From 1f3278be7e40d680e5c868502da399138fa65637 Mon Sep 17 00:00:00 2001
From: david6666666 <530634352@qq.com>
Date: Thu, 15 Oct 2026 22:40:06 +0000
Subject: [PATCH 1/6] Record C64 reference accounting and native feasibility
 evidence

Signed-off-by: david6666666 <530634352@qq.com>
---
 evidence/drama3-b300-scale150/patch/README.md |  4 +-
 .../results/diagnostic-lessons.md             | 60 +++++++++++++++++++
 .../drama3-b300-scale150/results/status.json  | 26 ++++++--
 ...concurrency-scaling-state-and-overlap.json | 26 ++++++--
 4 files changed, 102 insertions(+), 14 deletions(-)

diff --git a/evidence/drama3-b300-scale150/patch/README.md b/evidence/drama3-b300-scale150/patch/README.md
index effc32a13..6cd5833a5 100644
--- a/evidence/drama3-b300-scale150/patch/README.md
+++ b/evidence/drama3-b300-scale150/patch/README.md
@@ -4,6 +4,6 @@ No winning candidate is nominated. Each experiment records its source revision b
 
 The implementation and rejected trials are reviewable in
 [runtime draft PR #1](https://github.com/david6666666/drama-3-single-b300-inference/pull/1).
-Published evidence is pinned at
-[`2669d4c`](https://github.com/david6666666/drama-3-single-b300-inference/tree/e58c8418005af8972d1616d5368009bc07dbe651/evidence/20261015-scale150).
+Latest published evidence is pinned at
+[`ab2d901`](https://github.com/david6666666/drama-3-single-b300-inference/tree/ab2d9015d6b0115530e675f25bbf5f188b1d65b1/evidence/20261016-ceiling).
 This is an evidence publication revision, not a nominated performance winner.
diff --git a/evidence/drama3-b300-scale150/results/diagnostic-lessons.md b/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
index c10272192..ae7eb2e13 100644
--- a/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
+++ b/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
@@ -177,3 +177,63 @@ all-concurrency acceptance or proof about every historical quality failure.
 All new raw groups, diagnostics, source snapshots and limitations are indexed in
 [the published mixed investigation](https://github.com/david6666666/drama-3-single-b300-inference/tree/dfdd598a9810cf6e7b8aa6a102f6ab9270ff9cc6/evidence/20261016-mixed).
 The runtime PR remains draft and the campaign remains INCONCLUSIVE.
+
+
+## C64 reference-gap investigation, 2026-10-16
+
+The target remains6765.407 frames/s with native full-top8 computation and the
+frozen quality gates. The measured isolated B64 reference is approximately7734
+frames/s; it is neither a hardware upper bound nor a service guarantee. The
+complete1088 asynchronous trace replaces old synchronous timing proportions.
+Client continuous in-flight is about63.98, with only about6ms equivalent missing
+client-slot time. Client dispatch starvation is not supported as the major gap.
+
+Rich tracing separates183228 real prompt tokens from208384 captured rows and
+2122 codec frame-bucket groups from2265 actual graph calls. Only143 calls are
+avoidable same-bucket partitions; frame-bucket diversity accounts for833 groups
+beyond the emit count. Neither percentage is a time-saving estimate. The trace
+records multi-step decode explicitly instead of equating graph calls with steps.
+
+Compact GDN passes2099 exact checks. Mixed-layout and empty-convolution changes
+pass339 checks each and339 combined. Nevertheless, their full one-run GPU0
+candidate is4306.693 frames/s versus4371.393 for the same-source control; both
+score all1088 outputs and pass WER/SIM. The candidate's frame count is2.106%
+higher and drain1.092s longer. Preserve the failed full throughput score and the
+fixed-state graph improvement as different observations.
+
+Deferred codec host submission preserves the original producer event and passes
+414-event dynamic exact replay. Its GPU6 feasibility is4430.811 frames/s versus
+4366.914, but it produces2.677% more frames, takes1.197% longer and regresses TTFA.
+Both complete quality groups pass; this is not a fixed-work throughput win or a
+reason to change the default.
+
+A controlled native interference probe passes all60 code/PCM/state comparisons.
+Four B64 steps take33.33ms alone, versus34.65–36.92ms with three real F8 codec jobs
+of batch1–8. Concurrent pair completion is still better than serial. Model-high
+stream priority has no material gain; queried graph flags confirm launch-stream
+priority was the tested variable. This establishes contention for those fixed
+inputs, not the recoverability of every service overlap interval.
+
+The existing FI GDN adapter was tested on the current mixed path. An initial
+numeric gate fails on a private padding sequence and remains recorded as FAIL.
+A separate FP32-reference and padding-isolation diagnostic identifies the live
+versus unused domain, but final prefill hidden RRMS is0.09118. The local timing
+gain is only1.776%, below the predeclared5% criterion, so no full service trial or
+promotion follows. Do not silently reinterpret the original failure as success.
+
+Short reserved attention contexts pass 1023 exact checks across 9 cases and a
+complete GPU7 service/quality pair. Combining this with compact GDN, mixed layout
+and empty-convolution guards gives a same-source GPU0 one-pair result of
+4353.909 → 4591.038 frames/s. Candidate TTFA p50/p95 is 140.774/268.097 ms,
+mean RTF 0.295911, WER 1.925814%, SIM 63.693694. Both 1088-output quality groups
+pass, and the candidate's +0.858% generated frames are retained. The observed
+single-run +5.446% is not a repeated speedup claim. It is only 1.0179× the frozen
+vLLM baseline and fails the 1.5× target.
+
+Codec minimum-launch repartitioning changes 5/80 PCM outputs; the original exact
+gate remains FAIL even after identifying private scratch versus real final state.
+A new MoE residual fusion also fails its first exact static check; no timing is
+claimed for that candidate. Codec resource limiting remains experimental.
+No all-concurrency acceptance or successful README example is claimed; the
+campaign remains INCONCLUSIVE. Complete source, commands, numeric failures and
+full service/quality records are [published at the frozen runtime revision](https://github.com/david6666666/drama-3-single-b300-inference/tree/ab2d9015d6b0115530e675f25bbf5f188b1d65b1/evidence/20261016-ceiling).
diff --git a/evidence/drama3-b300-scale150/results/status.json b/evidence/drama3-b300-scale150/results/status.json
index f632caa34..b20f9c2d2 100644
--- a/evidence/drama3-b300-scale150/results/status.json
+++ b/evidence/drama3-b300-scale150/results/status.json
@@ -1,10 +1,10 @@
 {
   "verdict": "INCONCLUSIVE",
   "acceptance_issue": "https://github.com/ThinkFlowLab/OmniRSI/issues/10",
-  "reason": "The dedicated-capture native async candidate passes fixed dynamic replay and complete single-group quality, but C64 throughput still falls below the baseline and target; the required all-concurrency repeated matrix is not complete.",
+  "reason": "The four-change native C64 candidate passes complete one-group quality and observes 4591.04 frames/s, only 1.0179x the frozen vLLM baseline and below 6765.41; the required all-concurrency repeated matrix is not complete.",
   "runtime_pull_request": "https://github.com/david6666666/drama-3-single-b300-inference/pull/1",
-  "runtime_evidence_commit": "dfdd598a9810cf6e7b8aa6a102f6ab9270ff9cc6",
-  "runtime_evidence": "https://github.com/david6666666/drama-3-single-b300-inference/tree/dfdd598a9810cf6e7b8aa6a102f6ab9270ff9cc6/evidence/20261016-mixed",
+  "runtime_evidence_commit": "ab2d9015d6b0115530e675f25bbf5f188b1d65b1",
+  "runtime_evidence": "https://github.com/david6666666/drama-3-single-b300-inference/tree/ab2d9015d6b0115530e675f25bbf5f188b1d65b1/evidence/20261016-ceiling",
   "baseline_quality_groups": 21,
   "baseline_scored_waveforms": 22848,
   "packed_layer_graph_ownership_checks": 5076,
@@ -43,7 +43,21 @@
   "codec_workspace_shared_service_quality_passes": false,
   "codec_workspace_private_service_quality_passes_single_group": true,
   "codec_workspace_service_quality_samples_per_arm": 1088,
-  "latest_c64_gpu0_feasibility_frames_s": 4360.916512808546,
-  "latest_c64_gpu0_feasibility_wer_percent": 1.5908900611236707,
-  "latest_c64_gpu0_feasibility_sim": 64.19134978818543
+  "latest_c64_gpu0_feasibility_frames_s": 4591.037744827167,
+  "latest_c64_gpu0_feasibility_wer_percent": 1.9258142845181279,
+  "latest_c64_gpu0_feasibility_sim": 63.693693577804986,
+  "latest_c64_gpu0_feasibility_ttfa_p50_ms": 140.7738695270382,
+  "latest_c64_gpu0_feasibility_ttfa_p95_ms": 268.0968882865193,
+  "latest_c64_gpu0_feasibility_rtf_mean": 0.29591053403594225,
+  "latest_c64_feasibility_source_sha": "a607dff26c3ca5423746f7eebec4f5f5a88cd4c0",
+  "latest_c64_feasibility_repetitions": 1,
+  "latest_c64_feasibility_all_outputs_scored": 1088,
+  "latest_c64_target_frames_s": 6765.407263214912,
+  "latest_c64_target_met": false,
+  "compact_gdn_exact_checks": 2099,
+  "short_padding_exact_checks": 1023,
+  "controlled_codec_contention_checks": 60,
+  "codec_minlaunch_original_exact_gate_passes": false,
+  "moe_residual_initial_exact_gate_passes": false,
+  "all_concurrency_target_status": "INCONCLUSIVE"
 }
