What happened?
Confidence is assigned primarily by absolute score thresholds (score >= 14 → high) with no “only the decisive leader is high” rule. On real tasks this labels the entire top-8 as high confidence, which teaches agents that every listed file is equally safe to edit.
Example (Zod, v0.7.4):
npx -y @aryam/fixmap@0.7.4 plan --repo https://github.com/colinhacks/zod \
--issue "stringProcessor produces invalid JSON Schema" --format json
Observed: highCount=8/8 while topGap=19 and clustered=false. Top-1 is clearly ahead, but files #2–#8 still say high.
confidenceForEntry in packages/core/src/rank.ts never looks at rank position or margin vs runner-up (except clustering caps).
How can we reproduce it?
Zod command above; count confidence === "high".
What did you expect instead?
Reserve high for the leading file (or files within a small score epsilon of #1 / with definition-site evidence). Demote the rest to medium/low so “high” remains a scarce signal. Matches the product claim that high ≈ lead to verify, not a blanket on the whole list.
FixMap version or commit
v0.7.4
Where did this happen?
CLI
What happened?
Confidence is assigned primarily by absolute score thresholds (
score >= 14→ high) with no “only the decisive leader is high” rule. On real tasks this labels the entire top-8 as high confidence, which teaches agents that every listed file is equally safe to edit.Example (Zod, v0.7.4):
npx -y @aryam/fixmap@0.7.4 plan --repo https://github.com/colinhacks/zod \ --issue "stringProcessor produces invalid JSON Schema" --format jsonObserved:
highCount=8/8whiletopGap=19andclustered=false. Top-1 is clearly ahead, but files #2–#8 still say high.confidenceForEntryinpackages/core/src/rank.tsnever looks at rank position or margin vs runner-up (except clustering caps).How can we reproduce it?
Zod command above; count
confidence === "high".What did you expect instead?
Reserve
highfor the leading file (or files within a small score epsilon of #1 / with definition-site evidence). Demote the rest to medium/low so “high” remains a scarce signal. Matches the product claim that high ≈ lead to verify, not a blanket on the whole list.FixMap version or commit
v0.7.4
Where did this happen?
CLI