponsoring Organizations of the Treadway Commission.
Sarbanes-Oxley Act of 2002, Pub. L. 107-204, §404.
Panko, R. R. (1998). What we know about spreadsheet errors. Journal of End User Computing, 10(2), 15–21.
JPMorgan Chase & Co. (2013). Report of JPMorgan Chase & Co. Management Task Force Regarding 2012 CIO Losses.
Herndon, T., Ash, M., & Pollin, R. (2014). Does high public debt consistently stifle economic growth? A critique of Reinhart and Rogoff. Cambridge Journal of Economics, 38(2), 257–279.
Nigrini, M. J. (1996). A taxpayer compliance application of Benford's law. Journal of the American Taxation Association, 18(1), 72–91.
Cobbe, K., et al. (2021). Training verifiers to solve math word problems. arXiv:2110.14168.
Gao, L., et al. (2023). PAL: Program-aided language models. Proceedings of ICML 2023.
Yao, S., et al. (2023). ReAct: Synergizing reasoning and acting in language models. Proceedings of ICLR 2023.
Schick, T., et al. (2023). Toolformer: Language models can teach themselves to use tools. Advances in NeurIPS 36.

