How we measure accuracy

Every signal passes a progressive, source-anchored review before delivery. Stage 1 verifies each new signal against its own source filing (classification, figures, transaction type, completeness). Anything that does not pass cleanly, plus every high-stakes signal type, escalates to Stage 2 (an independent, stricter review that corrects against the filing where it can) and, if still unresolved, Stage 3 (our deepest audit of the full filing). Deterministic arithmetic and entity checks run first at each step. When even Stage 3 cannot fully confirm a detail, the signal is still delivered but carries a plain-English “Flagged for your review” note — we never withhold a signal, and never present an unverified detail as certain. A separate monthly deep audit — a six-step exhaustive review (classification, figures, transaction type, context, prior-signal consistency, and tone) on a stratified sample of delivered signals. We publish the post-audit accuracy rate rather than a single marketing number, because accuracy varies by signal type and improves over time. See the methodology page for the full detail.

Overall accuracy (0 verified of 33 audited)
0.0%
Flagged for your review
0.0%
Corrected before delivery

Our most recent deep audit sampled 33 signals, but their source filings could not be retrieved from EDGAR for independent verification — so no accuracy rate is published yet. (An unverifiable signal is never scored as 0%.) Verified rates will appear here as signals with retrievable source documents accumulate.

Accuracy by signal type

Accuracy data is being generated. Signal types appear here once at least 3 signals of that type have been verified against their source filings.

Improvement log
2026-07-25 · Form 4
Retroactive fix: comma/decimal mis-parse corrected for MSFT signal 21 (headline showed a decimal share count where commas were mis-read as a decimal point); signal removed from feed (is_duplicate=1).
Impact: Share count and dollar value were ~1,000x too small; corrected going forward by parse_number util + Claude number-formatting rule + comma-safe Form 4 extractor.
2026-07-25 · Form 4
Added explicit same-day exercise-and-sale classification rule. Corrected misclassification of Code M + Code S transactions as open-market purchases.
Impact: Prevents exercise-and-sale transactions from being classified as discretionary open-market purchases
Real-time QA (today)
1
Signals verified today
1
Auto-approved
0
Held for review
0.8
Avg confidence

Full signal classification methodology →