agentculture
diff --git a/‎.devague/current‎
Lines changed: 1 addition & 1 deletion b/‎.devague/current‎
Lines changed: 1 addition & 1 deletion
diff --git a/‎.devague/current_plan‎
Lines changed: 1 addition & 1 deletion b/‎.devague/current_plan‎
Lines changed: 1 addition & 1 deletion
diff --git a/‎.devague/frames/lobes-now-defaults-to-a-spark-duo-the-qwen3-6-27b.json‎
Lines changed: 314 additions & 0 deletions b/‎.devague/frames/lobes-now-defaults-to-a-spark-duo-the-qwen3-6-27b.json‎
Lines changed: 314 additions & 0 deletions
@@ -1 +1 @@
-lobes-adds-a-third-middle-model-tier-and-pressure
+lobes-now-defaults-to-a-spark-duo-the-qwen3-6-27b
@@ -1 +1 @@
-lobes-adds-a-third-middle-model-tier-and-pressure
+lobes-now-defaults-to-a-spark-duo-the-qwen3-6-27b
@@ -0,0 +1,314 @@
+{
+  "slug": "lobes-now-defaults-to-a-spark-duo-the-qwen3-6-27b",
+  "title": "lobes now defaults to a Spark duo: the Qwen3.6-27B-MTP primary paired with a Gemma 4 12B NVFP4 multimodal worker that takes over the normal/middle tier, replacing the text-only Qwen3-14B (demoted to a legacy candidate profile); the duo gives Spark multimodal senses, a diverse non-Qwen mind, and a DSpark speculative-decoding test target, with old models kept behind explicit legacy profiles.",
+  "schema_version": 1,
+  "status": "exported",
+  "created": "2026-06-30T12:07:56Z",
+  "updated": "2026-06-30T12:35:33Z",
+  "claims": [
+    {
+      "id": "c1",
+      "kind": "announcement",
+      "text": "lobes now defaults to a Spark duo: the Qwen3.6-27B-MTP primary paired with a Gemma 4 12B NVFP4 multimodal worker that takes over the normal/middle tier, replacing the text-only Qwen3-14B (demoted to a legacy candidate profile); the duo gives Spark multimodal senses, a diverse non-Qwen mind, and a DSpark speculative-decoding test target, with old models kept behind explicit legacy profiles.",
+      "origin": "user",
+      "status": "confirmed",
+      "honesty_conditions": [
+        {
+          "id": "h9",
+          "text": "a fresh lobes install with no overrides serves exactly the Qwen 'main' + Gemma 'multimodal' duo, and every legacy model (14B, minor, etc.) is reachable only behind an explicit profile/override",
+          "status": "confirmed"
+        }
+      ],
+      "hard_questions": [],
+      "links": []
+    },
+    {
+      "id": "c2",
+      "kind": "audience",
+      "text": "Spark's agent stack and the Culture mesh peers that consume the lobes OpenAI endpoint \u2014 they gain a multimodal worker and a diverse second mind alongside the Qwen primary",
+      "origin": "llm",
+      "status": "confirmed",
+      "honesty_conditions": [
+        {
+          "id": "h10",
+          "text": "a Spark agent / Culture mesh peer reaches the multimodal gear through the existing lobes OpenAI endpoint with no client change beyond sending model=multimodal",
+          "status": "confirmed"
+        }
+      ],
+      "hard_questions": [],
+      "links": []
+    },
+    {
+      "id": "c3",
+      "kind": "after_state",
+      "text": "lobes serve / fleet up brings up the Qwen3.6-27B-MTP primary + a Gemma 4 12B NVFP4 multimodal worker by default; callers send model=normal and reach Gemma (vision-capable), model=hard reaches the Qwen primary, with the 14B kept only as an opt-in legacy candidate",
+      "origin": "llm",
+      "status": "rejected",
+      "honesty_conditions": [],
+      "hard_questions": [],
+      "links": []
+    },
+    {
+      "id": "c4",
+      "kind": "before_state",
+      "text": "today the normal/middle tier is a text-only nvidia/Qwen3-14B-NVFP4 that is opt-in (COMPOSE_PROFILES=middle), never load-tested (status=configured), and same-family as the primary \u2014 so the fleet has no multimodal generate gear and no non-Qwen mind",
+      "origin": "llm",
+      "status": "confirmed",
+      "honesty_conditions": [
+        {
+          "id": "h11",
+          "text": "this matches today's shipped state, verifiable in the repo: the 14B middle is opt-in (COMPOSE_PROFILES=middle), status=configured in catalog.py, hermes parser, text-only \u2014 and there is no multimodal or non-Qwen generate gear in the fleet",
+          "status": "confirmed"
+        }
+      ],
+      "hard_questions": [],
+      "links": []
+    },
+    {
+      "id": "c5",
+      "kind": "why_it_matters",
+      "text": "a multimodal worker lets Spark see screenshots/UI and images (the MTP primary is text-only); a non-Qwen family gives genuinely diverse reasoning for review/triage; and Gemma 4 gives a concrete DSpark/MTP speculative-decoding acceleration target on Blackwell",
