diff --git a/README.md b/README.md
index f4d80dc..8675537 100644
--- a/README.md
+++ b/README.md
@@ -86,3 +86,5 @@ docs/superpowers/    design spec + implementation plan
 **Performance defaults** target ~60 FPS on discrete/integrated GPU (developed on Apple M4): desktop `steps=300`, `render_scale=0.75`, `bloom=High`, `disk=High`, `aa=Low`; web drops to `steps=200`, `render_scale=0.5`, `bloom=Low`, `disk=Low`, `aa=Off`. Release profile uses fat LTO + single codegen unit on desktop; the web profile overrides `opt-level="z"` for binary size.
 
 **Phase 4 (future work).** Full exact Kerr Cartesian pseudo-Hamiltonian (Σ/Δ/Carter-separable form) for sub-percent photon-orbit accuracy at high spin, retrograde spin, tilted spin axis, and adaptive integrator *order*. See `docs/superpowers/specs/2026-07-13-interstellar-blackhole-phase2-kerr-design.md` §9.
+
+> An earlier attempt (Stage A: exact Hamiltonian core, CPU + shader mirror) lives on the `phase4-exact-kerr-experimental` branch. Its CPU side reaches sub-percent `b_crit` accuracy (χ=0: 0.08%, χ=0.9: 1.5%) but the shader mirror has an unresolved CPU↔WGSL numerical divergence — the rendered disk shows no gravitational lensing despite the CPU trace showing correct 173° deflection. Resume there with per-pixel debug-color output to localize the divergence.