Engineering notebook

Measuring robustness across degradation severity

Keep recognition, serialization and cohort composition visible.

01

Recognition and sequence are different questions

The scoring policy makes order-independent token F1 the primary recognition metric. CER and WER remain reported as reading-order-sensitive sequence metrics. No composite score combines them.

02

Read the curve with its denominator

The historical cohorts contain 250 clean, 182 subtle, 147 moderate and 421 severe samples. Plotting their means shows performance by cohort, not the deterioration of one matched document over time.

Relative severe retention is severe mean token F1 divided by clean mean token F1. It is a descriptive ratio of unpaired cohort means, not a paired treatment effect or training uplift.

03

Keep diagnostic failures visible

Report empty responses and exact-token/sequence-error cases separately. A successful request is not necessarily a successful recognition result. A durable raw response is evidence even when its recognized text is empty.

Condition memberships overlap for mixtures. Summing condition counts does not yield a unique sample total, and individual condition means should not be treated as independent experiments.

Build with evidence

See what breaks
before production does.

Run StressBench against your document model and get a condition-level robustness readout.