diff --git a/src/omnirsi/data/repo_packs/vllm_omni/knowledge/cuda/audio/concurrency-scaling-state-and-overlap.json b/src/omnirsi/data/repo_packs/vllm_omni/knowledge/cuda/audio/concurrency-scaling-state-and-overlap.json
index d4cf563be..06303d3a2 100644
--- a/src/omnirsi/data/repo_packs/vllm_omni/knowledge/cuda/audio/concurrency-scaling-state-and-overlap.json
+++ b/src/omnirsi/data/repo_packs/vllm_omni/knowledge/cuda/audio/concurrency-scaling-state-and-overlap.json
@@ -10,8 +10,8 @@
   "summary": "Diagnose audio scaling with scheduler-step counts, admission/drain phases, mixed prefill/decode and codec overlap; isolate state, label and RNG ownership before qualifying same-source service throughput and complete audio quality.",
   "translations": {
     "zh-CN": {
-      "title": "音频并发扩展：混合预填充、状态归属和 codec 重叠",
-      "summary": "按调度轮而非图碎片计数，分开初始填充、持续接纳与排空阶段，分析混合预填充/解码和 codec 重叠；分别检查状态、标签及随机数归属，再以同源服务对照和完整音频质量验收。"
+      "title": "\u97f3\u9891\u5e76\u53d1\u6269\u5c55\uff1a\u6df7\u5408\u9884\u586b\u5145\u3001\u72b6\u6001\u5f52\u5c5e\u548c codec \u91cd\u53e0",
+      "summary": "\u6309\u8c03\u5ea6\u8f6e\u800c\u975e\u56fe\u788e\u7247\u8ba1\u6570\uff0c\u5206\u5f00\u521d\u59cb\u586b\u5145\u3001\u6301\u7eed\u63a5\u7eb3\u4e0e\u6392\u7a7a\u9636\u6bb5\uff0c\u5206\u6790\u6df7\u5408\u9884\u586b\u5145/\u89e3\u7801\u548c codec \u91cd\u53e0\uff1b\u5206\u522b\u68c0\u67e5\u72b6\u6001\u3001\u6807\u7b7e\u53ca\u968f\u673a\u6570\u5f52\u5c5e\uff0c\u518d\u4ee5\u540c\u6e90\u670d\u52a1\u5bf9\u7167\u548c\u5b8c\u6574\u97f3\u9891\u8d28\u91cf\u9a8c\u6536\u3002"
     }
   },
   "tags": [
@@ -33,7 +33,11 @@
     "rng_ownership",
     "hardware_counters",
     "static_sequence_bound",
-    "output_length_control"
+    "output_length_control",
+    "codec_batch_fragmentation",
+    "contention_control",
+    "reference_throughput",
+    "private_padding"
   ],
   "sources": [
     {
@@ -91,7 +95,13 @@
     "Use hardware counters on actual routed inputs to distinguish tensor arithmetic from expert-weight traffic. Report profiler replay policy, measured DRAM bytes and peak fraction, layer/input coverage and routing-count provenance; low tensor activity is not a directly recoverable speedup.",
     "Inspect static maximum sequence lengths separately from real packed token counts. Tighter graph bounds can remove inactive recurrent chunks and attention work without changing rows; validate every real and padding sequence fits, preserve a larger-bound fallback, and compare all state and outputs before service.",
     "For stochastic service A/B trials, retain generated-frame counts and request-length differences alongside frames/s and drain duration. Capture warmups and scheduling can change random draws, so a faster makespan alone does not prove a fixed-work kernel gain.",
-    "Separate replay streams and graph memory pools do not by themselves prove BLAS scratch isolation. Inspect the installed runtime capture/workspace ownership; where buffers are keyed by capture stream, test dedicated capture streams and record actual address ranges. Compare identical dynamic admissions, RNG, slot reuse and full mutable state before attributing a failure to workspace sharing."
+    "Separate replay streams and graph memory pools do not by themselves prove BLAS scratch isolation. Inspect the installed runtime capture/workspace ownership; where buffers are keyed by capture stream, test dedicated capture streams and record actual address ranges. Compare identical dynamic admissions, RNG, slot reuse and full mutable state before attributing a failure to workspace sharing.",
+    "An achieved isolated full-batch decode rate is a workload-specific reference, not a hardware upper bound or a service lower bound. Account for real prompt work, partial batches, codec interference and complete drain before judging the service target.",
+    "For asynchronous timelines, compute model/codec interval unions and their intersection before attributing elapsed time. Record actual decode steps, active rows, mixed token capacity and exact codec grouping rather than inferring GPU work from client concurrency.",
+    "Distinguish codec frame-bucket diversity from same-bucket partitioning. Log each ready group, chosen partition, actual request count and actual captured graph size; a planned oversized batch can be rounded differently by the graph dispatcher.",
+    "Compact packed recurrent chunk descriptors once per replay, preserve the original per-sequence recurrence ordering and arithmetic, and guard kernels that otherwise execute matrix work on invalid chunks. Recheck empty sequences, dynamic descriptor updates and larger-span fallback.",
+    "Reserved decode padding can use a shorter runtime attention context while preserving the largest capture plan, provided the graph explicitly identifies its private slots. Verify real outputs and all nonparticipating slots; do not infer a reserved range for arbitrary request-state callers.",
+    "Test stream contention on identical inputs and full restored state, comparing model-only, codec-only, serial and concurrent pair completion. A higher launch-stream priority is only a scheduling hint; query graph instantiation flags before interpreting priority trials."
   ],
   "validation_plan": [
     "L0: exercise partial batches, variable prompt boundaries, slot reuse, request cancellation and untouched-state checks before any service comparison.",
@@ -101,7 +111,9 @@
     "For stream overlap, hold request ordering, initial model/codec state and sampling counters fixed; compare generated codes, PCM, completion order and reused-slot state. Exact fixed-request replay does not replace quality under dynamic service admission.",
     "L1/L2: replay heterogeneous real prompts; inspect routing occupancy and graph kernels, and distinguish summed kernel time from elapsed time when streams overlap.",
     "L3/L4: score every output in the frozen corpus and repeated concurrency sweep; report acoustic frames/s, first playable audio latency, RTF, errors, WER and speaker similarity. Preserve rejected runs and source/environment provenance.",
-    "Promote a scheduling or kernel change only after a same-source end-to-end control/candidate comparison and complete repeated quality coverage meet the frozen gates. Operator checks, graph replay and a phase-selected trace cannot substitute for the full service result."
+    "Promote a scheduling or kernel change only after a same-source end-to-end control/candidate comparison and complete repeated quality coverage meet the frozen gates. Operator checks, graph replay and a phase-selected trace cannot substitute for the full service result.",
+    "Preserve an original bit-exact failure when investigating a changed batch shape or numerical backend. Identify real versus private scratch slots, quantify PCM and state errors, and use an independent higher-precision reference or causal isolation. A later diagnostic does not rewrite the failed gate and does not replace full WER/SIM.",
+    "When an isolated kernel improvement fails the full service throughput score, retain frame-count, length and drain differences and the failed score. Use fixed-input timing to test the operator hypothesis; do not rescue acceptance by discarding tails or substituting a selected-phase rate."
   ],
   "limitations": [
     "This record describes source-grounded mechanisms, not a reproduced speedup or a portable set of tuned constants.",
@@ -113,7 +125,9 @@
     "A whole-run average batch can be dominated by drain steps; a high batch in an admission subset does not establish an end-to-end speedup. Phase cuts are diagnostic scope, not permission to remove requests from acceptance metrics.",
     "A shared-execution prototype may change GEMM shapes, MoE implementation or random-number consumption. Keep those changes explicit, and keep timing from a device with observed external work unqualified even when numerical checks pass.",
     "A kernel reading nearly one active-expert weight footprint can remain bandwidth-limited despite low tensor activity; row padding or inactive experts must not be converted directly into a speedup claim.",
-    "A dedicated capture stream is a workspace-isolation hypothesis, not proof that prior quality failures were races. Require a fixed dynamic schedule and full service quality before attributing a cause or selecting overlap."
+    "A dedicated capture stream is a workspace-isolation hypothesis, not proof that prior quality failures were races. Require a fixed dynamic schedule and full service quality before attributing a cause or selecting overlap.",
+    "Configured or logged SM reservations may apply to a routing subkernel rather than expert GEMMs; inspect the actual dispatched branch before using such a setting as a resource-allocation explanation.",
+    "Native codec re-batching can change floating-point rounding. An error confined to a private scratch slot differs from a live-state error, but neither scratch analysis nor a high PCM SNR establishes full service quality."
   ],
   "verification": {
     "locally_verified": false,

From 39f9cc0b12e63a9468da7f8958d1c515ab4ba937 Mon Sep 17 00:00:00 2001
From: david6666666 <530634352@qq.com>
Date: Thu, 15 Oct 2026 23:41:07 +0000
Subject: [PATCH 2/6] Record native pipeline lifetime and full codec service
 lessons

