Clarify structured MoE pruning claim

Lead with the deployment-oriented 25% structured MoE parameter-removal claim, move atomic/group terminology into the method explanation, and clarify that compact checkpoint and kernel benchmarks remain future work.

README.md

@@ -12,31 +12,41 @@ tags:
   - inference
 ---
 
-# Laguna Martini - grouped atomic-expert pruning for Laguna XS.2
+# Laguna Martini - structured MoE pruning for Laguna XS.2
 
 > **25% fewer experts, served straight up.**
 
-This is a provisional research submission for the Poolside Research Hackathon. More precisely, the
-current artifact prunes **25% of Laguna XS.2's 64-wide atomic-expert groups**: `19,968 / 79,872`
-groups across 39 sparse layers. These are child groups inside routed parent experts, not 25% of the
-parent experts.
+This is a provisional research submission for the Poolside Research Hackathon. Laguna Martini
+identifies **25% of Laguna XS.2's routed MoE parameters** for removal while preserving a regular
+structure designed for efficient deployment kernels.
 
 > [!IMPORTANT]
-> The current implementation measures pruning quality, not deployment speed. It preserves Laguna's
-> original top-8 parent-expert router and tensor shapes, then zeroes pruned 64-wide blocks in place.
-> The original kernel still computes those zeroed blocks. A modified grouped MoE kernel is future
-> work.
+> The current artifact measures pruning quality, not deployment speed. It applies the structured
+> pruning mask by zeroing blocks inside the original tensors, so the released Laguna kernel still
+> computes those blocks. A modified grouped MoE kernel and physically compact deployment checkpoint
+> are future work.
 
 ## One-line claim
 
-Using HEAPr-style output-space importance scores, we can prune 25% of Laguna XS.2's 64-wide
-atomic-expert groups while moving MMLU from `0.733514` to `0.725965` and full-cache perplexity from
+Using HEAPr-style importance scores, we identify 25% of Laguna XS.2's routed MoE parameters for
+structured removal while moving MMLU from `0.733514` to `0.725965` and full-cache perplexity from
 `10.458797` to `11.783569`.
 
+## What structured pruning means here
+
+Laguna routes tokens to parent experts. Inside each parent expert, the computation can be decomposed
+into smaller atomic contributions. We sort those contributions by importance and bundle them into
+regular 64-wide blocks. The 25% pruning mask removes `19,968 / 79,872` of those blocks across 39
+sparse layers while retaining at least one block in every parent expert.
+
+This grouping keeps the pruned layout structured enough for a future deployment kernel to skip
+removed blocks. It is different from deleting 25% of the routed parent experts outright.
+
 ## Method
 
 Laguna XS.2 is a mixture-of-experts model with 39 sparse layers, 256 routed parent experts per sparse
-layer, and 512 atomic contributions per parent expert. We adapt
+layer, and 512 atomic contributions per parent expert. An **atomic expert** is one independently
+scorable contribution inside a routed parent expert. We adapt
 [HEAPr: Hessian-based Efficient Atomic Expert Pruning in Output Space](https://arxiv.org/abs/2509.22299):
 
 1. Decompose each routed parent expert into atomic expert contributions.
@@ -65,10 +75,10 @@ The 25% keep mask removes exactly `19,968` of `79,872` groups. Per-layer mask st
 Loss and perplexity use a held-out cache of 1,024 sequences x 4,096 tokens. MMLU uses the full
 14,042-example zero-shot task suite.
 
-### Native grouped pruning sweep
+### Original-routing structured pruning sweep
 
-The primary runtime keeps Laguna's original top-8 parent routing semantics and zeroes pruned child
-groups in place.
+The primary evaluation path, called **native** in the code and reports, keeps Laguna's original
+top-8 parent routing semantics and zeroes pruned blocks in place.
 
 | Pruned groups | Mean loss | Perplexity | Perplexity delta |
 | ---: | ---: | ---: | ---: |
@@ -82,13 +92,13 @@ The sweep deltas are rounded for readability. The detailed report records both t
 baseline and a directly comparable explicit-static-cache baseline; their perplexities differ by
 less than `0.04%`.
 
-### What we tried: repacked child routing
+### What we tried: repacked routing
 
-We also repacked the model into independently routed 64-wide child experts and selected a fixed
-top-64 retained child groups per token. This is useful as an exploratory runtime, but it changes the
-post-pruning behavior: selected groups can come from parent experts that the original top-8 router
-would not have selected. It therefore answers a different question and performs worse than native
-group pruning.
+We also tested **repacked** routing: treating each retained 64-wide block as an independently routed
+mini-expert and selecting a fixed top-64 blocks per token. This is useful as an exploratory runtime,
+but it changes the post-pruning behavior. Selected blocks can come from parent experts that the
+original top-8 router would not have selected. It therefore answers a different question and
+performs worse than the original-routing evaluation path.
 
 | Repacked child groups pruned | Mean loss | Perplexity | Perplexity delta |
 | ---: | ---: | ---: | ---: |
@@ -102,12 +112,12 @@ pruned child-group work.
 ## Future work: sentinel-group kernel
 
 The next implementation step is a modified grouped MoE kernel. For each selected parent expert, its
-pruned child-group indices should point to sentinel groups instead of materialized expert blocks.
-The kernel can then avoid loading and computing pruned groups while preserving the original parent
+pruned block indices should point to sentinel groups instead of materialized expert blocks. The
+kernel can then avoid loading and computing removed blocks while preserving the original parent
 router decisions.
 
-No parameter-count reduction, memory reduction, or runtime speedup is claimed until that kernel is
-implemented and benchmarked.
+The structured mask identifies the removable MoE parameters. Realized checkpoint-size, memory, and
+runtime improvements remain to be measured after the deployment checkpoint and kernel exist.
 
 ## Reproducibility artifacts
 

results/summary.json

@@ -1,6 +1,6 @@
 {
   "base_model": "poolside/Laguna-XS.2",
-  "method": "HEAPr-style output-space atomic-expert scoring with importance-sorted 64-wide native groups",
+  "method": "HEAPr-style output-space scoring with importance-sorted 64-wide structured MoE pruning blocks",
   "status": {
     "optimized_kernel": "pending",
     "gsm8k_cot": "pending",
@@ -68,5 +68,5 @@
       "num_fewshot": 0
     }
   },
-  "caveat": "The current implementation preserves original parent routing and tensor shapes, zeroes pruned blocks in place, and does not yet realize runtime speedup."
+  "caveat": "The structured mask identifies 25% of routed MoE parameters for removal. The current implementation preserves original parent routing and tensor shapes, zeroes pruned blocks in place, and does not yet realize checkpoint-size, memory, or runtime improvements."
 }