How Audex Caught a 33.5% Revenue Discrepancy in Intel's 10-K That LLM Extraction Missed

The Filing

Intel Corporation (INTC) filed its FY2025 10-K with the SEC (accession 0000050863-26-000011), reporting consolidated revenue of $52.853 billion. The filing reflects Intel's restructured segment reporting: Intel Products, Intel Foundry, and All Other — a structure adopted after the separation of Intel Foundry Services (IFS) into a distinct operating segment.

The Extraction

Audex's Layer 1 (structured extraction via Claude Sonnet, triple-pass consensus) correctly extracted all twelve financial claims from the filing. These included the standard income statement items — revenue, cost of revenue, gross profit, operating income, net income — along with balance sheet totals and segment revenue breakdowns:

Every individual number is correct. The LLM read the document accurately. A pipeline that stops at extraction would report these values with high confidence — and any downstream model that sums segment revenues would compute Intel's total revenue as $70.5 billion.

That number is wrong by $17.7 billion.

The Discrepancy

Audex's Layer 2 (mathematical consistency) performs a deterministic check: do the extracted segment revenues sum to the extracted total revenue? In this case:

The 33.5% delta triggered an immediate rejection. All twelve claims in the filing were flagged because the segment-to-total inconsistency contaminates the reliability of the entire extraction set.

Why It Happens

Intel Foundry manufactures semiconductor wafers for two types of customers: external foundry clients and Intel's own product divisions. When IFS produces a wafer for Intel Products, it records that as revenue in its segment. Intel Products records the finished chip sale to the end customer. Under GAAP consolidation, the internal transfer must be eliminated to avoid double-counting.

The result: Intel's segment disclosures report $70.5B in aggregate segment revenue, but the consolidated income statement reports $52.9B. The $17.7B difference is inter-segment eliminations — real accounting entries that exist in the footnotes but are invisible to any system that only reads the segment table.

This is not an error in the filing. It is a structural feature of GAAP segment reporting that creates a trap for automated extraction.

What This Means for Quant Pipelines

Consider the implications for any data pipeline that ingests segment-level revenue:

• A revenue-weighted sector model that sums Intel's segments would overweight Intel by 33%
• A segment growth analysis comparing IFS quarter-over-quarter would include internal transfer revenue, distorting the organic growth rate of the foundry business
• A peer comparison of Intel Foundry vs. TSMC or Samsung Foundry would use a revenue base inflated by captive demand
• Any factor model using segment-level data would propagate the $17.7B error into portfolio weights

The LLM extracted everything correctly. The error is not in the extraction — it is in the assumption that extracted values are ready for downstream use without verification.

The Audit Trail

Every rejected claim produced a persistent audit trail record with:

• The exact layer that triggered the rejection (Layer 2, math consistency)
• The specific check that failed (segment sum vs. total revenue)
• The computed delta (33.4569%)
• A unique audit trail ID for each claim, traceable through the verification database

This is what separates verification from analysis. A signal tool would give you a confidence score. Audex gives you a deterministic reason for rejection and the exact arithmetic that triggered it.

Broader Results

Running the same engine across 12 companies produced consistent patterns:

Companies without inter-segment transactions (AAPL, META, GOOG, SMCI) verified cleanly — 0.0000% delta against XBRL ground truth on every matched claim. Companies with vertically integrated segments (INTC, BA, WBD) were correctly flagged with specific, traceable reject reasons.

Conclusion

LLM extraction is a solved problem for well-structured documents. The unsolved problem is what happens after extraction: are the numbers internally consistent? Do they match the authoritative XBRL taxonomy? Are the cross-statement relationships satisfiable?

Intel's 10-K is a cleanly filed document with no errors. But a pipeline that extracts segment revenues without verification would produce a $17.7 billion overcount — silently, confidently, and with no audit trail explaining why.

Audex exists to make that kind of failure impossible.