Our extraction pipeline reads SEC XBRL. To verify it, an independent reader — an LLM with zero shared code — reads the rendered filing and we diff the two. One bug cannot corrupt both paths, so agreement is evidence, not circularity. Sampling is seeded and stratified across company sizes: re-runnable by anyone.
| EPS (diluted) | 99.5% |
| Total assets | 99.6% |
| Total liabilities | 99.5% |
| Stockholders' equity | 99.6% |
| Net income | 97.5% |
| Revenue | 98.1% |
The adjudicated figure counts the disagreements we have examined against the filing by hand: nearly all sit in two documented classes — bank revenue composition (we report the standard net-interest basis) and net income before vs after preferred dividends (we report the filed attributable-to-parent line). Both are definitional, both are published, none are smoothed over.
Every value the API returns carries its provenance — the exact accession number and filing URL — so any number can be checked against the source in one click. Found a discrepancy? [email protected] — adjudications feed the next audit cycle.