Cashflow Check Waterfall

Do earlier parsers agree with the last parser that ran? Skip-check and improvement opportunities per bank

snapshot

Range

Attempts per day

Compared attempts (match / mismatch vs. the truth parser) in solid colors · truth-parser runs muted

Waterfall & agreement by parser

Enabled bank-statement parsers in run order · the truth parser is per-unit: whichever parser ran last on that unit, not necessarily the last in this list

Bank × parser agreement matrix

Agreement with the truth parser over compared attempts · green = skip-check candidate · red = improvement opportunity · "truth n=" cells only ever ran last on their units (nothing to compare)

Improvement opportunities

Skip-check opportunities

Methodology. A unit is one file × account × statement period; a parser's result for a unit combines all its checks. The truth parser is per-unit: the last parser that actually ran on that unit — which may be earlier in the waterfall than the configured last parser (with fail_on_cashflow off, most units stop at their first parser). Every earlier parser's attempt is scored against it: a match when both pass, or when both extract the same cashflow (calculated credits and debits within tolerance) — matching numbers on a failing document still count as a parsing success for the earlier parser; a mismatch when its extracted cashflow differs from the truth parser's. A bank whose early parser consistently matches is a skip-check opportunity (trust the cheap parser without the cashflow gate); consistent mismatches are a parser improvement opportunity. A truth parser's own runs are not scored (nothing later to compare against); its failures are user errors (bad documents) — excluded from all metrics, shown muted for volume. The waterfall table shows the enabled parser sequence and fail_on_cashflow flags; while flags are off, failures don't cascade, so most units see one parser and comparisons are sparse. Banks are identified by institution_name (canonical) with bank_name fallback. Unlike the volume dashboards, weekends are included here: checks are pipeline events and agreement is valid signal whichever day the waterfall ran. Source: waterfall_cashflowcheckresult + waterfall_parser (enabled bank-statement parsers, mocks excluded), laminr-prod · Eastern-time days.