Studies you can re-run, not screenshots you have to trust.

Study 1 · Volume profile vs its folklore

Data:

GC=F (COMEX gold futures) 5m bars with real exchange volume, 50 full CME sessions (2026-04-29 → 2026-07-10), frozen in-repo

Sample:

50 session pairs; per-pair distance-band random-level controls (K=1,000)

Control:

Distance-band random levels [0.8d, 1.2d], real-level ±$2 excluded — redesigned after adversarial review broke the first null

Limitation:

One symbol, one 50-session window; bar-resolution volume approximation

Finding: POC touch rate 69.4% vs control 71.5% — the magnet is folklore; the “80% rule” failed all three principled readings; one narrow POC-respect effect graded “suggestive, unproven”

Study 2 · Our own SMC zone score, shuffle-tested

Data:

Same frozen gold tape, re-exported at bar level (13,640 bars), in-repo

Sample:

1,729 zones detected by a line-annotated port of the shipped Pine at defaults; 1,693 resolved

Control:

Scores shuffled across resolved zones ×2,000 (fixed seed) + 12-bar thinning for zone clustering + shipped age-prune robustness

Limitation:

One market and timeframe; a port, not the TradingView runtime; respect definition is mechanical, not a win rate

Finding: Composite score: no better than shuffled (+2.6 pts, p=0.21; thinned −1.9, p=0.72). Two ingredients real: displacement +18.0 pts (p < 0.0005), volume +10.5 (p=0.003)

Study 3 · Multi-timeframe agreement vs the next hour

Data:

Frozen gold tape (observation clock + 15m stack) + frozen 60m/730d and 1d/5y series for honest higher-TF warmup, all in-repo

Sample:

1,131 hourly stack readings (non-overlapping forward windows); 1,124 directional

Control:

Confluence column shuffled ×2,000 + session-cluster bootstrap for serial dependence

Limitation:

One market and window; UTC-bucketed 4h/weekly resamples; hourly readings remain serially dependent after guards

Finding: Full alignment held 28.6% of all hours (median run 3h) — common, not special; alignment level separated nothing (+0.4 pts, p=0.43; CI [−5.2, +10.0]); the 100% alert hit 44.8% vs 48.2% baseline