+      "origin": "llm",
+      "status": "confirmed",
+      "honesty_conditions": [
+        {
+          "id": "h12",
+          "text": "an image-bearing request the text-only primary cannot serve is answered by the multimodal gear, and the served family is observably Gemma (a different lineage from the Qwen primary/minor)",
+          "status": "confirmed"
+        }
+      ],
+      "hard_questions": [],
+      "links": []
+    },
+    {
+      "id": "c6",
+      "kind": "boundary",
+      "text": "not removing the Qwen3-14B or any catalog gear (kept as legacy candidates); not changing the embed/rerank pooling gears, the audio overlay, or the minor 4B LoRA base; not adding a lobes train verb; DSpark stays a disabled-by-default experiment",
+      "origin": "llm",
+      "status": "confirmed",
+      "honesty_conditions": [
+        {
+          "id": "h13",
+          "text": "after the change the embed/rerank/audio/minor gears and all other catalog entries are unchanged except the 14B role demotion and the new Gemma entry; no lobes train verb is added; DSpark ships off",
+          "status": "confirmed"
+        }
+      ],
+      "hard_questions": [],
+      "links": []
+    },
+    {
+      "id": "c7",
+      "kind": "success_signal",
+      "text": "a smoke test confirms both default models are reachable (model=hard text + model=normal multimodal image round-trip) and that a legacy profile (e.g. the 14B) can still be selected explicitly",
+      "origin": "llm",
+      "status": "confirmed",
+      "honesty_conditions": [
+        {
+          "id": "h14",
+          "text": "a committed smoke test asserts model=main (text) and model=multimodal (image round-trip) both return valid output, and that an explicit legacy profile (e.g. the 14B) still boots and serves",
+          "status": "confirmed"
+        }
+      ],
+      "hard_questions": [],
+      "links": []
+    },
+    {
+      "id": "c8",
+      "kind": "requirement",
+      "text": "Gemma serves WITH its vision tower enabled (NOT --language-model-only) so model=normal accepts image inputs \u2014 multimodal coverage is the headline capability",
+      "origin": "llm",
+      "status": "confirmed",
+      "honesty_conditions": [
+        {
+          "id": "h1",
+          "text": "vLLM on the production image serves Gemma 4 12B with the vision tower and returns a correct answer to an image+text request via the gateway model=normal route",
+          "status": "confirmed"
+        }
+      ],
+      "hard_questions": [],
+      "links": []
+    },
+    {
+      "id": "c9",
+      "kind": "requirement",
+      "text": "Gemma takes role_hint=middle and the normal tier alias in catalog.py + TIER_ROLE, replacing nvidia/Qwen3-14B-NVFP4 as what model=normal resolves to",
+      "origin": "llm",
+      "status": "rejected",
+      "honesty_conditions": [
+        {
+          "id": "h2",
+          "text": "catalog test (tool_parser==infer_parser, doc exists) stays green after Gemma replaces the 14B as middle, and model=normal resolves to Gemma at the gateway",
+          "status": "proposed"
+        }
+      ],
+      "hard_questions": [],
+      "links": []
+    },
+    {
+      "id": "c10",
+      "kind": "requirement",
+      "text": "the Qwen-primary + Gemma duo is the DEFAULT served fleet (Gemma moves from opt-in COMPOSE_PROFILES=middle to default-on); old gears require an explicit legacy profile/override to serve",
+      "origin": "llm",
+      "status": "rejected",
+      "honesty_conditions": [
+        {
+          "id": "h3",
+          "text": "after the change, a fresh lobes serve/fleet up with no extra flags brings up BOTH the primary and Gemma, and brings up NEITHER the 14B nor any other legacy gear",
+          "status": "proposed"
+        }
+      ],
+      "hard_questions": [],
+      "links": []
+    },
+    {
+      "id": "c11",
+      "kind": "requirement",
+      "text": "a disabled-by-default experiment wires the DeepSeek DSpark draft (deepseek-ai/dspark_gemma4_12b_block7) as a Gemma4 speculative-decoding target, off unless explicitly enabled",
+      "origin": "llm",
+      "status": "rejected",
+      "honesty_conditions": [
+        {
+          "id": "h4",
+          "text": "the DSpark experiment is provably off by default (no speculative-config on the default Gemma gear) and can be turned on by one documented toggle",
+          "status": "proposed"
+        }
+      ],
