Add guide for OpenAI unit-distance disproof as FAMM solve pattern

(cherry picked from commit afc269c076d6026cf6ccb974d673b084c4976238)

docs/research/OPENAI_UNIT_DISTANCE_FAMM_GUIDE_2026_05_20.md

@@ -0,0 +1,172 @@
+# OpenAI Unit-Distance Disproof as FAMM Solve Pattern
+
+Date: 2026-05-20
+Status: REVIEWED_EXTERNAL_RESULT_REFERENCE
+
+## External Source
+
+OpenAI article:
+https://openai.com/index/model-disproves-discrete-geometry-conjecture/
+
+Problem originally posed by Paul Erdos in 1946.
+
+---
+
+# Core Event
+
+A model-guided search discovered a counterexample family disproving the long-believed conjectural upper bound:
+
+u(n) <= n^(1+o(1))
+
+for the planar unit-distance problem.
+
+The discovered route used:
+
+- algebraic number theory
+- infinite class field towers
+- Golod-Shafarevich theory
+- richer number fields than Gaussian integers
+
+rather than the historically dominant square-grid constructions.
+
+---
+
+# Why This Matters for FAMM
+
+This is a canonical solved-scar pattern.
+
+The important lesson is NOT:
+
+"AI solved geometry."
+
+The important lesson is:
+
+cross-domain structural routes can dominate local intuition.
+
+---
+
+# FAMM Interpretation
+
+Native problem domain:
+
+- discrete geometry
+- planar point configurations
+- unit-distance counting
+
+Winning route domain:
+
+- algebraic number theory
+- class field tower structure
+- symmetry-rich number fields
+
+Meaning:
+
+The low-obstruction proof manifold existed outside the community-default search basin.
+
+---
+
+# Underverse Interpretation
+
+This is an explicit Underverse-style move.
+
+The model did not strengthen the dominant construction.
+
+Instead:
+
+- it searched outside the expected manifold,
+- bred counterexample constructions,
+- discovered an admissible family,
+- then verified the route mathematically.
+
+---
+
+# Architecture Mapping
+
+| External Result | FAMM Stack |
+|---|---|
+| conjecture | obstruction field |
+| counterexample family | braid/eigensolid family |
+| algebraic-number-field route | remote NUVMAP sector |
+| model search | BraidStorm traversal |
+| construction selection | eigensolid closure |
+| external mathematical verification | receipt verification |
+
+---
+
+# Core Search Lesson
+
+Do not overweight dominant local routes.
+
+Preserve low-prior remote manifold routes when they exhibit:
+
+- high symmetry
+- low residual pressure
+- reusable invariants
+- unusual closure behavior
+
+This should become a permanent FAMM heuristic.
+
+---
+
+# Attention Residuals Interpretation
+
+This result strongly supports selective residual retrieval.
+
+The successful route behaved like:
+
+- a distant off-diagonal retrieval,
+- not a local next-step continuation.
+
+Meaning:
+
+softmax-selected remote retrieval may outperform dense local accumulation.
+
+Project translation:
+
+alpha_(t,i) = softmax(M_i - Omega_i + S_i)
+
+should preserve rare but structurally rich remote manifold sectors.
+
+---
+
+# NUVMAP Lesson
+
+The proof route effectively traversed:
+
+Discrete Geometry
+→ Algebraic Number Theory
+→ Symmetry Construction Space
+→ Infinite Family Closure
+→ Counterexample
+
+That is a sparse topological jump.
+
+NUVMAP should therefore preserve:
+
+- remote-domain adjacency,
+- not merely local similarity.
+
+---
+
+# Eigensolid Lesson
+
+Finite examples are weaker than reusable infinite-family closures.
+
+Eigensolid ranking should therefore prioritize:
+
+- reusable construction families,
+- scalable symmetry structures,
+- infinite admissible replay.
+
+---
+
+# Final Keeper Lesson
+
+The shortest path to a proof or disproof may live outside the visible problem manifold.
+
+FAMM should therefore:
+
+- preserve remote structural routes,
+- store solved counterexample scars,
+- weight reusable symmetry constructions highly,
+- and treat low-obstruction cross-domain jumps as valuable.