165 learning candidates, 14 sources, zero rejections. The batch looked productive. The distribution said otherwise.
i.What landed in the review panel
A learning pipeline extracted candidate insights from 14 creator transcripts. The goal was a usable harvest: insights worth promoting to a knowledge base, with low-quality entries filtered out before any human spent time on them.
The pipeline returned 165 candidates. Formatted correctly. Metadata attached. Titles readable.
Something was off. One candidate’s title was “Click it.” Two others were single-sentence observations with no context. Nothing had been rejected.
The giveaway was not any single bad row. It was the shape of the full set.
ii.Reading the distribution
14 sources. 165 candidates. That is exactly 11.78 per source. In practice, every source had returned exactly 12.
No source produced 8. No source produced 17. No source produced 0. Every source, regardless of length, depth, or quality, produced the same count.
That number is not a product of judgment. It is a quota.
A model exercising real quality filtering will produce ragged counts. Some sources are dense with transferable insight. Some are thin. A talk by a practitioner mid-career who has solved a hard problem in public will yield more than a ten-minute product overview. A model that treats them identically has not evaluated them. It has filled a template.
On AI pipeline quality signals Distribution shape is a faster quality signal than row-level review. Uniform output from a judgment task means no judgment ran.
iii.Why this failure mode is quiet
Row-level review is expensive and slow. A human inspecting 165 rows will eventually catch the thin ones, but they will also spend time on the correct ones. The signal-to-noise ratio of the inspection itself degrades.
The distribution check takes thirty seconds. Count the items per source. Plot or sort them. If they cluster within a narrow band, the pipeline is not filtering. If every source returns the same count, the pipeline has a quota, not a quality bar.
This check can run before any human opens the review panel. It is a cheap gate at the pipeline’s output boundary that converts a slow row-level problem into a fast batch-level verdict.
The 165-candidate batch was rejected at the distribution check. The defect became the next sprint’s fix: replace the count-based extraction instruction with a quality-threshold instruction, and validate that the threshold actually produces rejection on thin sources before the pipeline runs at scale.
iv.The reusable principle
Any AI pipeline that produces structured output from a judgment task should have a distribution check before the output reaches a human reviewer.
The specific signals worth watching:
Uniform counts across heterogeneous sources are a quota flag. Real judgment produces variance.
Zero rejections are a threshold flag. If the model never rejects, the rejection criterion is either absent or ineffective.
Round numbers at the item level (exactly 10, exactly 5) are a prompt-echo flag. The model may be producing the number named in the instruction rather than the number the sources warrant.
None of these checks require reading the content. They require reading the metadata. That is the point: the distribution tells you whether to trust the content before you invest the time to evaluate it.
The pipeline was not broken at the logic level. The architecture was sound. The failure was that the extraction instruction named a count without naming a quality bar, and the model, correctly following the instruction, filled the count every time. The spec was the bug.
Drafted with Bishop, my AI partner.
Words picked, edited, and approved by me.
Model provenance: Claude Code (Claude Opus and Sonnet)