What this is

RuView's WiFi "edge skills" — the 64 tiny detectors that read the WiFi in a room and flag things (presence, breathing rate, gestures, intrusion, etc.) — are now (a) wired into one runnable pipeline and (b) honestly accuracy-tested. Done in PR #1044 (merged).

Plain-language writeup: https://gist.github.com/ruvnet/3e7caf63e3d45d3065b2ba4b117bcef0

Why it was needed

• The 64 skills were all real, working code — but only 3 ran on the default per-frame path; the rest were never connected into one system.
• Their accuracy was unproven. For a project people have called "AI slop," that's the gap that matters most.

What landed (PR #1044)

1. One pipeline runs all 64 skills. A unified EdgePipeline registers every skill and runs them on each WiFi frame, collecting all detections into one tagged stream (~9,200 events over a 300-frame demo). The 5 medical skills stay behind an opt-in feature flag so they can't ship by accident.

2. Honest accuracy, three buckets:

• ✅ Measured & working (10 skills): presence, occupancy, intrusion, hidden-breathing, vital-rate, gesture enrollment, person-matching (0 ID swaps/40 frames), peak/zone localization — 100% on synthetic ground-truth tests.
• ⚠️ Caught NOT working (published, not hidden): exo_time_crystal can't separate a sub-harmonic from a normal rhythm; sig_sparse_recovery slightly worsens its target (−2.2%). Reported as negatives.
• 🚫 Can't be honestly tested without real data (10 skills): seizure, sleep-apnea, cardiac, respiratory, gait, weapon-detect, emotion, happiness, dream-stage, sign-language. No accuracy claimed — each lists the real labeled data it needs; all stay disclaimer-gated.

Reproduce it yourself

git clone https://github.com/ruvnet/RuView && cd RuView/v2/crates/wifi-densepose-wasm-edge
cargo test --features std --test synthetic_validation -- --nocapture

Results: benchmarks/edge-skills/RESULTS.md. Library: 631 tests pass (669 with medical tier), 0 failures, 0 warnings.

The point

Real code, measured where measurement is honest, openly truthful about the rest — including the two detectors we caught not living up to their names. That's the opposite of "AI slop."

Tracking only — the work is merged. Follow-on (out of scope here): real-hardware accuracy needs the ESP32/WASM3 on-target path + labeled datasets for the gated skills.