Signed-off-by: david6666666 <530634352@qq.com>
---
 evidence/drama3-b300-scale150/patch/README.md |  2 +-
 .../results/diagnostic-lessons.md             | 43 +++++++++++++++++++
 .../drama3-b300-scale150/results/status.json  | 33 ++++++++++++--
 ...concurrency-scaling-state-and-overlap.json | 10 ++++-
 4 files changed, 82 insertions(+), 6 deletions(-)

diff --git a/evidence/drama3-b300-scale150/patch/README.md b/evidence/drama3-b300-scale150/patch/README.md
index 6cd5833a5..ccf83d075 100644
--- a/evidence/drama3-b300-scale150/patch/README.md
+++ b/evidence/drama3-b300-scale150/patch/README.md
@@ -5,5 +5,5 @@ No winning candidate is nominated. Each experiment records its source revision b
 The implementation and rejected trials are reviewable in
 [runtime draft PR #1](https://github.com/david6666666/drama-3-single-b300-inference/pull/1).
 Latest published evidence is pinned at
-[`ab2d901`](https://github.com/david6666666/drama-3-single-b300-inference/tree/ab2d9015d6b0115530e675f25bbf5f188b1d65b1/evidence/20261016-ceiling).
+[`171acc8`](https://github.com/david6666666/drama-3-single-b300-inference/tree/171acc82460f479b56372a57d9f7de7a154795aa/evidence/20261016-ceiling).
 This is an evidence publication revision, not a nominated performance winner.
diff --git a/evidence/drama3-b300-scale150/results/diagnostic-lessons.md b/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
index ae7eb2e13..ad7461508 100644
--- a/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
+++ b/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
@@ -237,3 +237,46 @@ claimed for that candidate. Codec resource limiting remains experimental.
 No all-concurrency acceptance or successful README example is claimed; the
 campaign remains INCONCLUSIVE. Complete source, commands, numeric failures and
 full service/quality records are [published at the frozen runtime revision](https://github.com/david6666666/drama-3-single-b300-inference/tree/ab2d9015d6b0115530e675f25bbf5f188b1d65b1/evidence/20261016-ceiling).
+
+
+## Native pipeline, rounding and resource-allocation follow-up
+
+The corrected MoE residual fusion preserves the reference compiler's native BF16
+FMA and passes2374 exact checks. The original separate-operation failure remains
+archived. Explicit reserved-slot routing passes2574 production checks while
+preserving every real full-top8 route and native activation/scales; its local
+benefit is limited to partially populated decode batches.
+
+The checkpoint embeds semantic/code1 for the next backbone. A native two-frame
+pipeline can therefore overlap its next backbone with eight remaining tower
+layers. Its production gate passes3689 exact checks, including slot reuse,
+max_frames and a separately injected first-EOS boundary. A callback-local buffer
+had been allocated on the main stream and consumed on an auxiliary stream after
+its reference expired. Retaining its lifetime is supported by a one-variable335
+exact-check diagnostic. An additional validation failure came from an eager call
+rebinding a Python tensor attribute; graph replay still updated the old captured
+storage. Both failed records remain, and the corrected validator pins addresses.
+These are model correctness gates; complete service evaluation is still separate.
+
+An actual emission ledger leaves47.58% of supported ordinary calls not already
+blocked by known audio boundaries. This is an upper bound, not actual pair use.
+Correlated chunk phases contradict an independent-row estimate. New capture
+warmups also advance the global sampling counter, so the service comparison must
+retain output lengths and cannot be labeled fixed-output merely because inputs
+match.
+
+Native codec resource caps require capture-time kernel assignment. NCU verifies
+all284 nodes of the inspected native B4/F8 graph. The full dynamic diagnostic
+covers46 captures and28 actual shapes, preserving414 trace entries,80 full PCM
+outputs and all model/codec state across normal/capped-sync/poll/thread modes.
+The complete service pair observes4412.554→4470.317 frames/s, with fewer output
+frames and arm-specific measured-cost DP feedback. Both1088-output quality groups
+pass (WER1.909068%/1.565771%, SIM63.902018/64.244919); the candidate remains below
+the baseline and target. Adjacent codec group coalescing likewise passes complete
+quality but regresses4513.299→4414.243 frames/s. Neither changes the campaign verdict.
+
+The best complete single-run quality-qualified observation remains4591.038frames/s,
+and the all-concurrency three-repeat requirement remains incomplete. All new raw
+results, preserved failures, source snapshots and limitations are
+[pinned in runtime evidence](https://github.com/david6666666/drama-3-single-b300-inference/tree/171acc82460f479b56372a57d9f7de7a154795aa/evidence/20261016-ceiling).
+No successful README example is added; the campaign remains INCONCLUSIVE.
diff --git a/evidence/drama3-b300-scale150/results/status.json b/evidence/drama3-b300-scale150/results/status.json
index b20f9c2d2..6cf566d89 100644
--- a/evidence/drama3-b300-scale150/results/status.json
+++ b/evidence/drama3-b300-scale150/results/status.json
@@ -3,8 +3,8 @@
   "acceptance_issue": "https://github.com/ThinkFlowLab/OmniRSI/issues/10",
   "reason": "The four-change native C64 candidate passes complete one-group quality and observes 4591.04 frames/s, only 1.0179x the frozen vLLM baseline and below 6765.41; the required all-concurrency repeated matrix is not complete.",
   "runtime_pull_request": "https://github.com/david6666666/drama-3-single-b300-inference/pull/1",
-  "runtime_evidence_commit": "ab2d9015d6b0115530e675f25bbf5f188b1d65b1",
-  "runtime_evidence": "https://github.com/david6666666/drama-3-single-b300-inference/tree/ab2d9015d6b0115530e675f25bbf5f188b1d65b1/evidence/20261016-ceiling",
+  "runtime_evidence_commit": "171acc82460f479b56372a57d9f7de7a154795aa",
+  "runtime_evidence": "https://github.com/david6666666/drama-3-single-b300-inference/tree/171acc82460f479b56372a57d9f7de7a154795aa/evidence/20261016-ceiling",
   "baseline_quality_groups": 21,
   "baseline_scored_waveforms": 22848,
   "packed_layer_graph_ownership_checks": 5076,
@@ -59,5 +59,32 @@
   "controlled_codec_contention_checks": 60,
   "codec_minlaunch_original_exact_gate_passes": false,
   "moe_residual_initial_exact_gate_passes": false,
-  "all_concurrency_target_status": "INCONCLUSIVE"
+  "all_concurrency_target_status": "INCONCLUSIVE",
+  "moe_residual_corrected_exact_checks": 2374,
+  "moe_residual_corrected_exact_gate_passes": true,
+  "reserved_padding_router_production_exact_checks": 2574,
+  "native_frame_pipeline_production_exact_checks": 3689,
+  "native_frame_pipeline_service_complete": false,
+  "recent_codec_service_trials": {
+    "adjacent_coalescing": {
+      "source_sha": "69d636ef11875add6e4edf7c9bb028b9574143a3",
+      "repeats_per_arm": 1,
+      "quality_samples_per_arm": 1088,
+      "control_frames_s": 4513.299150726729,
+      "candidate_frames_s": 4414.243311954199,
+      "both_complete_quality_pass": true,
+      "target_met": false,
+      "raw_summary": "coalescing-native-service-summary.json"
+    },
+    "native_48sm": {
+      "source_sha": "69d636ef11875add6e4edf7c9bb028b9574143a3",
+      "repeats_per_arm": 1,
+      "quality_samples_per_arm": 1088,
+      "control_frames_s": 4412.554410530426,
+      "candidate_frames_s": 4470.316688552834,
+      "both_complete_quality_pass": true,
+      "target_met": false,
+      "raw_summary": "green-sm-service-summary.json"
+    }
+  }
 }
diff --git a/src/omnirsi/data/repo_packs/vllm_omni/knowledge/cuda/audio/concurrency-scaling-state-and-overlap.json b/src/omnirsi/data/repo_packs/vllm_omni/knowledge/cuda/audio/concurrency-scaling-state-and-overlap.json
index 06303d3a2..3fc329b94 100644
--- a/src/omnirsi/data/repo_packs/vllm_omni/knowledge/cuda/audio/concurrency-scaling-state-and-overlap.json
+++ b/src/omnirsi/data/repo_packs/vllm_omni/knowledge/cuda/audio/concurrency-scaling-state-and-overlap.json
@@ -101,7 +101,11 @@
     "Distinguish codec frame-bucket diversity from same-bucket partitioning. Log each ready group, chosen partition, actual request count and actual captured graph size; a planned oversized batch can be rounded differently by the graph dispatcher.",
     "Compact packed recurrent chunk descriptors once per replay, preserve the original per-sequence recurrence ordering and arithmetic, and guard kernels that otherwise execute matrix work on invalid chunks. Recheck empty sequences, dynamic descriptor updates and larger-span fallback.",
     "Reserved decode padding can use a shorter runtime attention context while preserving the largest capture plan, provided the graph explicitly identifies its private slots. Verify real outputs and all nonparticipating slots; do not infer a reserved range for arbitrary request-state callers.",
-    "Test stream contention on identical inputs and full restored state, comparing model-only, codec-only, serial and concurrent pair completion. A higher launch-stream priority is only a scheduling hint; query graph instantiation flags before interpreting priority trials."
+    "Test stream contention on identical inputs and full restored state, comparing model-only, codec-only, serial and concurrent pair completion. A higher launch-stream priority is only a scheduling hint; query graph instantiation flags before interpreting priority trials.",
+    "For cross-stream model pipelines, establish both execution dependencies and allocation lifetimes. Retain a producer tensor until every consumer stream has finished; an event wait alone does not stop the allocator from reusing callback-local storage.",
+    "Derive inter-frame dependencies from the actual embedding and sampled-code paths before declaring a whole tower serial. Keep next-frame positions, attention workspaces, recurrent states and sampling counters independent until their required joins.",
+    "When changing codec resource allocation, verify actual captured kernel assignments and preserve the normal partition policy. If its measured cost table changes partitions, report that feedback explicitly rather than claiming a fixed-partition comparison.",
+    "Estimate temporal batching eligibility from observed chunk phases, admission and completion boundaries. Requests can have correlated emission phases; an independent per-row probability model can badly misestimate useful coverage."
   ],
   "validation_plan": [
     "L0: exercise partial batches, variable prompt boundaries, slot reuse, request cancellation and untouched-state checks before any service comparison.",
@@ -113,7 +117,9 @@
     "L3/L4: score every output in the frozen corpus and repeated concurrency sweep; report acoustic frames/s, first playable audio latency, RTF, errors, WER and speaker similarity. Preserve rejected runs and source/environment provenance.",
     "Promote a scheduling or kernel change only after a same-source end-to-end control/candidate comparison and complete repeated quality coverage meet the frozen gates. Operator checks, graph replay and a phase-selected trace cannot substitute for the full service result.",
     "Preserve an original bit-exact failure when investigating a changed batch shape or numerical backend. Identify real versus private scratch slots, quantify PCM and state errors, and use an independent higher-precision reference or causal isolation. A later diagnostic does not rewrite the failed gate and does not replace full WER/SIM.",
-    "When an isolated kernel improvement fails the full service throughput score, retain frame-count, length and drain differences and the failed score. Use fixed-input timing to test the operator hypothesis; do not rescue acceptance by discarding tails or substituting a selected-phase rate."
+    "When an isolated kernel improvement fails the full service throughput score, retain frame-count, length and drain differences and the failed score. Use fixed-input timing to test the operator hypothesis; do not rescue acceptance by discarding tails or substituting a selected-phase rate.",
+    "A capture-versus-eager validator must retain references to the captured buffers: replay updates captured storage, while eager Python execution can rebind object attributes. Preserve any failing run and prove a validator-handle mistake before replacing its result.",
+    "Check compiler FMA contraction and intermediate rounding when fusing native arithmetic. A source-level cast sequence need not imply separate machine operations. Keep original failures and distinguish an explicit correction from a separately declared higher-precision numerical diagnostic."
   ],
   "limitations": [
     "This record describes source-grounded mechanisms, not a reproduced speedup or a portable set of tuned constants.",

