bong-water-water-bong · Nov 11, 2026 · Nov 11, 2026 · Nov 11, 2026 · Nov 11, 2026 · Nov 11, 2026
diff --git a/docs/arch-gaps.md b/docs/arch-gaps.md
@@ -1087,3 +1087,14 @@ the single undocumented entry out of 37.
   MLA-style geometry (`kv_lora_rank` 512, `q_lora_rank` 1536, `qk_rope_head_dim` 64, `v_head_dim` 128, 32 heads /
   16 KV, IM 4096). **Not an alias of `deepseek_v41`**: nothing at the pin decodes AW-SECD weights, so mapping it to
   that architecture would name a backend that cannot read the file. Reported, not alerted (`"watch": "report"`).
+
+## 2026-11-11: `OpensciForCausalLM` — reviewed remote code (engine#2171)
+
+* `OpensciForCausalLM` (`model_type` `opensci`, 121 models in the 2026-11-11 census) — the open-sci reference models
+  (for example `open-sci/open-sci-1.7b-nemotron-cc-1T-oh-dcft`: hidden 2048, 24 layers, 32 heads / 32 KV,
+  `qk_layernorm`, vocab 50304). **Published as remote code**: `config.json` carries an `auto_map` to its own
+  `configuration_opensci.OpensciConfig` and `modeling_opensci.OpensciForCausalLM`, so no pinned backend accepts the
+  class without that code. **Not an alias** of any supported family. Reported, not alerted (`"watch": "report"`).
+  The other spelling, `OpenSciForCausalLM` (7 models, same `model_type` and the same `auto_map`), is a separate exact
+  key: `census_watch.py` matches architecture classes exactly, so that spelling needs its own entry and is not covered
+  by this one.
diff --git a/docs/npu.md b/docs/npu.md
@@ -1916,6 +1916,34 @@ route loads and a captured `--device npu` run stands behind it; the regenerated
 GGUF-on-NPU forward), while every D3-* pack in this body declares the dense `qwen3_5` family, so
 the two are not the same answer and the registry keeps them apart.
 
+The answers are also machine-readable, in `registry/flm_answers.json`, because CI keeps no watcher state: every scheduled
+run sees the whole org as new, and before that file each family answered above printed `!! NPU FLM` again and re-filed
+the census issue the day after it closed (engine#2171; the 2026-11-11 run alerted on 31 packs, every one already
+answered or tracked). A pack whose declared architecture is listed there prints `. NPU FLM` / `. FLM` with the issue
+that tracks it and does not alert; a family with no entry still alerts. The test fails on an entry for a family the
+registry serves on the NPU route, so a family that starts to load drops its answer. Keyed by the architecture each
+pack declares (a config with no `architectures` declares its `model_type`):
+
+| declared architecture | packs on 2026-11-11 | answer | tracked by |
+|---|---|---|---|
+| `MiniCPMV4_7ForConditionalGeneration` | MiniCPM-V-4.7-1B, MiniCPM-V-4.7-35B-A3B | not served: a VLM class the pin does not register | engine#1788 |
+| `MiniCPMOcrForConditionalGeneration` | MiniCPM-OCR-1.0 | not served: an OCR/vision pack; the route claims text decoders only | engine#1786 |
+| `HYV3ForCausalLM` | Hy-Memory-30B-A3B | not served: `hy_v3`, hrx only | engine#1786 |
+| `DFlashDraftModel` | Nemotron-3.5-Bolt-E4B-DFlash | not served: a draft head, not a decoder | engine#1780 |
+| `Gemma4ForConditionalGeneration`, `Gemma4UnifiedForConditionalGeneration` | Nemotron-3.5-Bolt-E4B, Gemma4-E4B-IT, Gemma4-E2B-IT, Gemma4-12B-IT | not served: `gemma4`, hrx only | engine#1786 |
+| `HunYuanDenseV1ForCausalLM` | Hy-MT2-1.8B | not served: `hunyuan-dense`, hrx only | engine#1786 |
+| `Gemma3ForCausalLM`, `gemma3_text` | Gemma3-270M, Gemma3-1B, Gemma3-4B, Gemma3-4B-Text, medgemma-4b-it, medgemma-1.5-4b-it, Translategemma-4B-Instruct | not served: `gemma3`, hrx only | engine#1786 |
