AUDEXRequest Early Access
PRE-RELEASE — SEEKING EARLY TESTERS

Verification infrastructure
for financial claims

Audex is a neurosymbolic verification engine for SEC filings. It extracts financial claims using an LLM, then routes every output through a deterministic verification stack — mathematical consistency checks, XBRL taxonomy validation, multi-pass consensus scoring, and Z3 constraint solving — before emitting a verified assertion with a full audit trail.

The product is not a dashboard or a score. It is a verified assertion with a traceable reject reason — something defensible enough to act on.

audex verify AAPL
Fetching latest 10-K for AAPL...
Filing: 0000320193-25-000079 | Period: CY2025
Running 6-layer verification pipeline...
us-gaap:Assets$359.2B VERIFIED ✓ XBRL Δ=0.0000%
us-gaap:Liabilities$285.5B VERIFIED ✓ XBRL Δ=0.0000%
us-gaap:StockholdersEquity$73.7B VERIFIED ✓ XBRL Δ=0.0000%
us-gaap:Revenues$416.2B VERIFIED ✓ XBRL Δ=0.0000%
us-gaap:NetIncomeLoss$112.0B VERIFIED ✓ XBRL Δ=0.0000%
14 claims · 8 verified · 6 flagged · 0 rejected · 72.9s
Request Early Access

The 6-layer verification stack

A claim that fails any layer is rejected with a specific, machine-readable reason code. No layer trusts the layer above it. LLMs live only in Layer 1 — everything else is deterministic.

1

Structured Extraction

Anthropic Claude + Pydantic v2

3x multi-pass extraction at temperature=0. Every LLM output parsed into a typed Pydantic model. If it can't be parsed, it's a reject. Consensus scoring across passes.

2

Mathematical Consistency

Pure Python

Balance sheet equation holds. Segment revenues sum to consolidated total. Income statement flows consistently. No LLM involvement — deterministic arithmetic only.

3

XBRL Ground Truth

SEC EDGAR XBRL API

Every extracted value cross-referenced against the company's own XBRL-tagged data. XBRL is treated as canonical. Discrepancies produce a reject with the exact delta.

4

Multi-Pass Consensus

Divergence scoring

Field-level agreement across all extraction runs. Claims where any numeric value disagrees across passes are flagged with the specific divergence.

5

Domain Rules Engine

Extensible Python rules

Growing library of GAAP-specific verification rules. Revenue recognition consistency, segment completeness, asset positivity. Rules are append-only and versioned.

6

SMT Constraint Verification

Z3 theorem prover

Formal mathematical proof that the set of extracted claims is internally consistent. Z3 returns an unsat core when constraints conflict — identifying exactly which claims are in tension.

What the output looks like

Every claim produces a VerifiedAssertion or RejectedClaim — never a probability without context. Each includes the full audit trail: every layer result, the raw LLM responses from all 3 extraction passes, the XBRL ground truth values used for comparison, which rules triggered, and a confidence score with a traceable decomposition.

This audit trail is the product. It's what separates a probabilistic signal from a defensible claim.

Read: How Audex caught a $17.7B segment discrepancy in Intel's 10-K

Request early access

Audex is in pre-release. We're looking for quantitative researchers and alternative data teams who want to test the verification pipeline on their real filings. Free access in exchange for honest feedback.

CLI + API access. No dashboard, no screenshots of things that don't exist yet.