diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 0000000..00f121a
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1,37 @@
+# CLAUDE.md
+
+Project memory for Claude sessions working in this repo.
+
+## Memory: NPF normalization — ADF vs second-moment (2026-07)
+
+Context: a Discord post analyzed **NPF** (net position fraction = net
+non-commercial positioning / open interest) built from CFTC Legacy COT data
+(the same data `src/cotdata/providers/cftc.py` downloads).
+
+Feedback received, and accepted as correct:
+
+- Dividing net positioning by OI is a *level-stationary* transformation, but
+  raw net is already ADF-stationary in most markets — positioning oscillates
+  around a mean by construction — so the gain shown by an ADF test is modest
+  (it mainly rescues markets with strong secular OI growth).
+- The real benefit is **second-moment** and ADF cannot detect it: net measured
+  in contracts scales with market size (±20k contracts was an extreme in 1995,
+  noise in 2015), so full-history raw net is a mixture of a narrow early era
+  and a wide late era. That inflates kurtosis and lets the recent
+  high-variance era dominate percentile-based thresholds. NPF collapses the
+  eras onto one scale: era variance stabilizes, kurtosis drops, and thresholds
+  (and the "gray zone" between long/short triggers) calibrated on full history
+  stay meaningful across decades.
+- Remaining caveat: NPF fixes *scale*, not *composition*. The 2004–2006
+  commodity-index influx changed who holds positions in ags/energy; no
+  denominator fixes that structural break — consider post-break calibration
+  windows.
+
+Verification plan (script `npf_check.py`, handed off to a trusted-network
+session via the trigger "NPF variance-stability check" because some sandbox
+network policies block cftc.gov): for gold, silver, WTI, nat gas, corn,
+EUR FX, 10Y note, and E-mini S&P over 1990–present, compare raw net vs NPF on
+(1) ADF p-values, (2) era variance ratio = std(last third of history) /
+std(first third), (3) excess kurtosis. Expected pattern if the feedback holds:
+similar ADF p-values with most markets stationary in both forms; raw variance
+ratios well above 1 vs NPF ratios near 1; lower kurtosis under NPF.