+| `Lfm2ForCausalLM` | LFM2-1.2B, LFM2-2.6B, LFM2-1.2B-Tool, LFM2-2.6B-Transcript, LFM2.5-1.2B, LFM2.5-1.2B-Thinking | not in the registry's NPU set: `lfm2`, hrx only (LFM2.5-1.2B on the per-op route is engine#2170, which adds no `NPU_ARCHS` entry) | engine#1786 |
+| `GptOssForCausalLM` | GPT-OSS-20B, GPT-OSS-Safeguard-20b | not served: `gpt-oss`, hrx only | engine#1786 |
+| `Qwen3_5MoeForConditionalGeneration` | Qwen3.6-35B-A3B | not served for this class: the route serves `qwen3_5_moe` through `Qwen3_5MoeForCausalLM` (GGUF on the NPU); the VLM wrapper class is hrx only, and FastFlowLM's own pack is not measured on the route | engine#1786 |
+| `Qwen2_5_VLForConditionalGeneration` | Qwen2.5-VL-3B-Instruct | not served: `qwen2vl`, hrx only (the route serves the text-only `qwen2` class) | engine#1786 |
+| `Phi3ForCausalLM` | Phi4-mini-Instruct | not served: `phi3`, hrx only | engine#1786 |
+| `Gemma3TextModel` | Embedding-Gemma-300M | not served: `gemma-embedding`, an embedding model, hrx only | engine#1786 |
+| `WhisperForConditionalGeneration` | Whisper-V3-Turbo | not served: a speech encoder-decoder the registry does not map | engine#1783 |
+
+`FastFlowLM/smolvla-NPU2` is answered by its repo id (`repos` in the same file): it has no config and no `*.q4nx`
+file, so there is no architecture to key on and nothing for the route to load.
+
 ### FastFlowLM's own containers, served as published: measured rows (engine#1786)
 
 The tables above answer "does the registry map this family". This one answers a different question for the packs

diff --git a/registry/flm_answers.json b/registry/flm_answers.json
@@ -0,0 +1,91 @@
+{
+ "about": "Recorded answers for FastFlowLM packs this engine does not serve, read by tools/census_watch.py (engine#2171). CI keeps no watcher state, so without this file every scheduled run re-alerted on families docs/npu.md had already answered. A pack whose declared architecture (or, for a repo with no config, whose id) is listed here prints `. NPU FLM` / `. FLM` with the issue that tracks it and does not alert. Hand-written: add an entry only with a row for it in docs/npu.md, and remove it when the family is served (tests/census_watch_test.py fails on an answer for a family the registry serves on the NPU route) or when the issue that tracks it closes without serving it (no test can see that: check the `issue` numbers here whenever one of them closes).",
+ "architectures": {
+  "MiniCPMV4_7ForConditionalGeneration": {
+   "answer": "not served: a vision-language pack whose class the pin does not register; serving it on the NPU route is engine#1788",
+   "doc": "docs/npu.md#fastflowlms-npu2-drops-the-families-the-route-does-not-serve",
+   "issue": 1788
+  },
+  "MiniCPMOcrForConditionalGeneration": {
+   "answer": "not served: an OCR/vision pack whose class the pin does not register; the route claims text decoders only",
+   "doc": "docs/npu.md#fastflowlms-npu2-drops-the-families-the-route-does-not-serve",
+   "issue": 1786
+  },
+  "HYV3ForCausalLM": {
+   "answer": "not served: hy_v3 is mapped for hrx only; no hy_v3 model type or kernels in the NPU route",
+   "doc": "docs/npu.md#fastflowlms-npu2-drops-the-families-the-route-does-not-serve",
+   "issue": 1786
+  },
+  "DFlashDraftModel": {
+   "answer": "not served: a draft head, not a decoder; running the DFlash drafts as NPU sources is engine#1780",
+   "doc": "docs/npu.md#fastflowlms-npu2-drops-the-families-the-route-does-not-serve",
+   "issue": 1780
+  },
+  "Gemma4ForConditionalGeneration": {
+   "answer": "not served: gemma4 is mapped for hrx only; no gemma4 model type or kernels in the NPU route",
+   "doc": "docs/npu.md#fastflowlms-npu2-drops-the-families-the-route-does-not-serve",
+   "issue": 1786
+  },
+  "Gemma4UnifiedForConditionalGeneration": {
