Case study: A finding that took four passes and two sessions

On 2026-08-12, two sessions revised one finding four times. Two statements were wrong, and each session corrected the other.

Both sessions gave confident answers that a command could disprove. This example applies when you already run more than one session; a solo session has no second reader.

Treat the first finding as a hypothesis. Ask another session to run the check before you file it; see what to do with this.


The sessions were checking whether any automated test read the repository’s always-loaded conventions file.

The four statements

Pass Session The claim Verdict
1 A A docs-only pull request skips the suite, so any gate protecting the file would not run Wrong
2 B Nothing reads the file at all: 76 test files mention it, zero read it Wrong
3 A Two guards read it, both are allowlisted, both run, and neither checks content Correct
4 Joint The defect is that the allowlist is an enumeration, so a check landed outside it goes green by absence Correct, narrower

Passes 2 and 3 each contradicted the previous claim. Pass 4 made the correct claim more precise.

Both sessions were ready to file a wrong finding. Session A would have filed “unguarded”; Session B would have filed “no reader exists”.

A reviewer could refute B by opening one test file. Reading one workflow path filter would refute A.

What moved each pass

Commands changed the claims in passes 1 to 3. Neither session corrected itself by reasoning or rereading; pass 4 narrowed pass 3 without another measurement.

Pass 2 used a grep that required the filename and a read verb on the same line. The matching code built the path first, then read it later through a variable.

The filter could not detect that code pattern. The session treated its empty result as proof that no reader existed.

What the sequence does not establish

Not established Why
A rate One finding is one finding. Nothing here says how often a first statement is wrong
That two sessions are better than one reader The comparison was never run. A single session that re-measured might have reached pass 3
That the channel caused it How the corrections travelled between sessions was not measured
That the correct statement is final Passes 3 and 4 are verdicts as of the date above, and nothing here re-establishes them

The errors were all one shape

Recorded the same day, across the same two sessions:

  • Two false zeros, from commands that failed rather than scanned.
  • A three-rule precedence chain reported from one regex. One rule was tested, and the result was written up as the chain’s behaviour.
  • A breakdown whose eight figures summed to 1202, from a tool whose own output read found 646 section citations. Written up under reconcile the parts against the total.

Each result came from a tool nobody had checked. Every case had an available positive control that would have cost nothing to run.

What to do with this

Do not file a confident first reading before someone checks it.

Keep a single session’s finding provisional until a second session runs the command. If both sessions already exist, send one message.

The second run gives you an independent measurement. In passes 2 and 3, conflicting results exposed the defect.

Limit: this is not a claim that peer review finds everything, and it does not replace the mechanical checks in when you write a guardrail.

Start with the code doing the reading: enumerate from the callee. Opening a guard shows which files it reads.

Finding every possible reader requires a filter over all callers. That broader filter failed here.