diff --git a/6-Documentation/famm/NUVMAP_DELTA_DAG_SEARCH_COMPRESSOR.md b/6-Documentation/famm/NUVMAP_DELTA_DAG_SEARCH_COMPRESSOR.md
new file mode 100644
index 00000000..3f3dd52d
--- /dev/null
+++ b/6-Documentation/famm/NUVMAP_DELTA_DAG_SEARCH_COMPRESSOR.md
@@ -0,0 +1,145 @@
+# NUVMAP Delta-DAG Search Compressor
+
+## Purpose
+
+Make the NP-search acceleration stack explicitly shardable.
+
+The stack is now:
+
+```text
+SEMANTIC_MASS_ROUTE_PLOW
+→ 16D_CHAOS_GAME_FIELD_SHRINKER
+→ DELTA_STREAM
+→ NUVMAP_DAG
+→ NOGOOD_SCAR_CACHE
+→ EXACT_RECEIPT
+```
+
+The key distinction:
+
+```text
+Delta compression shrinks the path.
+NUVMAP DAG shrinks the search topology.
+```
+
+## Why this tilts the needle
+
+A normal backtracking solver creates a tree. Many branches can produce equivalent remaining subproblems. A tree repeats them.
+
+A NUVMAP DAG stores projected state addresses and merges equivalent frontier basins:
+
+```math
+D_{\mathrm{NUVMAP}}
+=
+(V_{\mathrm{state}},E_{\Delta},h)
+```
+
+Each node is an addressed projected state:
+
+```math
+\nu_t
+=
+h(
+G,
+\text{assignment frontier},
+\text{domains},
+\text{scars},
+\text{residual}
+)
+```
+
+Each edge is only a legal delta:
+
+```math
+e_t=(\nu_t,\Delta_t,\nu_{t+1})
+```
+
+## FAMM object
+
+```math
+\mathfrak C_{\Delta\mathrm{DAG}}
+=
+A_{16}(u_G)
+\otimes
+[
+\Sigma_G
++
+\Sigma_{\Delta}
++
+\Sigma_{\mathrm{NUV}}
++
+\Sigma_{\mathrm{DAG}}
++
+\Sigma_{\mathrm{nogood}}
++
+\Sigma_{\mathrm{resid}}
++
+\Sigma_{\mathrm{receipt}}
+]
+```
+
+## NUVMAP address
+
+```math
+h_{\mathrm{NUV}}(s)
+=
+h(
+\Pi_{\mathrm{frontier}}(s),
+\Pi_{\mathrm{domain}}(s),
+\Pi_{\mathrm{scar}}(s),
+\Pi_{\mathrm{residual}}(s),
+\Pi_{\mathrm{sym}}(s)
+)
+```
+
+This is stronger than raw memoization because it does not ask only whether the state is byte-identical. It asks whether the projected route basin is equivalent.
+
+## Sharding model
+
+| Shard | Contents | Local verification |
+|---|---|---|
+| Base instance shard | graph / exact-cover / SAT instance | hash matches manifest |
+| Route-rule shard | deterministic FAMM route rule and weights | rule hash matches manifest |
+| NUVMAP node shard | projected frontier/domain/scar/residual address | node hash recomputes |
+| Delta-edge shard | legal operation from parent to child | replay parent + delta → child |
+| Nogood/scar shard | failed basin, contradiction, or forbidden transition | scar hash and projection match |
+| Solution-path shard | ordered edge hashes | path joins from root to receipt |
+| Exact receipt shard | final verifier output | residual = 0 |
+
+## Multiplicative gain
+
+```math
+G_{\mathrm{total}}
+=
+G_{\mathrm{route}}
+\cdot
+G_{\Delta}
+\cdot
+G_{\mathrm{DAG}}
+```
+
+The earlier estimate was:
+
+```text
+FAMM route shortening: >= 393.7×
+delta trace shrink:    ~30×
+```
+
+So before DAG merging:
+
+```text
+>= 11,800× effective route-trace shrink
+```
+
+With NUVMAP DAG merging:
+
+```text
+40% redundant frontier states → ~20,000×
+70% redundant frontier states → ~39,000×
+```
+
+These are architecture estimates, not proof of general-case NP collapse.
+
+## No-drift boundary
+
+This is not a P vs NP claim. It is a route/topology/receipt compression layer. Exactness still comes from the final verifier: graph coloring conflict count zero, exact cover residual zero, SAT formula satisfied, Lean proof accepted, or other domain-specific receipt.