+   "answer": "not served: gemma4 is mapped for hrx only; no gemma4 model type or kernels in the NPU route",
+   "doc": "docs/npu.md#fastflowlms-npu2-drops-the-families-the-route-does-not-serve",
+   "issue": 1786
+  },
+  "HunYuanDenseV1ForCausalLM": {
+   "answer": "not served: hunyuan-dense is mapped for hrx only",
+   "doc": "docs/npu.md#fastflowlms-npu2-drops-the-families-the-route-does-not-serve",
+   "issue": 1786
+  },
+  "Gemma3ForCausalLM": {
+   "answer": "not served: gemma3 is mapped for hrx only (medgemma, translategemma and the Gemma3 text packs declare it)",
+   "doc": "docs/npu.md#fastflowlms-npu2-drops-the-families-the-route-does-not-serve",
+   "issue": 1786
+  },
+  "gemma3_text": {
+   "answer": "not served: gemma3 is mapped for hrx only (medgemma, translategemma and the Gemma3 text packs declare it)",
+   "doc": "docs/npu.md#fastflowlms-npu2-drops-the-families-the-route-does-not-serve",
+   "issue": 1786
+  },
+  "Lfm2ForCausalLM": {
+   "answer": "not served by the registry yet: lfm2 is mapped for hrx only; LFM2.5-1.2B on the per-op route is engine#2170, which adds no NPU_ARCHS entry",
+   "doc": "docs/npu.md#fastflowlms-npu2-drops-the-families-the-route-does-not-serve",
+   "issue": 1786
+  },
+  "GptOssForCausalLM": {
+   "answer": "not served: gpt-oss is mapped for hrx only",
+   "doc": "docs/npu.md#fastflowlms-npu2-drops-the-families-the-route-does-not-serve",
+   "issue": 1786
+  },
+  "Qwen3_5MoeForConditionalGeneration": {
+   "answer": "not served for this class: the route serves qwen3_5_moe through Qwen3_5MoeForCausalLM (GGUF on the NPU); the VLM wrapper class FastFlowLM declares is mapped for hrx only, and its own Q4NX pack is not measured on the route",
+   "doc": "docs/npu.md#fastflowlms-npu2-drops-the-families-the-route-does-not-serve",
+   "issue": 1786
+  },
+  "Qwen2_5_VLForConditionalGeneration": {
+   "answer": "not served: qwen2vl is mapped for hrx only; the route serves the text-only qwen2 class",
+   "doc": "docs/npu.md#fastflowlms-npu2-drops-the-families-the-route-does-not-serve",
+   "issue": 1786
+  },
+  "Phi3ForCausalLM": {
+   "answer": "not served: phi3 is mapped for hrx only",
+   "doc": "docs/npu.md#fastflowlms-npu2-drops-the-families-the-route-does-not-serve",
+   "issue": 1786
+  },
+  "Gemma3TextModel": {
+   "answer": "not served: gemma-embedding (an embedding model, not a decoder) is mapped for hrx only",
+   "doc": "docs/npu.md#fastflowlms-npu2-drops-the-families-the-route-does-not-serve",
+   "issue": 1786
+  },
+  "WhisperForConditionalGeneration": {
+   "answer": "not served: a speech encoder-decoder the registry does not map; the Whisper encoder on our runtime is engine#1783",
+   "doc": "docs/npu.md#fastflowlms-npu2-drops-the-families-the-route-does-not-serve",
+   "issue": 1783
+  }
+ },
+ "repos": {
+  "FastFlowLM/smolvla-NPU2": {
+   "answer": "not a pack: a config-less VLA repo with no *.q4nx file, so there is no architecture to map and nothing for the route to load",
+   "doc": "docs/npu.md#fastflowlms-npu2-drops-the-families-the-route-does-not-serve"
+  }
+ }
+}
diff --git a/registry/significant.json b/registry/significant.json
@@ -172,6 +172,11 @@
    "model_type": "deepseek_v4_flash_superexpert",
    "family": "DeepSeek V4.1 Flash SuperExpert, published as remote code and with AW-SECD-compressed weights (8.5 GB, claimed 0.962 retention); MLA-style geometry (kv_lora_rank 512, q_lora_rank 1536, qk_rope_head_dim 64, v_head_dim 128, 32 heads / 16 KV, IM 4096). Not an alias of deepseek_v41: nothing at the pin decodes AW-SECD weights, so mapping it would name a backend that cannot read it",
    "watch": "report"
+  },
+  "OpensciForCausalLM": {
+   "model_type": "opensci",