From 69761e278887260e967b70ad6b2f56555445a4f1 Mon Sep 17 00:00:00 2001
From: david6666666 <530634352@qq.com>
Date: Fri, 16 Oct 2026 00:35:00 +0000
Subject: [PATCH 3/6] Record complete native C64 trials and distinguish model
 cost from codec gains

Signed-off-by: david6666666 <530634352@qq.com>
---
 .../results/diagnostic-lessons.md             |  53 +++++++
 .../drama3-b300-scale150/results/status.json  | 137 +++++++++++++++++-
 ...concurrency-scaling-state-and-overlap.json |  12 +-
 3 files changed, 195 insertions(+), 7 deletions(-)

diff --git a/evidence/drama3-b300-scale150/results/diagnostic-lessons.md b/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
index ad7461508..cba96b4ac 100644
--- a/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
+++ b/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
@@ -280,3 +280,56 @@ and the all-concurrency three-repeat requirement remains incomplete. All new raw
 results, preserved failures, source snapshots and limitations are
 [pinned in runtime evidence](https://github.com/david6666666/drama-3-single-b300-inference/tree/171acc82460f479b56372a57d9f7de7a154795aa/evidence/20261016-ceiling).
 No successful README example is added; the campaign remains INCONCLUSIVE.
+
+
+## Native structure follow-up: correctness and performance remain separate
+
+Issue#10 remains INCONCLUSIVE. Distinguishing queued admission from runnable
+admission raises two-frame eligibility from139/1902 to625/1246 lifetime ordinary
+launch decisions, including warmup. The complete1088-request pair observes
+4481.280→4583.673frames/s with both quality groups passing, but p50/p95 first
+audio rises. One feasibility pair is not a repeated gain or final acceptance.
+
+Private native-BF16 Linear parameter caching removes40 repeat autocast kernels
+from the codec. Original FP32 codec parameters, norm/layer-scale arithmetic and
+reference encoding remain unchanged. Local273 exact checks,46 captured shapes,
+and fixed80-request/414-entry dynamic replay pass. Yet full service observes
+4726.771→4585.126frames/s with all1088 outputs scored in each arm and quality
+passing. Output lengths and delivery tails differ; do not attribute the entire
+service delta to codec compute. The cache remains default off. The control is
+the best complete-quality single observation, not a claimed repeatable winner.
+
+Preserving native GDN warp/reduction order restores exact real-layer and full
+state results, while fused recurrence/norm is slower. Resource accounting must
+use the production specialization (182 registers/thread), not a debug output
+variant (177). A CPU lifetime change raises allocation to185 and is stopped.
+Shared MoE metadata removes duplicate routing work but the complete partition
+path still loses to native. Neither fewer launches nor exact outputs imply a
+performance gain. New resident-expert proposals remain unverified CPU designs.
+
+Autotuning comparisons must control algorithm-plan caches, not merely switch a
+flag. The first native cuDNN comparison reused a thread-local cache populated by
+control; its benchmark flag was absent from the key. Fresh host threads expose
+a PCM exact mismatch before timing. Preserve the original result and explicit
+correction, and keep a failed exact check distinct from unmeasured task quality.
+No user quality bound, precision boundary, expert count or acceptance gate is
+changed. Detailed artifacts are pinned by status.json's runtime evidence commit.
+
+The cache control follows the installed [pinned PyTorch implementation](https://github.com/pytorch/pytorch/blob/cf30153c4c131c8164ee7798e5022d810682e2cb/aten/src/ATen/native/cudnn/Conv_v8.cpp); this dependency source is kept in issue-scoped evidence rather than the vLLM-only imported source list.
+
+
+A fixed-work replay now separates causal codec contention from variable-length
+service scores. All4,177 exact checks pass on80 requests/414 trace entries;
+normal overlap lengthens model spans yet remains faster than serialization.
+Native weight caching reduces isolated codec time but only0.402% of combined
+wall time in this fixture. Its packed-prefill/single-step scope excludes the
+current mixed/frame-pipeline executor, whose separate event ledger is retained.
+Never present this replay as a complete service benchmark.
+
+Prefix reuse opportunities must be evaluated against arrival groups. Adjacent
+identical references can be simultaneous cold misses; completed-prefix hits and
+same-group deduplication require different state ownership. Cache identity must
+include all acoustic code rows. A valid suffix executor needs prefix K/V,
+convolution history and recurrent state, correct teacher labels and independent
+writable branch state. Count extra model traversals and precision boundaries
+before translating duplicate-row counts into a latency forecast.
diff --git a/evidence/drama3-b300-scale150/results/status.json b/evidence/drama3-b300-scale150/results/status.json
index 6cf566d89..ccb64f08a 100644
--- a/evidence/drama3-b300-scale150/results/status.json
+++ b/evidence/drama3-b300-scale150/results/status.json
@@ -1,10 +1,10 @@
 {
   "verdict": "INCONCLUSIVE",
   "acceptance_issue": "https://github.com/ThinkFlowLab/OmniRSI/issues/10",
-  "reason": "The four-change native C64 candidate passes complete one-group quality and observes 4591.04 frames/s, only 1.0179x the frozen vLLM baseline and below 6765.41; the required all-concurrency repeated matrix is not complete.",
+  "reason": "Best complete-quality C64 observation is4726.77 frames/s (one control arm, not a repeatable improvement claim), below6765.41. Native cache and partition candidates fail local/service performance promotion; the required all-concurrency repeated matrix remains incomplete.",
   "runtime_pull_request": "https://github.com/david6666666/drama-3-single-b300-inference/pull/1",
-  "runtime_evidence_commit": "171acc82460f479b56372a57d9f7de7a154795aa",
-  "runtime_evidence": "https://github.com/david6666666/drama-3-single-b300-inference/tree/171acc82460f479b56372a57d9f7de7a154795aa/evidence/20261016-ceiling",
+  "runtime_evidence_commit": "245629abae234866de4cbb589bff6587d0a5c85a",
+  "runtime_evidence": "https://github.com/david6666666/drama-3-single-b300-inference/tree/245629abae234866de4cbb589bff6587d0a5c85a/evidence/20261016-ceiling",
   "baseline_quality_groups": 21,
   "baseline_scored_waveforms": 22848,
   "packed_layer_graph_ownership_checks": 5076,
@@ -64,7 +64,7 @@
   "moe_residual_corrected_exact_gate_passes": true,
   "reserved_padding_router_production_exact_checks": 2574,
   "native_frame_pipeline_production_exact_checks": 3689,
-  "native_frame_pipeline_service_complete": false,
+  "native_frame_pipeline_service_complete": true,
   "recent_codec_service_trials": {
     "adjacent_coalescing": {
       "source_sha": "69d636ef11875add6e4edf7c9bb028b9574143a3",
@@ -86,5 +86,134 @@
       "target_met": false,
       "raw_summary": "green-sm-service-summary.json"
     }
+  },
+  "native_frame_pipeline_service_trial": {
+    "source_sha": "a7118fb7eb748d08c3b774fa6c9189b6eab0d9cc",
+    "repeats_per_arm": 1,
+    "quality_samples_per_arm": 1088,
+    "control_frames_s": 4576.835318240403,
+    "candidate_frames_s": 4580.029569444877,
+    "candidate_wer_percent": 2.277484719082308,
+    "candidate_sim": 63.90760645840098,
+    "both_complete_quality_pass": true,
+    "pairs_selected": 159,
+    "ordinary_launch_decisions": 1710,
+    "decision_scope": "64 warmup plus1088 measured requests; combined waiting-or-pending blocks cannot identify recoverable waiting-only cases",
+    "target_met": false,
+    "raw_summary": "pipeline-service-summary.json"
+  },
+  "gdn_fused_norm_expanded_numeric_gate_pass": false,
+  "gdn_fused_norm_timing_qualified": false,
+  "expert_partition_native_checks": 135,
+  "expert_partition_local_performance_gate_pass": false,
+  "native_deferred_waiting_trial": {
+    "source_sha": "95fcb0ef2662f2caf9bfb7fe6863dc3ad7805b1b",
+    "control_frames_s": 4481.280462849354,
+    "candidate_frames_s": 4583.673173280921,
+    "candidate_ttfa_p50_ms": 146.27732400549576,
+    "candidate_ttfa_p95_ms": 274.84844140708003,
+    "candidate_rtf_mean": 0.2895720557859185,
+    "candidate_wer_percent": 1.833710123084652,
+    "candidate_sim": 64.12095218902344,
+    "repeats_per_arm": 1,
+    "quality_samples_per_arm": 1088,
+    "both_complete_quality_pass": true,
+    "target_met": false,
+    "raw_summary": "pipeline-admission-service-summary.json"
+  },
+  "native_codec_cached_linear": {
+    "source_sha": "e7cda1eba083603b88e1f2e89688e1ff2ced7fc2",
+    "local_exact_checks": 273,
+    "full_captured_shapes_exact": 46,
+    "dynamic_requests": 80,
+    "dynamic_trace_entries": 414,
+    "all_dynamic_modes_exact": true,
+    "original_fp32_codec_preserved": true,
+    "service_complete": true,
+    "target_met": false,
+    "raw_gate": "codec-cache-dynamic/result.json",
+    "quality_samples_per_arm": 1088,
+    "both_complete_quality_pass": true,
+    "control_frames_s": 4726.77149517109,
+    "candidate_frames_s": 4585.126355696841,
+    "candidate_wer_percent": 2.0179184459516035,
+    "candidate_sim": 63.86089451267617,
+    "repeats_per_arm": 1,
+    "performance_promoted": false,
+    "raw_summary": "codec-cache-service-summary.json"
+  },
+  "best_observed_complete_quality_c64": {
