Skip to main content
Understanding coverage describes demonstrated understanding across eligible source code at a fixed revision. Open Repositories, choose a repository, and browse folders or a file to see the available evidence. Repository summaries use the default branch’s resolved revision. A selected branch uses its own revision. A result’s source view shows the author’s evidence at that result’s original revision.

​
Read the measures

For example, a file has 100 eligible lines. There is evidence on 20 lines, with an average understanding estimate of 75% on those lines. Its evidence coverage is 20%, understanding on measured code is 75%, and understanding coverage across the file is 15%. These measures summarize evidence from explanations. They are not probabilities that a person understands all possible behavior.

​
Team and personal views

You uses your current, applicable evidence. Team takes the highest individual understanding estimate for each eligible line among current authorized members. Folder and repository totals combine the underlying eligible and demonstrated units. They do not average file percentages. The source heat map can show average understanding among measured members. That display does not change the canonical team total.

​
Distinguish evidence states

  • No evidence yet: Eligible code has no usable understanding evidence. This is different from a measured zero.
  • Measured zero: An explanation was graded and its applicable evidence contributes zero understanding.
  • No eligible code: The selection has no eligible source lines, so a coverage percentage does not apply.
  • Unmappable: Eligible source has no safe question target. The file has no percentage.
  • Calculating, Updating, Incomplete, or Unavailable: A current value is not ready or cannot be read. These states do not mean zero.
Blank lines and comment-only lines do not contribute eligible units. Documentation, generated output, dependencies, and other excluded files do not contribute. Files that cannot be analyzed are skipped. Code without an explanation can still be eligible and remain without evidence.

​
How explanations contribute

Questions assess explicit source behavior and lines. A citation or supporting file alone does not award understanding coverage. External context from Linear, Notion, or Mintlify can ground a question, but it does not become repository source coverage. A later completed explanation for the same behavior and applicable lines replaces the earlier estimate for those targets. Questions about another behavior add evidence for that behavior. Coverage can increase or decrease; repeated equally strong explanations do not accumulate extra credit. A follow-up check affects only the behavior and lines it assesses. Earlier results remain immutable.

​
When code changes or merges

Historical results stay attached to their original revision. Current repository coverage uses evidence that applies to the displayed revision. Second Flow carries evidence forward only when it can verify revision lineage and the relevant source correspondence. Verified unchanged source or a pure rename can preserve evidence. Merge, squash, and rebase histories can qualify when the original pull-request head and merged result are verified. Changed behavior or supporting context, ambiguous copies, and unverified refactors can prevent evidence from carrying forward. Similar-looking unmerged code does not receive transferred evidence. New eligible code starts without evidence. Carrying evidence forward does not rewrite an earlier result or make an old passing result satisfy the policy for a new pull-request head.

​
Compare with ownership-signal coverage

Analysis & Signals also reports ownership-signal coverage: the share of eligible, ingested pull requests with a submitted explanation and recorded result on the current head. That report counts the presence of a result, regardless of passing or failing the policy. Source understanding coverage measures demonstrated understanding across eligible code. A pull request can have a recorded result while much of the repository still has no evidence.