A 5-factor formula your auditors can reproduce on a calculator
Every account reconciliation receives a confidence score between 0 and 1. The formula is deterministic. Not an LLM, not a neural network. Five weighted factors: variance materiality (35%), auto-reconciliation result (20%), matching coverage (20%), materiality level (15%), and historical pattern (10%). Same inputs, same score, every time. Your external auditors can verify any score independently.
The problem
Risk scores from a black box don't survive an audit. Yours need to be explainable.
Most platforms assign risk levels using opaque ML models. When your auditor asks why an account is rated “medium risk,” nobody can answer. Arvexi's confidence score is a formula, not a model. Every factor, every weight, every threshold is visible and documented. Your auditors don't need to trust the AI. They can verify the math.
The 5 factors
Five weighted inputs, one deterministic score
Variance materiality | 35%
The largest factor. Uses an exponential decay formula: as the variance approaches the materiality threshold, the score drops sharply. A $500 variance on a $1M account scores near 1.0. A $50,000 variance on the same account scores near 0.3. The decay curve is steeper for high-materiality accounts.
Auto-recon result | 20%
Did the account pass auto-reconciliation? Zero-balance accounts and tolerance-matched accounts receive full credit. Accounts that failed auto-recon receive partial credit based on how close they were to passing. No-activity accounts receive a configurable default.
Matching coverage | 20%
What percentage of transactions are matched between GL and subledger? 100% matched = full credit. If only 60% of transactions have matches, this factor reflects the 40% gap. Transaction matching runs before confidence scoring, so this factor captures real matching results.
Materiality level | 15%
Accounts classified as critical (cash, revenue, debt) receive stricter scoring than immaterial accounts. This factor ensures that a $10,000 variance on a cash account is treated differently than a $10,000 variance on a low-risk prepaid account. Materiality levels are configurable per account.
Historical pattern | 10%
How has this account performed over the last 3–6 close cycles? Accounts with consistently clean reconciliations earn a bonus. Accounts that were RED last month start with a deficit. Cortex tracks false positives. If an account was flagged RED but turned out clean, the historical factor adjusts upward.
Controller calibration
Not a factor in the formula. A feedback loop that tunes the formula over time. Controllers mark scores as correct, too high, too low, or wrong. Calibration adjustments are per-organization and per-account, accumulating across close cycles to personalize scoring to your business.
Color classification
GREEN, AMBER, RED | three levels that drive everything
GREEN (≥ 0.95): Low risk. Cortex generates a template-based work paper at zero AI cost. No investigation needed. AMBER (0.70–0.94): Review recommended. Cortex generates an AI narrative explaining the account's risk factors and suggesting where to focus. RED (< 0.70): Full investigation triggered. Cortex runs the investigation agent with all 7 data tools, generates anomaly reasoning, and produces a detailed work paper with findings. Read our confidence scoring glossary. Part of Cortex.
Special cases
Two overrides that prevent false confidence
Zero-variance override: when GL and subledger balance exactly, the score is set to 0.97 (GREEN) regardless of other factors. A zero variance means no unexplained difference exists. Though the 0.97 cap (not 1.0) acknowledges that offsetting errors remain possible. Subledger-pending override: when a subledger import is stale or in progress, the score is capped at 0.80 (AMBER). This prevents falsely high confidence from incomplete data. The cap lifts automatically when fresh data arrives.
Auditability
Every score shows its work
Each confidence score includes a full breakdown: the raw value for each of the 5 factors, the weighted contribution, any applicable overrides, and the final composite score. Your auditors can see that Account 1200 scored 0.82 because variance materiality contributed 0.28 (of 0.35), auto-recon contributed 0.20 (full credit), matching coverage contributed 0.14 (of 0.20), materiality level contributed 0.10 (of 0.15), and historical pattern contributed 0.10 (of 0.10).
