The Error Cascade

Five stages, each one good. Chain them, then choose where the humans look. Watch a pipeline of individually strong stages deliver a mediocre corpus, and see why the errors land on exactly the documents that matter.

The pipeline · five stages

Good stages, compounding losses

Each stage keeps most documents right. But accuracy multiplies through the chain, so five strong stages still hand you a corpus with a real error rate. Drag any stage.

How the errors fall

Spread evenly Cluster on bad docs

The toggle. Switch from "spread evenly" to "cluster on bad documents" and the headline accuracy barely moves, yet the errors pile onto the worst 5% of the corpus. The average hides the harm.

The review gate · where humans look

Buy back fidelity per hour

Put a human-review gate after one stage, and route the lowest-quality documents to it first. The same review budget buys back far more when the errors are clustered.

Review after which stage

Review budget 5%

no reviewworst 20% reviewed

Route by quality, not at random. The budget reviews the worst documents first: all of the awful band, then into the degraded band. In clustered mode that tail carries most of the errors, so each review-hour is worth several times more.

The numbers · expected values over 10,000 documents

Readouts

End-to-end intact 77.4% before any review

Documents with an error 2,262 at least one, of 10,000

Errors on the worst 5% 5.0% share carried by the awful band

Fixed per review-hour 6.8 errors caught per hour

End-to-end after review 78.5% intact once review fixes its queue

The corpus, worst on the right

Each cell is 100 documents, ordered clean → degraded → awful. Brighter red means more expected error. The bracket shows the tail sent to human review.

Expected error density Routed to review

The stage waterfall

One hundred percent of documents enter. Each bar is the share still fully intact after that stage; the red slice above it is the loss that stage adds.

Still intact Lost at this stage