+   "family": "OpenSci (1.7B 'nemotron-hq' recipe). A REMOTE-CODE class: its config carries an auto_map pointing at modeling_opensci.py / configuration_opensci.py, so the pinned backends cannot accept the architecture without that custom code. Not an alias of any supported family, and not mappable until the code is vendored or upstreamed.",
+   "watch": "report"
   }
  }
 }
diff --git a/registry/support_pins.json b/registry/support_pins.json
@@ -7,9 +7,15 @@
    "why": "The official gguf Q4_0/Q4_1 -> Q4NX converter with per-architecture configs (configs/*.json, 18 of them: gemma3, gemma4, gpt-oss, lfm2, llama, nanbeige, phi4, qwen2, qwen2vl, qwen3, qwen3.5_{0.8b,2b,4b,9b,a3b,omni}, qwen3vl). docs/npu.md's repack path runs it (scripts/repack_gguf.py, ONEBIT_Q4NX_CONVERTER) and fails its declared-scale guard when the converter or config drops a critical scale. A per-arch config here is what makes a family packable on our side at all.  Reviewed 2026-11-08 (d1d5232 -> 93b8398, 2 commits, \"Qwen3.5 / Qwen3.5-MoE / Omni converters on stock llama.cpp GGUFs\"): three changes touch the family this route serves. (a) full-attention layers are now detected from `attn_q.weight` instead of \"every Nth layer\", because the d3 layouts are irregular -- the commit names full attention at 2, 5, 8, 9, 10, 11 for d3:edge, which is the same irregularity our own converter takes from `layer_types`; (b) an MTP block (nextn_predict_layers > 0, counted in block_count) no longer leaks out as a bogus extra decoder layer; (c) the Qwen35 value-head reorder is decided from the v/qk head ratio instead of feed_forward_length <= 6144, and a ratio > 2 raises. Stock Qwen3.5 0.8B/2B/4B (Q4_K, Q4_1) are tested 124/124 in that PR, and the output is byte-identical for GGUFs without an MTP block, so the repack path's declared-scale guard is unaffected; the entry point and configs/ layout that scripts/repack_gguf.py invokes are unchanged."
   },
   "vllm-project/semantic-router": {
-   "commit": "a9219240e6f9dd4d9e1bea83596565cc8f18989b",
-   "reviewed": "2026-11-09",
-   "why": "The Decision 3.0 (d3) source: the router calls d3 for routing, guardrails and tool selection, and FastFlowLM v1.0.11 serves those weights through /v1/systemone with the same HTTP API as the upstream d3 server. Engine #653 is our side of that; a change here is what a new d3 pack would follow.  Reviewed 2026-11-08 (632f087 -> b7adcfa, 152 commits, 300 files): **the d3 contract is unchanged** -- `d3_format.py`, `d3_server.py` and `d3_runtime.py` are not in the delta at all, and the only System One files touched are dashboard frontend pages (`SystemOnePlaygroundPage.tsx`, `systemOneTargets.ts`). What moved is router-side: `config/fragments/signal/decision/vela{2,3}-labels-spans.yaml` (their own newer decision models), a `config/recipes/decision-balance/` recipe, and a `canonical_decision_model` CRD. So what this route must speak is unchanged, and the vela2/vela3 names are upstream's packs, not ours.  Reviewed 2026-11-09 (b7adcfa -> a921924, 9 commits, engine #1667): the d3 contract is unchanged again -- no d3_* file is in the delta. The one System One file is the router's client, pkg/systemone/backend.go (#5144): a backend marked trusted_gateway is now refused before dispatch (errTrustedGatewayUnsupported), because System One dispatch does not carry the request's gateway credential. The wire protocol string stays vllm-sr/systemone@1, and FastFlowLM's /v1/systemone is not a trusted gateway, so nothing this route speaks or serves moves. The rest is router-side (header_mutation append, the lifecycle client's 5xx retry, contrastive jailbreak embedding validation, adaptive thinking on Responses backends, a repeated [DONE] in protocolcodec) plus a Claude Code plugin and maintainer docs."