+    "source_sha": "e7cda1eba083603b88e1f2e89688e1ff2ced7fc2",
+    "configuration": "codec-cache-service control: cached-linear off; native frame pipeline and deferred waiting on",
+    "gpu": "GPU6",
+    "frames_s": 4726.77149517109,
+    "ttfa_p50_ms": 145.6451045232825,
+    "ttfa_p95_ms": 273.6818287463342,
+    "rtf_mean": 0.28813005589917623,
+    "wer_percent": 1.5992631667085322,
+    "sim": 64.41275850157527,
+    "repetitions": 1,
+    "all_outputs_scored": 1088,
+    "is_formal_acceptance": false,
+    "target_met": false,
+    "raw_summary": "codec-cache-service-summary.json"
+  },
+  "gdn_native_warp_tiles": {
+    "layer_cases_exact": 180,
+    "full_graph_cases_exact": 6,
+    "steps_per_graph_case": 8,
+    "local_time_relative_increase": 0.06369973429084452,
+    "production_registers_per_thread": 182,
+    "late_layout_registers_per_thread": 185,
+    "local_performance_pass": false,
+    "runtime_promoted": false,
+    "raw_report": "gdn-fused-norm/warp-tiles-result-report.md"
+  },
+  "expert_metadata_rebase": {
+    "exact_checks": 1373,
+    "local_performance_pass": false,
+    "runtime_promoted": false,
+    "raw_report": "expert-metadata-rebase/summary.md"
+  },
+  "codec_fresh_cudnn_selection": {
+    "original_same_thread_cache_confounded": true,
+    "fresh_thread_exact_pass": false,
+    "timing_run": false,
+    "runtime_promoted": false,
+    "raw_report": "codec-cudnn-fresh-thread/summary.md"
+  },
+  "current_pipeline_timeline": {
+    "source_sha": "e7cda1eba083603b88e1f2e89688e1ff2ced7fc2",
+    "requests": 1088,
+    "is_acceptance": false,
+    "model_union_s": 21.289383056640624,
+    "codec_only_s": 0.970009521484375,
+    "neither_model_codec_s": 0.613250732421875,
+    "raw_report": "current-pipeline-timeline-results.md"
+  },
+  "fixed_native_work_contention": {
+    "source_sha": "e7cda1eba083603b88e1f2e89688e1ff2ced7fc2",
+    "requests": 80,
+    "trace_entries": 414,
+    "exact_checks": 4177,
+    "all_modes_exact": true,
+    "includes_current_mixed_pipeline": false,
+    "is_service_acceptance": false,
+    "raw_report": "cpu-service-ceiling/fixed-work/results.md"
+  },
+  "reference_prefix_feasibility": {
+    "cpu_only": true,
+    "runtime_implemented": false,
+    "cold_reuse_prompt_row_fraction": 0.34421045803919875,
+    "admission_groups_are_fifo_proxy": true,
+    "raw_report": "reference-prefix/feasibility.md"
+  },
+  "resident_fc1_prototype": {
+    "original_exact_gate_pass": false,
+    "cpu_layout_corrected_bytes": 8192,
+    "original_scales_differ": 35,
+    "timing_run": false,
+    "runtime_promoted": false,
+    "raw_report": "expert-fc1-native/layout-audit.json"
   }
 }
diff --git a/src/omnirsi/data/repo_packs/vllm_omni/knowledge/cuda/audio/concurrency-scaling-state-and-overlap.json b/src/omnirsi/data/repo_packs/vllm_omni/knowledge/cuda/audio/concurrency-scaling-state-and-overlap.json
index 3fc329b94..889ba80f5 100644
--- a/src/omnirsi/data/repo_packs/vllm_omni/knowledge/cuda/audio/concurrency-scaling-state-and-overlap.json
+++ b/src/omnirsi/data/repo_packs/vllm_omni/knowledge/cuda/audio/concurrency-scaling-state-and-overlap.json
@@ -105,7 +105,11 @@
     "For cross-stream model pipelines, establish both execution dependencies and allocation lifetimes. Retain a producer tensor until every consumer stream has finished; an event wait alone does not stop the allocator from reusing callback-local storage.",
     "Derive inter-frame dependencies from the actual embedding and sampled-code paths before declaring a whole tower serial. Keep next-frame positions, attention workspaces, recurrent states and sampling counters independent until their required joins.",
     "When changing codec resource allocation, verify actual captured kernel assignments and preserve the normal partition policy. If its measured cost table changes partitions, report that feedback explicitly rather than claiming a fixed-partition comparison.",
-    "Estimate temporal batching eligibility from observed chunk phases, admission and completion boundaries. Requests can have correlated emission phases; an independent per-row probability model can badly misestimate useful coverage."
+    "Estimate temporal batching eligibility from observed chunk phases, admission and completion boundaries. Requests can have correlated emission phases; an independent per-row probability model can badly misestimate useful coverage.",
+    "Separate queued work from runnable admission when deciding whether to launch a bounded multi-step graph. Reuse the actual capacity, minimum-batch, minimum-running and oldest-age predicate; preserve pending-arrival and audio/completion boundaries. Record mutually exclusive decisions separately from overlapping context counters, and measure any extra admission delay.",
+    "Inspect compiled codec graphs for repeated casts of immutable Linear weights. Caching the existing autocast result can remove traffic without changing GEMM precision, but retain original reference/encoding modules and native norm/layer-scale precision. A compiler option alone is not proof that casts disappeared; inspect dispatched kernels and validate full codec state.",
+    "Inspect production-specialization register/shared-memory allocation when fusing recurrent operations. A debug output can change allocation, and removing launches can lose throughput through reduced residency or added synchronization; compile and measure the actual deployed specialization.",
+    "Control native algorithm-plan caches when comparing autotuning modes. Verify the installed cache key and lookup order: a flag change after control warmup may reuse the same plan. Isolate per-thread caches using fresh host threads where the implementation supports that scope, retain compilation/preparation costs, and do not clear shared system caches."
   ],
   "validation_plan": [
     "L0: exercise partial batches, variable prompt boundaries, slot reuse, request cancellation and untouched-state checks before any service comparison.",
@@ -119,7 +123,8 @@
     "Preserve an original bit-exact failure when investigating a changed batch shape or numerical backend. Identify real versus private scratch slots, quantify PCM and state errors, and use an independent higher-precision reference or causal isolation. A later diagnostic does not rewrite the failed gate and does not replace full WER/SIM.",
     "When an isolated kernel improvement fails the full service throughput score, retain frame-count, length and drain differences and the failed score. Use fixed-input timing to test the operator hypothesis; do not rescue acceptance by discarding tails or substituting a selected-phase rate.",
     "A capture-versus-eager validator must retain references to the captured buffers: replay updates captured storage, while eager Python execution can rebind object attributes. Preserve any failing run and prove a validator-handle mistake before replacing its result.",
-    "Check compiler FMA contraction and intermediate rounding when fusing native arithmetic. A source-level cast sequence need not imply separate machine operations. Keep original failures and distinguish an explicit correction from a separately declared higher-precision numerical diagnostic."
+    "Check compiler FMA contraction and intermediate rounding when fusing native arithmetic. A source-level cast sequence need not imply separate machine operations. Keep original failures and distinguish an explicit correction from a separately declared higher-precision numerical diagnostic.",
+    "Expand numerical checks across independently prepared real fixtures, layers and partial batches before timing. A single fixture can pass while another exposes small-magnitude per-head errors or a worse independent-reference error; retain both outcomes and preserve the predeclared gate."
   ],
   "limitations": [
     "This record describes source-grounded mechanisms, not a reproduced speedup or a portable set of tuned constants.",
@@ -133,7 +138,8 @@
     "A kernel reading nearly one active-expert weight footprint can remain bandwidth-limited despite low tensor activity; row padding or inactive experts must not be converted directly into a speedup claim.",
     "A dedicated capture stream is a workspace-isolation hypothesis, not proof that prior quality failures were races. Require a fixed dynamic schedule and full service quality before attributing a cause or selecting overlap.",
     "Configured or logged SM reservations may apply to a routing subkernel rather than expert GEMMs; inspect the actual dispatched branch before using such a setting as a resource-allocation explanation.",
-    "Native codec re-batching can change floating-point rounding. An error confined to a private scratch slot differs from a live-state error, but neither scratch analysis nor a high PCM SNR establishes full service quality."
+    "Native codec re-batching can change floating-point rounding. An error confined to a private scratch slot differs from a live-state error, but neither scratch analysis nor a high PCM SNR establishes full service quality.",
+    "A selected-phase delivery rate can explain a changing service score but is not GPU effective batch or a replacement acceptance metric. High client occupancy only rules out client starvation within its measured scope."
   ],
   "verification": {
     "locally_verified": false,

