The Empty Stage-1 Output: When a Pipeline Fails Without an Error

A JSON file landed on the desk this afternoon. The schema was complete — braces in place, fields named, brackets closed. Inside, every field was empty. Article Title: N/A. Article Source: N/A. Article Type: Unclassified. Information Points: an empty list. Against Source Quality it said "judge from the source fields" — except there were no source fields to judge from. One value in the entire document carried weight: Domain Label: cricket_world. I read it twice, then wrote one line in the notebook: this is not a broken output, it is a valid shell — schema-valid, content-empty.

What Stage-1 was actually asked to produce

Stage-1 does no creative work. It pulls fact-units out of raw text. Its output is a fixed list: title, source, publication date, article type, one-sentence summary, author stance, article purpose, information_points, entities, time sensitivity, source quality.

One field carries the load: information_points. Each point has three parts — claim, source_field (whose claim it is: official release, match report, or author inference), and timestamp where relevant. An empty list means not a single fact-unit survived extraction.

Every other field depends on that list. The Entities instruction reads "identify from the information points above" — nothing above, so no object to identify. Time sensitivity cannot be assessed without a publication date. Source quality cannot be assessed without a source. Author stance and purpose cannot be assessed without sentences. The entire analysis chain hangs on one list, and the list is empty. So there is no analysis here, only a filled-in template.

There is an internal inconsistency worth logging. Domain Label: cricket_world is the only populated value, but with Article Type Unclassified and zero information points, there is no internal evidence for it. Whether it came from body extraction or a routing preset cannot be determined. The one value in the file is the one value that cannot be verified.

Where 2,368 words come from

The arithmetic is simple. A decent match report carries at least one verifiable fact every 100 to 150 words — a scorecard number, an over's ball-by-ball shape, a field setting, a quotable line. That puts 2,368 words at roughly 40 to 60 fact-units. Available right now: zero.

I cannot close that gap with style. If I sat down and wrote cricket-shaped prose — a formation, a bowler's angle, a middle order — it would read fine and every sentence would be invented. That error does not get caught in editing, because it is plausible. Which is exactly what makes it the most expensive kind.

The real risk is not failure but invisible failure

The pipeline did not report a failure. It returned something that satisfies every schema condition. In data engineering that distinction is decisive: loud failures get fixed, silent ones get shipped.

A 200 status with information_points: [] counts as success in monitoring. Downstream, two paths remain. The writer fills the gap with inference, and the inference becomes plausible fiction in print. Or the writer stalls and the deadline goes. Either way Stage-1 caused it and the desk pays.

The fix is not in the prompt, it is in the assertion. Three hard checks belong on the Stage-1 output:

• Fail if the length of information_points is zero.
• Fail if title is N/A.
• Fail if any information_point arrives without a source_field.

A fourth, softer check: if information_points are populated but entities is empty, raise a warning — any real cricket fact contains at least one name, one venue, or one match.

Takeaway

Send one of three things and the report can be written. The original text, most directly — I will extract from zero. Or a filled Stage-1 with at minimum title, source, publication date, article type, information_points carrying claim and source_field, and entities. Or, if the extraction pipeline itself fell over, the error log or the raw fetch, and I will help locate whether the fault is at input or extraction.

Until one of those arrives, 2,368 words is not available. That is not stubbornness; it is the condition on which a byline rests. We do not write a bowler's pace from inference. The same rule applies to text.