+   "commit": "15b4fb5ad46b5c40a3bd99f35d0feb28997dac68",
+   "reviewed": "2026-11-11",
+   "why": "The Decision 3.0 (d3) source: the router calls d3 for routing, guardrails and tool selection, and FastFlowLM v1.0.11 serves those weights through /v1/systemone with the same HTTP API as the upstream d3 server. Engine #653 is our side of that; a change here is what a new d3 pack would follow.  Reviewed 2026-11-08 (632f087 -> b7adcfa, 152 commits, 300 files): **the d3 contract is unchanged** -- `d3_format.py`, `d3_server.py` and `d3_runtime.py` are not in the delta at all, and the only System One files touched are dashboard frontend pages (`SystemOnePlaygroundPage.tsx`, `systemOneTargets.ts`). What moved is router-side: `config/fragments/signal/decision/vela{2,3}-labels-spans.yaml` (their own newer decision models), a `config/recipes/decision-balance/` recipe, and a `canonical_decision_model` CRD. So what this route must speak is unchanged, and the vela2/vela3 names are upstream's packs, not ours.  Reviewed 2026-11-09 (b7adcfa -> a921924, 9 commits, engine #1667): the d3 contract is unchanged again -- no d3_* file is in the delta. The one System One file is the router's client, pkg/systemone/backend.go (#5144): a backend marked trusted_gateway is now refused before dispatch (errTrustedGatewayUnsupported), because System One dispatch does not carry the request's gateway credential. The wire protocol string stays vllm-sr/systemone@1, and FastFlowLM's /v1/systemone is not a trusted gateway, so nothing this route speaks or serves moves. The rest is router-side (header_mutation append, the lifecycle client's 5xx retry, contrastive jailbreak embedding validation, adaptive thinking on Responses backends, a repeated [DONE] in protocolcodec) plus a Claude Code plugin and maintainer docs.  Reviewed 2026-11-11 (a9219240 -> 15b4fb5a, 88 commits; engine#2171): **the d3 contract is unchanged** -- `vela3_d3/package.py`, `model.py` and `decode.py` have the same blob at both commits. Three commits touch the family: #5886 adds router config fragments for the decision model (config/fragments/signal/systemone/*, no runtime code), #6113 records ROCm golden answers for the d3 sizes (a new registry/golden_answers_vela3_d3.json and the test that reads it), and #6192 versions the System One HTTP API (systemone/0.2, vllm-sr/5.0; vllm-sr/4.0 stays a selectable compatible superset, `/v1/models` now reports api_version 5.0.0, a request may send `roles`, and `roles: null` makes an object state the user's content for span questions). Its own description says model input stays byte-identical on every family. Nothing here reaches npu/systemone.h, which copies the pack's d3_format.py/d3_runtime.py (choice/noul/score prompts and answers; no span questions and no api_version field). From this review on the pin watches `paths` only: the d3 family directory, the runtime's systemone.py and api/ (the HTTP contract), and config/schemas/systemone; router-side commits no longer re-file the census issue every day.",
+   "paths": [
+    "src/model-runtime/vllm_srun/families/vela3_d3",
+    "src/model-runtime/vllm_srun/systemone.py",
+    "src/model-runtime/vllm_srun/api",
+    "config/schemas/systemone"
+   ]
   }
  },
  "d3_sources": {