From e0c5026bf80d22e8544814cdcb89a84fec0190fe Mon Sep 17 00:00:00 2001
From: david6666666 <530634352@qq.com>
Date: Fri, 16 Oct 2026 01:00:02 +0000
Subject: [PATCH 4/6] Document native arithmetic gates and unsuccessful GDN H/O
 fusion

Signed-off-by: david6666666 <530634352@qq.com>
---
 .../results/diagnostic-lessons.md             | 38 +++++++++++++
 .../drama3-b300-scale150/results/status.json  | 53 ++++++++++++++++++-
 2 files changed, 89 insertions(+), 2 deletions(-)

diff --git a/evidence/drama3-b300-scale150/results/diagnostic-lessons.md b/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
index cba96b4ac..ec35e3799 100644
--- a/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
+++ b/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
@@ -333,3 +333,41 @@ include all acoustic code rows. A valid suffix executor needs prefix K/V,
 convolution history and recurrent state, correct teacher labels and independent
 writable branch state. Count extra model traversals and precision boundaries
 before translating duplicate-row counts into a latency forecast.
+
+
+## Native arithmetic compatibility is an explicit implementation boundary
+
+The resident-MoE arithmetic foundation now has expanded component evidence.
+FC1's first discrepancy contains a physical16-element store permutation and a
+quantization-scale formula mismatch. Correcting those passes a single expert,
+but a subsequent zero-input expansion still fails. Native SASS lacks the
+candidate's tiny-amplitude clamp. A separate correction passes438 checks over
+15 native layer/routing fixtures plus original-native zero and near-threshold
+references. FC2's corrected physical BF16 store passes306 expanded checks, and
+a shared-memory activation body passes2063. Every original failure is retained.
+Do not substitute a plausible library reimplementation for the actual frozen
+native arithmetic reference; test layouts, scales, signed zero, partial routes,
+changed graph metadata and retained storage before timing.
+
+Resource capacity must include dynamic register allocation and actual TMEM,
+not only a cubin's static register header or shared bytes. The current FC1 roles
+request53248 registers per CTA, so its small shared-memory working set does not
+establish two-CTA residency. Shared intermediates require producer publication,
+consumer-completion synchronization and explicit lifetime reuse. Individually
+exact stages do not establish a connected cluster or end-to-end performance.
+
+An independent H/O GDN fusion keeps original BF16 intermediate rounding and
+passes90 real layer cases plus complete graph/state checks per variant. Yet
+three repetitions show no useful complete-MixedGraph reduction. The first V32
+variant doubles QK calculation relative to original O V64. Preserving V64
+removes that duplication but raises shared/TMEM allocation and still regresses
+the primary shape slightly. Eliminated intermediate bytes are not guaranteed
+DRAM savings or speed: live values, resident blocks and chunk parallelism also
+change. Both candidates stay out of production. No more tile sweep follows
+these negative results, and none changes service acceptance or quality bounds.
+
+The component evidence is
+[pinned with all raw gates and repetitions](https://github.com/david6666666/drama-3-single-b300-inference/tree/f1dc31d57e4d29699a5bfbe5b0fc29ed5e163333/evidence/20261016-ceiling).
+The complete resident implementation remains in progress. C64 still needs
+6765.407frames/s, all-concurrency repeated acceptance remains incomplete, and
+no successful README example is claimed.
diff --git a/evidence/drama3-b300-scale150/results/status.json b/evidence/drama3-b300-scale150/results/status.json
index ccb64f08a..607fb71c0 100644
--- a/evidence/drama3-b300-scale150/results/status.json
+++ b/evidence/drama3-b300-scale150/results/status.json
@@ -3,8 +3,8 @@
   "acceptance_issue": "https://github.com/ThinkFlowLab/OmniRSI/issues/10",
   "reason": "Best complete-quality C64 observation is4726.77 frames/s (one control arm, not a repeatable improvement claim), below6765.41. Native cache and partition candidates fail local/service performance promotion; the required all-concurrency repeated matrix remains incomplete.",
   "runtime_pull_request": "https://github.com/david6666666/drama-3-single-b300-inference/pull/1",
-  "runtime_evidence_commit": "245629abae234866de4cbb589bff6587d0a5c85a",
-  "runtime_evidence": "https://github.com/david6666666/drama-3-single-b300-inference/tree/245629abae234866de4cbb589bff6587d0a5c85a/evidence/20261016-ceiling",
+  "runtime_evidence_commit": "f1dc31d57e4d29699a5bfbe5b0fc29ed5e163333",
+  "runtime_evidence": "https://github.com/david6666666/drama-3-single-b300-inference/tree/f1dc31d57e4d29699a5bfbe5b0fc29ed5e163333/evidence/20261016-ceiling",
   "baseline_quality_groups": 21,
   "baseline_scored_waveforms": 22848,
   "packed_layer_graph_ownership_checks": 5076,
@@ -215,5 +215,54 @@
     "timing_run": false,
     "runtime_promoted": false,
     "raw_report": "expert-fc1-native/layout-audit.json"
+  },
+  "resident_native_components": {
+    "source_sha": "95fcb0ef2662f2caf9bfb7fe6863dc3ad7805b1b",
+    "fc1_v2_single_expert_exact_checks": 42,
+    "fc1_zero_expansion_original_gate_pass": false,
+    "fc1_v3_expanded_exact_checks": 438,
+    "fc1_v3_native_oracle_cases": 15,
+    "fc1_v3_synthetic_zero_and_near_threshold_cases": 3,
+    "fc2_original_gate_pass": false,
+    "fc2_v2_single_expert_exact_checks": 40,
+    "fc2_expanded_exact_checks": 306,
+    "fc2_native_oracle_cases": 15,
+    "shared_activation_exact_checks": 2063,
+    "shared_activation_cases": 137,
+    "fc1_dynamic_register_budget_per_cta": 53248,
+    "full_resident_chain_exact_pass": false,
+    "timing_run": false,
+    "runtime_promoted": false,
+    "raw_report": "expert-resident-implementation-plan.md",
+    "preserved_failures": [
+      "expert-fc1-native",
+      "expert-fc1-expanded",
+      "expert-fc2-native"
+    ]
+  },
+  "gdn_prefill_ho_fusion": {
+    "source_sha": "e7cda1eba083603b88e1f2e89688e1ff2ced7fc2",
+    "variants": [
+      "V32",
+      "V64"
+    ],
+    "synthetic_exact_checks_per_variant": 57,
+    "real_exact_checks_per_variant": 522,
+    "timed_trajectory_exact_checks_per_variant": 1026,
+    "actual_gdn_layer_cases_per_variant": 90,
+    "complete_mixed_graph_shapes": 3,
+    "timing_repetitions_per_shape": 3,
+    "replays_per_timing_measurement": 20,
+    "both_local_continuation_gates_pass": false,
+    "v32_d58p6_native_ms": 26.35557861328125,
+    "v32_d58p6_candidate_ms": 26.49036560058594,
+    "v64_d58p6_native_ms": 26.355218505859376,
+    "v64_d58p6_candidate_ms": 26.420693969726564,
+    "runtime_promoted": false,
+    "is_service_acceptance": false,
+    "raw_reports": [
+      "gdn-fused-prefill/report.md",
+      "gdn-fused-prefill-o64/report.md"
+    ]
   }
 }

From d003e2aa2e83b9da480d313076afd0db333ad85c Mon Sep 17 00:00:00 2001
From: david6666666 <530634352@qq.com>
Date: Fri, 16 Oct 2026 01:30:33 +0000
Subject: [PATCH 5/6] Record resident-chain validation and M128 native resource
 evidence

