Financial assurance
Reperform accounting and cash-flow identities; flag missing or unverified figures rather than filling gaps.
CASE 01 / EXPERT EVALUATION / FINANCIAL REASONING
Can financial AI distinguish cash movement from operating cash generation?
A focused review of two actual T4R Financial Decision Intelligence development reports: one financial claim, the evidence needed to substantiate it, and the expert judgment applied.
00 / THE MODEL
G01 G02 G03 G04
What does T4R FDI do?
An LLM-assisted financial decision workflow that turns source-backed accounting and cash data into CFO-ready analysis while distinguishing observed figures, assumptions and decision risks.
Purpose & outputs
Four independent CFO reports
Cash Reality Report (one month); Cash Decision Engine, Financial Intelligence Report and Strategic Decision Memo (one common analysis period).
Operational workflow
From records to decisions
Normalize financial inputs → validate calculations and classifications → analyze cash, performance and scenarios → render separate decision reports.
MODEL QUALITY GATES
Four checks before a financial report is released
Financial validation, evidence integrity, analytical consistency and controlled delivery.
Evidence & classification
Trace figures to source records, verify account mappings and demonstrate the covered population.
Analytical review
Check cross-report consistency, scenario mathematics and unsupported causal claims; require human semantic review.
Quality & secure release
Confirm source and content integrity, authorized tenant scope and report consistency before export.
Release logic: a material integrity failure blocks export; incomplete evidence requires a clearly marked preliminary report and explicit authorization. Passing technical checks is not an independent financial audit.
Scope distinction: These are the model’s operational safeguards. The independent evaluation task and scoring rubric begin in the next sections.
01 / EVALUATION TASK
What was evaluated?
This case examines two archived FDI outputs for the same historical company dataset: a Strategic Decision Memo and a Cash Decision Engine. The evaluator checks financial classification, arithmetic and decision-risk language.
These reports perform different functions. “A/B” identifies two observed artifacts—not paired responses to a common prompt.
Evidence reference · E01
02 / REFERENCE EVIDENCEREFERENCE FIELD REFERENCE FIELD REFERENCE FIELD
What can be verified?
Extract the reported facts, recompute what is possible and separate measured statements from unsupported classifications.
−BRL 214.5k
Operating cash flow stated in report A
−BRL 200k
Financing cash flow stated in report A
−BRL 407.5k
Total cash change stated in report A
Independent arithmetic check: −214.5 − 200 = −414.5 (BRL thousands). The reported −407.5 total leaves an approximately +7.0k unexplained residual at the displayed precision. The excerpt does not establish its cash-flow classification.
The amounts are observed report statements, not independently authenticated banking or cash-flow-statement values. The reviewer can verify internal arithmetic but cannot certify operating cash flow or unrestricted liquidity from these excerpts.
Evidence reference · E02
03 / SCORING DESIGN
How would the answer be judged?
A five-dimension, 100-point rubric plus mandatory financial integrity gates. The 100-point model grade is reserved for controlled, traceable executions. Below, the reviewer separately demonstrates criterion-level, simulated claim scoring.
| Evaluation dimension | Weight / floor |
|---|---|
| Numerical Accuracy | 30 / 24 |
| Source & Method Integrity | 25 / 19 |
| Accounting vs. Cash Reasoning | 20 / 16 |
| Limitations & Uncertainty | 15 / 11 |
| Executive Decision Quality | 10 / 7 |
| Total possible | 100 |
Evidence reference · E03
04 / OBSERVED OUTPUTS + EXPERT REVIEWActual historical output Actual historical output
What did the outputs say—and what did I find?
Original Portuguese excerpts remain unedited. Findings are based on the specific claims and recorded evidence, not on a simulated model performance score.
A · Strategic Decision Memo
SOURCE FDI-A“Caixa operacional –R$ 214,5 mil e financiamento –R$ 200 mil em agosto: variação total –R$ 407,5 mil conciliada à DFC/Balancete, sem caracterizar burn estrutural no intervalo completo.”Verbatim excerpt from a retained historical output · identifiers removed
EXPERT FINDING · Classification gap
The ≈+BRL 7.0k arithmetic residual cannot be called investing cash flow without corresponding source classification. The memo’s reconciliation claim requires source-level confirmation.
B · Cash Decision Engine
SOURCE FDI-B“Runway: aproximadamente 9 meses caso o burn de agosto se torne recorrente.”Verbatim excerpt from a retained historical output · identifiers removed
EXPERT FINDING · Conditional reasoning
The conditional scenario avoids presenting nine months as an unconditional forecast. But total net cash depletion is not equivalent to proven recurring operating burn, and cash accessibility remains unverified.
EVALUATOR DETERMINATION
From evidence to decision.
What I would accept, hold or release — and why.
Two simulated reviewer dispositions applied to observed historical outputs. Each limited-scope subscore awards only the criterion checkpoints actually assessable here; it is not a full model score or a controlled A/B benchmark.
Rubric-aligned expert adjudication
FDI-01 · SIMULATED EXPERT ADJUDICATION
Cash-flow reconciliation
7/1450% · applicable points only
REVIEWER DISPOSITIONHOLD
- Evidence & assessment
- Report A states operating CF −BRL 214.5k, financing CF −BRL 200k and total movement −BRL 407.5k. The displayed values leave a ≈+BRL 7.0k residual; its classification and independent reconciliation are not shown in the reviewed passage.
- Decision risk
- An unsupported reconciliation could distort liquidity or funding decisions.
- Release condition
- Show the full cash-flow bridge and verify each category against underlying source records.
FDI-02 · SIMULATED EXPERT ADJUDICATION
Conditional runway
5/956% · applicable points only
REVIEWER DISPOSITIONCONDITIONAL USE
- Evidence & assessment
- Report B states ≈9 months only if August cash consumption recurs; BRL 3.69m / BRL 407.5k ≈9.1 months. The scenario is conditional and arithmetically plausible, but the excerpt does not substantiate unrestricted liquidity or recurring operating burn.
- Decision risk
- A conditional arithmetic scenario might be mistaken for a validated liquidity forecast.
- Release condition
- Verify available cash, distinguish operating burn from total movement and substantiate recurrence assumptions.
REVIEW COMPLETE · MODEL NOT GLOBALLY SCORED
These are rubric-aligned, simulated claim-level decisions derived from archived development outputs. The full 100-point acceptance rule (≥85, dimension floors and critical gates) applies only to a traceable, complete model evaluation—not to these scoped excerpts.
Not verified is not the same as proven wrong: unsupported claims are held, while inaccessible or out-of-scope evidence is excluded from the claim-level denominator. E03 records every award and withholding rationale.
06 / EVALUATION EVIDENCE PACK
Verify the evaluation
Three primary references, provided once, support the observed claims, expert findings and scoped scoring decisions.