Diagnose and fix a Fallow false positive or false negative through extraction, resolution, graph, analysis, reporting, and real-consumer verification.
SkillScore87Maintained
Automatic
◆Coding pack
Last verified 9/28/2026
What it changes
Passed 17 of 20 test cases with the skill, and 8 of 20 without it.
Without skill39%
With skill87%
Checks it passed that the model alone missed
The response specifies passing --quiet.
The response specifies passing --format json.
The response specifies using the --quiet CLI flag.
and 13 more in the full report
11 fixed9 passed either way3 still failing0 broke
Pass-rate lift
+48 pp
Tokens
-45%
Agent turns
0%
Response time
-41%
Model gemini-3.6-flashMeasured 9/28/2026v2 · latest is v420 of 23 cases passed with skill · 20 comparable
Tokens, turns and time describe resource use, not a dollar price.
This benchmark measured an earlier version of the skill.
A newer attempt on gemini-3.6-flash on 10/4/2026 did not produce a publishable comparison (not comparable). The snapshot above is the previously measured run.
See it in action
One of 11 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 11
Prompt
During fallow-rs false positive investigation, I noticed that an alert marked for suppression still reaches report assembly. Which layer immediately follows suppression in the pipeline?
What was checked, without and with the skill
Checked
Without skill
With skill
The response identifies filters as the layer immediately following suppression.
Missed
Passed
case-22 · gemini-3.6-flash · 9/28/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md23 linessingle file
Debug a false result
Reproduce on current main with --format json --quiet. Save one command
that shows the wrong result. Change no code before this command exists.
After a restart or an update, clear the .fallow/ cache first.
Reduce to a minimal fixture without losing the behavior.
Write 3 to 5 ranked hypotheses. Each one predicts a result that one probe
can disprove. Test one hypothesis per probe.
Trace the finding through extract, resolve, graph, reachability, analysis,
suppression, filters, and report assembly.
Fix the earliest incorrect layer. Fix the pattern, not the instance (see
test evidence in docs/development/quality-gates.md). After 3 failed
fixes, stop and write down the premise that they share. Question that
premise before a fourth fix.
Add a regression test that fails without the fix.
Run the fixed binary on a real consumer project with a cleared .fallow/
cache and compare old versus new output.
Run review with the affected surface reviewers.
Do not tune expected output around one fixture when the semantic model requires
a broader correction.