Signed-off-by: david6666666 <530634352@qq.com>
---
 evidence/drama3-b300-scale150/patch/README.md |  2 +-
 .../results/diagnostic-lessons.md             | 28 +++++++++
 .../drama3-b300-scale150/results/status.json  | 58 +++++++++++++++++--
 3 files changed, 83 insertions(+), 5 deletions(-)

diff --git a/evidence/drama3-b300-scale150/patch/README.md b/evidence/drama3-b300-scale150/patch/README.md
index ccf83d075..62ef71059 100644
--- a/evidence/drama3-b300-scale150/patch/README.md
+++ b/evidence/drama3-b300-scale150/patch/README.md
@@ -5,5 +5,5 @@ No winning candidate is nominated. Each experiment records its source revision b
 The implementation and rejected trials are reviewable in
 [runtime draft PR #1](https://github.com/david6666666/drama-3-single-b300-inference/pull/1).
 Latest published evidence is pinned at
-[`171acc8`](https://github.com/david6666666/drama-3-single-b300-inference/tree/171acc82460f479b56372a57d9f7de7a154795aa/evidence/20261016-ceiling).
+[`02aa919`](https://github.com/david6666666/drama-3-single-b300-inference/tree/02aa9198d8c1d0e0ef5ba28b1cc64d2c0776fb8b/evidence/20261016-ceiling).
 This is an evidence publication revision, not a nominated performance winner.
diff --git a/evidence/drama3-b300-scale150/results/diagnostic-lessons.md b/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
index ec35e3799..4dbe00b03 100644
--- a/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
+++ b/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
@@ -371,3 +371,31 @@ The component evidence is
 The complete resident implementation remains in progress. C64 still needs
 6765.407frames/s, all-concurrency repeated acceptance remains incomplete, and
 no successful README example is claimed.
+
+
+## Resident composition and M128 layout/resource evidence
+
+The selected-expert resident FC1/activation/FC2 chain now passes489 exact checks
+for valid rows0–8, including graph replay, five original numeric boundaries and
+ownership. The native finalizer check substitutes only tested branches into an
+otherwise native oracle and is explicitly oracle-assisted. Full256 routing and
+performance are not implied. Actual complete launch uses79,488bytes perCTA and
+one outer-owned TMEM allocation, not a sum of individual component estimates.
+The initial host invocation passed compile-time constants to the runtime ABI;
+its failure occurred before candidate kernel execution and remains archived.
+
+An independently validated CuTe activation translation passes2066 checks; the
+previous C++ helper's2063 checks were not inherited. A separate M128 FC1 attempt
+fails its exact gate because four TMEM slices were stored as adjacent channels.
+A fixed32-channel permutation explains the observed bytes but does not pass the
+checker. The corresponding corrected M128 FC2 store preserves original K32/K128
+arithmetic and passes306 expanded checks across15native oracles/all256experts.
+Four160-register epilogue warps plus eight48-register roles budget32,768registers
+perCTA, versus53,248 for the M64 organization. That offers a two-CTA capacity
+hypothesis, not measured occupancy or service throughput.
+
+The GDN H/O negative experiments were independently recomputed. Their pre/post
+GPU process checks do not prove uninterrupted exclusivity; reports now state
+that limitation. Both original local continuation gates still fail. Every new
+component result remains below full-service acceptance scope, and C64's6765.407
+frames/s target and confirmed WER/SIM requirements are unchanged.
diff --git a/evidence/drama3-b300-scale150/results/status.json b/evidence/drama3-b300-scale150/results/status.json
index 607fb71c0..bcdc9099b 100644
--- a/evidence/drama3-b300-scale150/results/status.json
+++ b/evidence/drama3-b300-scale150/results/status.json
@@ -3,8 +3,8 @@
   "acceptance_issue": "https://github.com/ThinkFlowLab/OmniRSI/issues/10",
   "reason": "Best complete-quality C64 observation is4726.77 frames/s (one control arm, not a repeatable improvement claim), below6765.41. Native cache and partition candidates fail local/service performance promotion; the required all-concurrency repeated matrix remains incomplete.",
   "runtime_pull_request": "https://github.com/david6666666/drama-3-single-b300-inference/pull/1",
-  "runtime_evidence_commit": "f1dc31d57e4d29699a5bfbe5b0fc29ed5e163333",
-  "runtime_evidence": "https://github.com/david6666666/drama-3-single-b300-inference/tree/f1dc31d57e4d29699a5bfbe5b0fc29ed5e163333/evidence/20261016-ceiling",
+  "runtime_evidence_commit": "02aa9198d8c1d0e0ef5ba28b1cc64d2c0776fb8b",
+  "runtime_evidence": "https://github.com/david6666666/drama-3-single-b300-inference/tree/02aa9198d8c1d0e0ef5ba28b1cc64d2c0776fb8b/evidence/20261016-ceiling",
   "baseline_quality_groups": 21,
   "baseline_scored_waveforms": 22848,
   "packed_layer_graph_ownership_checks": 5076,
@@ -238,7 +238,24 @@
       "expert-fc1-native",
       "expert-fc1-expanded",
       "expert-fc2-native"
-    ]
+    ],
+    "shared_activation_dsl_exact_checks": 2066,
+    "shared_activation_dsl_cases": 137,
+    "resident_single_expert_exact_checks": 489,
+    "resident_single_expert_valid_rows": [
+      0,
+      1,
+      2,
+      3,
+      4,
+      5,
+      6,
+      7,
+      8
+    ],
+    "resident_single_expert_total_shared_bytes": 79488,
+    "resident_full_256_experts_gate_pass": false,
+    "original_host_abi_attempt_failed_before_kernel": true
   },
   "gdn_prefill_ho_fusion": {
     "source_sha": "e7cda1eba083603b88e1f2e89688e1ff2ced7fc2",
@@ -263,6 +280,39 @@
     "raw_reports": [
       "gdn-fused-prefill/report.md",
       "gdn-fused-prefill-o64/report.md"
-    ]
+    ],
+    "independent_cpu_review_pass_with_evidence_limits": true,
+    "process_ownership_evidence": "Pre/post process checks only; no continuous audit proves full-interval exclusivity",
+    "independent_review": "gdn-fused-prefill/independent-review.md"
+  },
+  "m128_native_resource_experiment": {
+    "source_sha": "95fcb0ef2662f2caf9bfb7fe6863dc3ad7805b1b",
+    "original_fc1_layout_gate_pass": false,
+    "fc2_single_expert_exact_checks": 39,
+    "fc2_expanded_exact_checks": 306,
+    "fc2_oracles": 15,
+    "layers": [
+      0,
+      19,
+      39
+    ],
+    "experts": 256,
+    "mma": [
+      128,
+      8,
+      32
+    ],
+    "original_mma": [
+      64,
+      8,
+      32
+    ],
+    "K_scale_block": 128,
+    "actual_threads_per_cta": 384,
+    "declared_register_budget_per_cta": 32768,
+    "observed_occupancy_claim": false,
+    "runtime_promoted": false,
+    "timing_run": false,
+    "raw_report": "expert-m128-fc2-expanded/report.md"
   }
 }

From b247efd61d1f8fd709f5abb535c37e55fda96367 Mon Sep 17 00:00:00 2001
From: david6666666 <530634352@qq.com>
Date: Fri, 16 Oct 2026 01:54:40 +0000
Subject: [PATCH 6/6] Record complete native MoE outcomes and correct occupancy
 inference

Signed-off-by: david6666666 <530634352@qq.com>
---
 evidence/drama3-b300-scale150/patch/README.md |  2 +-
 .../results/diagnostic-lessons.md             | 59 +++++++++++--
 .../drama3-b300-scale150/results/status.json  | 88 ++++++++++++++++---
 3 files changed, 128 insertions(+), 21 deletions(-)

diff --git a/evidence/drama3-b300-scale150/patch/README.md b/evidence/drama3-b300-scale150/patch/README.md
index 62ef71059..4f3507e8e 100644
--- a/evidence/drama3-b300-scale150/patch/README.md
+++ b/evidence/drama3-b300-scale150/patch/README.md
@@ -5,5 +5,5 @@ No winning candidate is nominated. Each experiment records its source revision b
 The implementation and rejected trials are reviewable in
 [runtime draft PR #1](https://github.com/david6666666/drama-3-single-b300-inference/pull/1).
 Latest published evidence is pinned at
-[`02aa919`](https://github.com/david6666666/drama-3-single-b300-inference/tree/02aa9198d8c1d0e0ef5ba28b1cc64d2c0776fb8b/evidence/20261016-ceiling).
+[`13d0566`](https://github.com/david6666666/drama-3-single-b300-inference/tree/13d0566bf71b2ccf634c79cdef3aea938decc96e/evidence/20261016-ceiling).
 This is an evidence publication revision, not a nominated performance winner.
diff --git a/evidence/drama3-b300-scale150/results/diagnostic-lessons.md b/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
index 4dbe00b03..3759043c2 100644
--- a/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
+++ b/evidence/drama3-b300-scale150/results/diagnostic-lessons.md
@@ -349,10 +349,10 @@ Do not substitute a plausible library reimplementation for the actual frozen
 native arithmetic reference; test layouts, scales, signed zero, partial routes,
 changed graph metadata and retained storage before timing.
 
-Resource capacity must include dynamic register allocation and actual TMEM,
-not only a cubin's static register header or shared bytes. The current FC1 roles
-request53248 registers per CTA, so its small shared-memory working set does not
-establish two-CTA residency. Shared intermediates require producer publication,
+Resource capacity must be checked in final machine code and launch attributes.
+Correction: the FC1 source hints sum to53248, but the tested CUBIN reports REG32
+and contains no SASS register-reconfiguration instruction. The source-hint sum
+does not prove actual allocation or a one-CTA restriction. Shared intermediates require producer publication,
 consumer-completion synchronization and explicit lifetime reuse. Individually
 exact stages do not establish a connected cluster or end-to-end performance.
 
@@ -368,7 +368,7 @@ these negative results, and none changes service acceptance or quality bounds.
 
 The component evidence is
 [pinned with all raw gates and repetitions](https://github.com/david6666666/drama-3-single-b300-inference/tree/f1dc31d57e4d29699a5bfbe5b0fc29ed5e163333/evidence/20261016-ceiling).
-The complete resident implementation remains in progress. C64 still needs
+The component-only stage is superseded by the complete-call results below. C64 still needs
 6765.407frames/s, all-concurrency repeated acceptance remains incomplete, and
 no successful README example is claimed.
 
