Verify financial claims. Deterministically.

Aritiq parses numeric assertions from SEC filings using language models, then re-derives and checks every value using deterministic calculation code.

Run an auditView benchmark

No models are used in the verification engine.

The problem

Language models extract information well.
Arithmetic requires deterministic execution.

Language models are effective for unstructured text extraction, but struggle with complex multi-step math and accounting rules. To verify financial metrics, extraction must be separated from mathematical validation using a deterministic computing layer.

Architecture

Isolated parser and calculation layers.

Parser

1 · Extract

A language model parses claims, numbers, and formulas from filings into structured JSON.

Boundary
Calculation

2 · Verify

Deterministic code re-derives every claim using primary XBRL data and accounting rules.

Calculation

3 · Trace

Claims are assigned verification states (Verified, Wrong Math, Insufficient Evidence) with complete audit paths.

The calculation layer is deterministic and has zero model dependencies.

Evaluation

Measured benchmark results.

0.0%

Precision

The ratio of correctly verified or rejected assertions to total verified or rejected outputs.

0.0%

False-Positive Rate

The proportion of incorrect assertions accepted as valid.

0.0%

Coverage

The percentage of parsed assertions with sufficient primary source data to perform calculations.

0

SEC Filers evaluated

A representative set of filings across technology, industrial, and consumer sectors.

0+

Integration Tests

Automated test suite asserting calculation correctness and edge case validation.

Calculations are performed by local python libraries using verified financial schemas.

The pipeline fails safe: assertions are flagged as having insufficient evidence if source data is missing or ambiguous.

Analysis

Case study: AMD 10-K discrepancy.

Issue:An EPS reconciliation assertion was flagged as incorrect during automated evaluation of AMD’s FY2024 filing.
Root Cause:The extraction model parsed an incorrect net income figure from a nearby table block, causing the re-calculation formula to mismatch.
Resolution:Improved table boundary definitions in the parser step. The verification calculation rules remained unmodified.
Outcome:Re-evaluation succeeded without further discrepancy. The verification logic continues to reject invalid inputs rather than modifying rules to fit extraction anomalies.

While the verification layer is completely deterministic, Aritiq has the following operational boundaries:

  • Model-dependent parsing: Parsing narrative disclosures and unstructured tables relies on LLM-based layout analysis and token extraction.
  • Evidence constraints: The system fails safe. If assertions lack matching source disclosures or explicit values, they are classified as INSUFFICIENT_EVIDENCE rather than guessed.
  • SEC filing focus: The present verification ruleset and XBRL schema compiler are optimized for US SEC EDGAR filings (10-K, 10-Q, 8-K) and do not support international standards (IFRS) or private company reports.

Applications

Designed for automated financial pipelines.

AI Analyst Tools & Copilots

Pre-output calculations and truth checks to verify parsed metrics before rendering to end users.

Research & Extraction Pipelines

High-throughput parsing systems checking machine-extracted values against primary SEC database sources.

Compliance & Audit Systems

Automated verification checkpoints ensuring internal alignment between different tables and narrative disclosures.

Console

Evaluate a filing.

Select a ticker to load the latest SEC filing and review re-calculated metrics.

Run audit

Loads target filing from SEC EDGAR database.