The Metadata You Never See Is Doing All the Work

Anyone who has spent time with digital files knows the strange truth that the visible content is the smaller half. Behind every image sits EXIF data, behind every document sits revision history, behind every download sits a chain of information about where the thing came from and what happened to it along the way. Most people never look. The metadata decides more than the content does.

That principle just cost American healthcare a great deal of money, and the story is a useful reminder for anyone who works with files, feeds, or automated systems of any kind.

The content everyone looked at

American health insurers covering more than thirty million older adults are paid by the government according to the medical conditions documented for each member. Sicker documented populations bring larger monthly payments.

The conditions live in clinical notes, and there are years of them per patient, so software does the reading. For over a decade, that software was evaluated on its content output: how many conditions it surfaced, how fast, at what cost per chart. Dashboards counted findings. Everyone looked at the content.

The metadata nobody kept

What almost nobody understood was the provenance. Where did this specific conclusion come from? Which sentence in which note supports it? Which version of the model produced it, under which rules, and which human confirmed it before submission?

Those questions had no answers, because nothing in the pipeline was designed to store them. And for years it did not matter, because nobody asked.

Then somebody asked. Federal auditors, now roughly two thousand certified coders working a rolling quarterly cycle, began selecting individual outputs and demanding full justifications assembled from data retained at the time. Reviews of three insurance plans published this spring found that 81 to 91 percent of certain sampled high-risk diagnosis codes lacked adequate supporting documentation. A major insurer settled federal claims for 117.7 million dollars.

The content had been fine, mostly. The metadata was missing, entirely, and the missing metadata is what turned a defensible position into an indefensible one.

The rebuild

The correction has been architectural rather than cosmetic. Current evaluations of ai tools for medicare risk adjustment coding accuracy test provenance as a first-class property: every suggested code ships with the exact source sentence, the documentation rule it satisfies, a confidence value, and a record of the reviewer who approved it. Model versions and rule sets are pinned so a decision from three years ago can be reproduced under the logic that existed then.

There is a second requirement worth noting, because it is the one people find counterintuitive. The system must flag conditions that should be removed with the same diligence it applies to finding conditions to add. Investigators treated one-directional correction as evidence of intent, which means the metadata about what the system chose not to flag turned out to matter as much as what it did.

Why this generalises to anything with files

The pattern is not specific to healthcare. It shows up wherever automated systems produce outputs that someone might later question.

Content is what you deliver. Metadata is what lets you defend it. A photograph without provenance is a claim. A dataset without lineage is an assertion. A model output without its derivation is a guess wearing a confidence score, and the moment anyone contests it, the absence of the trail is the whole problem.

The practical version for anyone building or operating systems: decide now what questions someone might ask about a specific output in three years, and store what you would need to answer them. Not aggregate logs, not summary statistics. The per-item trail. It is boring, it is cheap while you are building, and it is close to impossible to reconstruct afterwards.

Healthcare learned this at a price that makes the lesson widely available for free. The stuff you never look at is doing all the work, right up until the day it is the only thing anyone wants to see.