@@ -390,12 +390,55 @@ fails its exact gate because four TMEM slices were stored as adjacent channels.
 A fixed32-channel permutation explains the observed bytes but does not pass the
 checker. The corresponding corrected M128 FC2 store preserves original K32/K128
 arithmetic and passes306 expanded checks across15native oracles/all256experts.
-Four160-register epilogue warps plus eight48-register roles budget32,768registers
-perCTA, versus53,248 for the M64 organization. That offers a two-CTA capacity
-hypothesis, not measured occupancy or service throughput.
+The32,768 versus53,248 source-hint sums originally motivated a capacity
+hypothesis. Final machine-code inspection rejects that occupancy inference;
+the explicit correction and actual resource facts are recorded below.
 
 The GDN H/O negative experiments were independently recomputed. Their pre/post
 GPU process checks do not prove uninterrupted exclusivity; reports now state
 that limitation. Both original local continuation gates still fail. Every new
 component result remains below full-service acceptance scope, and C64's6765.407
 frames/s target and confirmed WER/SIM requirements are unchanged.
+
+
+## Complete-call evidence rejects both initial native MoE candidates
+
+The resident chain now passes1371 full256-expert boundary/ownership checks;
+its build without global debug exports passes1262 more, including empty work.
+Yet all15 complete-call scenarios regress across three alternating repetitions.
+Real layers0/19/39 are14.16%,21.51%,37.11% slower than the optimized vendor path.
+Global intermediate traffic is a small part of expert-weight traffic. A fused
+implementation also changes phase serialization, cluster synchronization and
+PDL behavior; eliminating buffers alone does not imply a net speedup.
+
+The M128 global-buffer body passes1253 checks and626 timed-run checks but only
+reduces latency by1.304%,0.569%,2.871% in the three real cases. Each misses the
+predeclared5% continuation gate. Preserve that failure even though some synthetic
+routing cases improve more. Neither implementation is promoted into the service.
+
+Native routing metadata can legally change physical row order between calls.
+Check actual-map bijection, token/expert ownership, bounds and prefix counts
+before comparing logical intermediate bytes. An archived physical mapping is
+not an invariant. Both original harness failures remain preserved; no mapping
+was inferred from output values and no numeric tolerance was relaxed.
+
+Final SASS also corrects the earlier register explanation. The tested M64/M128
+Prims TS and resident CUBINs report REG30–32 and no register-reconfiguration
+instructions. Vendor FC1 reports REG128 and eight `USETMAXREG` instructions.
+Searching only for `SETMAXNREG` misses that Blackwell spelling. The source
+53,248→32,768 hint sums do not establish occupancy doubling. Actual thread and
+shared-memory changes remain facts; performance attribution requires controls.
+
+A more concrete prefix proposal preserves original prefill on misses and exports
+FP32 recurrent state at an aligned internal boundary. Its CPU-only conservative
+ready-cache budget is roughly0.243s; adding same-group forks gives roughly1.235s.
+These are optimistic estimates, not observed savings. Same-group sharing needs
+per-layer state forks and suffix-aware kernels; ordinary cache insertion does
+not provide it. This route alone does not explain the missing service budget.
+
+C64's target remains6765.407 frames/s, best complete-quality observation remains
+4726.771, and all-concurrency repeated acceptance is incomplete. The next PDL
+candidate has a separate protocol; it cannot retroactively pass either failed
+complete-call experiment.
+
+Completed-call evidence and the resource correction are [pinned at runtime 13d0566](https://github.com/david6666666/drama-3-single-b300-inference/tree/13d0566bf71b2ccf634c79cdef3aea938decc96e/evidence/20261016-ceiling).
diff --git a/evidence/drama3-b300-scale150/results/status.json b/evidence/drama3-b300-scale150/results/status.json
index bcdc9099b..bd15d1c6d 100644
--- a/evidence/drama3-b300-scale150/results/status.json
+++ b/evidence/drama3-b300-scale150/results/status.json
@@ -3,8 +3,8 @@
   "acceptance_issue": "https://github.com/ThinkFlowLab/OmniRSI/issues/10",
   "reason": "Best complete-quality C64 observation is4726.77 frames/s (one control arm, not a repeatable improvement claim), below6765.41. Native cache and partition candidates fail local/service performance promotion; the required all-concurrency repeated matrix remains incomplete.",
   "runtime_pull_request": "https://github.com/david6666666/drama-3-single-b300-inference/pull/1",
-  "runtime_evidence_commit": "02aa9198d8c1d0e0ef5ba28b1cc64d2c0776fb8b",
-  "runtime_evidence": "https://github.com/david6666666/drama-3-single-b300-inference/tree/02aa9198d8c1d0e0ef5ba28b1cc64d2c0776fb8b/evidence/20261016-ceiling",
+  "runtime_evidence_commit": "13d0566bf71b2ccf634c79cdef3aea938decc96e",
+  "runtime_evidence": "https://github.com/david6666666/drama-3-single-b300-inference/tree/13d0566bf71b2ccf634c79cdef3aea938decc96e/evidence/20261016-ceiling",
   "baseline_quality_groups": 21,
   "baseline_scored_waveforms": 22848,
   "packed_layer_graph_ownership_checks": 5076,
@@ -206,7 +206,12 @@
     "runtime_implemented": false,
     "cold_reuse_prompt_row_fraction": 0.34421045803919875,
     "admission_groups_are_fifo_proxy": true,
-    "raw_report": "reference-prefix/feasibility.md"
+    "raw_report": "reference-prefix/feasibility.md",
+    "conservative_c128_cpu_boundary_checks": 3645,
+    "ready_only_optimistic_seconds": 0.243,
+    "ready_plus_same_group_optimistic_seconds": 1.235,
+    "estimates_are_not_measured_savings": true,
+    "implementation_proposal": "reference-prefix/implementation-proposal.md"
   },
   "resident_fc1_prototype": {
     "original_exact_gate_pass": false,
@@ -229,15 +234,16 @@
     "fc2_native_oracle_cases": 15,
     "shared_activation_exact_checks": 2063,
     "shared_activation_cases": 137,
-    "fc1_dynamic_register_budget_per_cta": 53248,
-    "full_resident_chain_exact_pass": false,
-    "timing_run": false,
+    "full_resident_chain_exact_pass": true,
+    "timing_run": true,
     "runtime_promoted": false,
     "raw_report": "expert-resident-implementation-plan.md",
     "preserved_failures": [
       "expert-fc1-native",
       "expert-fc1-expanded",
-      "expert-fc2-native"
+      "expert-fc2-native",
+      "expert-resident-chain",
+      "expert-resident-timing"
     ],
     "shared_activation_dsl_exact_checks": 2066,
     "shared_activation_dsl_cases": 137,
@@ -254,8 +260,47 @@
       8
     ],
     "resident_single_expert_total_shared_bytes": 79488,
-    "resident_full_256_experts_gate_pass": false,
-    "original_host_abi_attempt_failed_before_kernel": true
+    "resident_full_256_experts_gate_pass": true,
+    "original_host_abi_attempt_failed_before_kernel": true,
+    "fc1_source_register_hint_sum_not_allocation": 53248,
+    "full_resident_chain_checks": 1371,
+    "zero_debug_checks": 1262,
+    "local_performance_gate_pass": false,
+    "raw_full_report": "expert-resident-full/summary.md",
+    "raw_timing_report": "expert-resident-timing-v2/summary.md",
+    "register_interpretation_correction": "register-lowering-audit/report.md",
+    "real_complete_call_medians": [
+      {
+        "case": "L0/real",
+        "primary_real": true,
+        "median_ms": {
+          "native": 0.12129887580871582,
+          "resident": 0.13847359657287597
+        },
+        "latency_reduction": -0.14159010666549054,
+        "case_continuation_pass": false
+      },
+      {
+        "case": "L19/real",
+        "primary_real": true,
+        "median_ms": {
+          "native": 0.1005072021484375,
+          "resident": 0.122128324508667
+        },
+        "latency_reduction": -0.21512012968282201,
+        "case_continuation_pass": false
+      },
+      {
+        "case": "L39/real",
+        "primary_real": true,
+        "median_ms": {
+          "native": 0.08140735626220703,
+          "resident": 0.11161855697631835
+        },
+        "latency_reduction": -0.3711114339200905,
+        "case_continuation_pass": false
+      }
+    ]
   },
   "gdn_prefill_ho_fusion": {
     "source_sha": "e7cda1eba083603b88e1f2e89688e1ff2ced7fc2",
@@ -309,10 +354,29 @@
     ],
     "K_scale_block": 128,
     "actual_threads_per_cta": 384,
-    "declared_register_budget_per_cta": 32768,
     "observed_occupancy_claim": false,
     "runtime_promoted": false,
-    "timing_run": false,
-    "raw_report": "expert-m128-fc2-expanded/report.md"
+    "timing_run": true,
+    "raw_report": "expert-m128-body-timing/report.md",
+    "source_register_hint_sum_not_allocation": 32768,
+    "fc1_selected_expert_checks": 542,
+    "fc1_expanded_checks": 447,
+    "full_body_checks": 1253,
+    "timing_checks": 626,
+    "local_performance_gate_pass": false,
+    "real_latency_reduction_percent": [
+      1.304,
+      0.569,
+      2.871
+    ],
+    "register_interpretation_correction": "register-lowering-audit/report.md"
+  },
+  "register_allocation_audit": {
+    "cpu_only": true,
+    "cubins_reproduced": 6,
+    "prototype_register_reconfiguration_instructions": 0,
+    "vendor_fc1_register_reconfiguration_instructions": 8,
+    "source_hint_occupancy_inference_withdrawn": true,
+    "raw_report": "register-lowering-audit/report.md"
   }
 }