diff --git a/README.md b/README.md
index bbfc228..9baaa39 100644
--- a/README.md
+++ b/README.md
@@ -118,7 +118,7 @@ for the full tool catalog, request schemas, and env vars (`PINEFORGE_ALLOW_ANYWH
 
 ## Why PineForge?
 
-- 🎯 **TradingView-exact.** 307 of 311 reference strategies match TV trade-for-trade on the public corpus (the other 4 hold ≥99.9% count parity; 1 further probe is a documented TV-side anomaly at the 1× margin boundary where TV's broker emulator is non-deterministic — engine is correct, and it is excluded from the headline). On a private 415-strategy real-world TradingView scrape, the engine validates **396/396 excellent (100%)** — with **19 TV-side anomalies discovered and documented** during that research. **100 of 100** PineForge excellent vs PyneCore + PineTS on the public three-way benchmark (~167,000 TV trades; PyneCore: 85 of 100; PineTS indicator-only).
+- 🎯 **TradingView-exact.** 307 of 311 reference strategies match TV trade-for-trade on the public corpus (the other 4 hold ≥99.9% count parity; 1 further probe is a documented TV-side anomaly at the 1× margin boundary where TV's broker emulator is non-deterministic — engine is correct, and it is excluded from the headline). On a **closed test set** of 415 community-shared TradingView scripts (kept private under TradingView’s Terms of Service — not redistributable), the engine validates **396/396 excellent (100%)** across **~520k trades** — with **19 TV-side anomalies discovered and documented** during that research. **100 of 100** PineForge excellent vs PyneCore + PineTS on the public three-way benchmark (~167,000 TV trades; PyneCore: 85 of 100; PineTS indicator-only).
 - ⚡ **Microsecond-class.** Median **162× faster than PyneCore** across 99 commonly-timed strategies (full 41,307-bar OHLCV, magnifier-on hot loop; see [benchmarks/results/speed.md](benchmarks/results/speed.md)). Parameter sweeps load one `.so` and re-run with new inputs — no recompile, no fork, no IPC.
 - 🔒 **Stable C ABI.** 28 functions, one header (`<pineforge/pineforge.h>`). Append-only across minor versions, `static_assert`-pinned struct layouts, hidden-visibility hygiene. Drop a strategy `.so` in any harness; it just runs.
 - 🧪 **Reproducible to the bit.** Deterministic float ordering, deterministic bar magnifier, no internal RNG seeded from time. Two runs with the same inputs produce bit-identical trade lists.