Drop the low-signal reviewers from gadfly's own swarm per the scoreboard:
- m5/qwen3.6:35b-mlx (slow local Mac; removes the last foreman/local lane,
so GADFLY_ENDPOINT_M5 + the m5 concurrency lane are gone)
- gemma4:cloud, gpt-oss:120b-cloud, kimi-k2.7-code:cloud (weak cloud models)
Remaining: minimax-m3, glm-5.2, glm-5.1, deepseek-v4-pro, nemotron-3-super,
qwen3-coder:480b + claude-code/sonnet — 6 cloud + the Claude Code engine,
all fast cloud lanes. Takes effect once this merges to main.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A lens whose verdict is "No material issues found" still tends to write
path:line references — but as verification notes ("verified X at
file:line is safe"), not problems. The telemetry was extracting those as
findings, which (a) pollutes the gadfly-reports store with false
positives and (b) unfairly penalizes thorough reviewers that do clean
passes — the FP penalty hit clean security passes from claude-code/sonnet,
deepseek, and minimax even though they correctly found nothing.
emit() now skips findings extraction for a clean-verdict lens (the run is
still recorded). Surfaced by grading the dogfood reviews: a large share
of "false positives" were exactly these clean-verification bullets.
Added TestEmit_SkipsCleanVerdictLens; README telemetry section updated.
gofmt clean, go vet quiet, go test -race green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
steve