Attribution produces three categories: AI-authored, human-authored and unknown. There is constant pressure to make the third disappear, because a chart with two bars looks finished and a chart with three looks like work in progress.
Collapsing it is the worst decision available.
What ends up in unknown
Code that predates recording. The largest bucket at first, and it shrinks only with time.
Code from unmonitored machines. A developer without the recorder installed produces code nobody can place.
Heavily reformatted code. Matching is on content; a formatter that rewrites lines breaks the match.
Genuine ambiguity. Text that appears in several interactions and cannot be resolved by adjacency.
Why folding it into human is wrong
It is the tempting move: if we cannot prove an AI wrote it, assume a person did. It is wrong for three reasons.
It is not true. Much of that residue is AI-written code from before recording started, and assigning it to humans understates AI authorship exactly where the history is longest.
It hides coverage problems. If unmonitored machines quietly increase the human column, you will never notice that half your engineers are not recording.
It destroys credibility. The first auditor who tests the number against a file they know will find code attributed to a human who did not write it, and every other figure in the report becomes suspect.
What a large unknown actually tells you
It is diagnostic rather than embarrassing. A high unknown share on an old repository means the history predates recording, which is expected and shrinks over time. A high share on a new repository means coverage is broken — somebody is not recording. A share that grows suddenly usually means a large refactor or a formatter change.
Each of those is useful. None of them is available if the residue has been redistributed.
The reporting rule
Three numbers, always, with the definition next to them. If a figure was not measured, leave it blank and say why. Never default an unmeasured value to zero, and never let silence read as a pass.
It makes for a less confident report and a much more useful one.