Trust Report
Pix Reconciliation Analyst
v0.1.4 · skill · super-agent-skill
Evidence for v0.1.4
No signed release recorded for this version.
- 2026-10-24T02:33:38.803088+00:00 · mlx-community/Qwen3.5-9B-4bit@8b2b98c00a6b4d291155e4890773ca8f769aee53 · 13/13 passed · 0 failed
The aggregate trust score and usage below can include other versions. Missing evidence is not a failing test.
Trust vector
Published score = quality × confidence. More adversarial runs, case coverage and real-world executions raise confidence — so a few lucky runs can't earn a high score.
Latency (30d)
Robustness
Model heatmap (30d)
No telemetry yet. Connected agents can call report_execution via MCP after using this skill.
Robustness findings
CVE-style report of adversarial failures discovered by the SkillForge red-team pipeline. Public by default — we publish what we find so you can trust what you ship.
Compatibility matrix
Independent cross-model probe — runs every published example on each major frontier model and lets a neutral judge score the outcome. Updated by the SkillForge compatibility sweep.
- passmlx-community/Qwen3.5-9B-4bit@8b2b98c00a6b4d291155e4890773ca8f769aee53/redact-route-generate-v113/13 passed · judge —/100100%
How this trust score is computed
Trust Score v2 is evidence-gated: published score = quality × confidence, where quality is a weighted blend of four dimensions (safety, competence, freshness, coverage). Pass and success rates use the Wilson lower confidence bound, so large samples beat a handful of lucky runs, and an untested package shows Unverified rather than a default number. Real-world signals come from agent executions reported via the MCP report_execution tool. The formula is pure and reproducible offline — see src/lib/trust/scoring.ts.