CASCADE Routing audit, KYC extraction Seal 1151f5a1cfaae0c0 · measured, then frozen
Finding 01two rates over one deskFinding 02one field changes readerFinding 03the break-even for silenceFinding 04withheld for non-reproducibilityFinding 05what the engagement delivers

Both numbers are true. Only one leaves your desk.

Cost separates the two routings. Accuracy does not.

A blank gets read again. A wrong value gets filed.

Every count held across two passes. Every duration moved.

Every routing enumerated. One report you can argue with.

Seven readers down, five fields across. Five green cells mark the published routing, one per field.Seven readers down, five fields across. One cell steps down: the name field moves from large to gen-4b. That is the only difference between the two routings.Seven readers down, five fields across. Where the rules reader returned 0%, it now returns nothing: the two red cells become gaps. Abstention, drawn.Seven readers down, five fields across. No cell is green: nothing here is held up as a result.Seven readers down, five fields across. Every cell is green: the full enumeration, one routing at a time.
Figure 2The two rates, on the same drawn scale
60708090100
17.7 points
94.4%

The dashboard mean.

76.7%

On your desk: all five fields right on the same file, 92 of 120. Wilson 95%: 68.3 to 83.3.

Figure 2Cost per 100 000 documents, on assumed prices
050100150200
3.5× cheaper
$191

The published routing.

$54

The routing aimed at the file: no file gets worse, 3 gained, 0 lost. p = 0.25, accuracy undecided.

Figure 2Precision, before and after silence
0255075100
after abstention
30%

Every value delivered.

62.3%

With silence, on the hard corpus: 33 of 53 right. 85 wrong removed, 12 right lost.

Figure 2Divergence between two passes of the same bench, in per cent
0204060
withheld
identical

Every count, to the digit.

16–60%

Every median duration moved. 32 conclusions retracted, 11 caught first.

Figure 2The routing space, enumerated end to end
016 807
the full span
16 807

Every combination of seven readers over five fields. Not a sample.

120 files · 596 tests

Held out and frozen; the tests run on your machine.

What this does not proveNeither number is wrong: they answer different questions, and only one of them is the question at your desk.Figure 1 — Five green cells mark the published routing, one per field.

What this does not proveThe sample decides cost, not accuracy: cheaper is decided, better is not — and both dollar figures rest on assumed prices.Figure 1 — One cell steps down: the name field moves from large to gen-4b. That is the only difference between the two routings.

What this does not proveWhere the break-even sits is your decision, not ours: we publish the curve, you place the point.Figure 1 — Where the rules reader returned 0%, it now returns nothing: the two red cells become gaps. Abstention, drawn.

What this does not proveRemoving a measurement is a discipline, not a guarantee about the figures we keep.Figure 1 — No cell is green: nothing here is held up as a result.

What this does not proveThe corpus here is synthetic: which is exactly why the real measurement runs on your records, on your machine.Figure 1 — Every cell is green: the full enumeration, one routing at a time.

On your records, on your machine. Nothing leaves the network.