+      "hard_questions": [],
+      "links": []
+    },
+    {
+      "id": "c12",
+      "kind": "requirement",
+      "text": "the duo + pooling gears fit the 128GB GB10 unified memory budget with measured gpu-memory-utilization values (Gemma multimodal weights + ViT + image KV vs the 14B's 0.12)",
+      "origin": "llm",
+      "status": "confirmed",
+      "honesty_conditions": [
+        {
+          "id": "h5",
+          "text": "a measured run shows primary + Gemma + embed + rerank (+ minor if kept) all healthy and co-resident under 1.0 total util on the 128GB GB10",
+          "status": "confirmed"
+        }
+      ],
+      "hard_questions": [],
+      "links": []
+    },
+    {
+      "id": "c13",
+      "kind": "after_state",
+      "text": "lobes serve (and fleet up) bring up the Qwen3.6-27B-MTP 'main' gear + the Gemma 4 12B NVFP4 'multimodal' gear by default; the three live generate gears are addressed as main / minor / multimodal, and model=multimodal reaches Gemma for image+text",
+      "origin": "user",
+      "status": "confirmed",
+      "honesty_conditions": [
+        {
+          "id": "h6",
+          "text": "model=multimodal routes to Gemma at the gateway, model=main routes to the Qwen primary, and the catalog test stays green with role_hint='multimodal' present and the 14B as a candidate",
+          "status": "confirmed"
+        }
+      ],
+      "hard_questions": [],
+      "links": []
+    },
+    {
+      "id": "c14",
+      "kind": "requirement",
+      "text": "Gemma gets a new role_hint='multimodal' in catalog.py; the old 'middle' role and 'normal' alias are deprecated; the normal tier slot resolves to Gemma for back-compat, and nvidia/Qwen3-14B-NVFP4 is demoted to role_hint=candidate (legacy, kept)",
+      "origin": "user",
+      "status": "confirmed",
+      "honesty_conditions": [
+        {
+          "id": "h7",
+          "text": "a clean lobes serve with no flags yields exactly two generate gears up (main + multimodal) and zero legacy gears, verifiable via /v1/models and the pressure/status output",
+          "status": "confirmed"
+        }
+      ],
+      "hard_questions": [],
+      "links": []
+    },
+    {
+      "id": "c15",
+      "kind": "requirement",
+      "text": "lobes serve no longer means single-model: with no extra flags it brings up BOTH main (Qwen primary) and multimodal (Gemma); minor, the 14B, and every other legacy gear require an explicit profile/override to serve",
+      "origin": "user",
+      "status": "confirmed",
+      "honesty_conditions": [
+        {
+          "id": "h8",
+          "text": "the default Gemma gear boots with native-MTP speculative-config active (measurable draft acceptance > 0), and flipping the DSpark toggle off is the default state with no DSpark weights loaded",
+          "status": "confirmed"
+        }
+      ],
+      "hard_questions": [],
+      "links": []
+    },
+    {
+      "id": "c16",
+      "kind": "requirement",
+      "text": "the default Gemma 'multimodal' gear serves NVFP4 + its native MTP assistant (speculative decoding ON); the DeepSeek DSpark draft (dspark_gemma4_12b_block7) is a separate, disabled-by-default experiment toggle",
+      "origin": "user",
+      "status": "confirmed",
+      "honesty_conditions": [
+        {
+          "id": "h15",
+          "text": "the default Gemma gear's compose command carries a native-MTP --speculative-config and NO DSpark draft, and a single documented env toggle (off by default) swaps in deepseek-ai/dspark_gemma4_12b_block7",
+          "status": "confirmed"
+        }
+      ],
+      "hard_questions": [],
+      "links": []
+    },
+    {
+      "id": "c17",
+      "kind": "requirement",
+      "text": "the default 'multimodal' gear ALSO serves Gemma 4's native AUDIO modality (audio-in understanding / ASR via the Gemma4Unified multimodal embedder, which vLLM supports), so model=multimodal accepts audio content in a chat request alongside text+image \u2014 distinct from and NOT replacing the existing /v1/audio/* overlay (Parakeet STT / Chatterbox TTS) per c6",
+      "origin": "user",
+      "status": "confirmed",
+      "honesty_conditions": [
+        {
+          "id": "h16",
+          "text": "vLLM on our image serves the chosen Gemma 4 12B build with the audio modality active, and model=multimodal returns correct text for an audio-bearing chat request (e.g. a short transcription) without disabling vision or the existing /v1/audio/* overlay",
+          "status": "confirmed"
+        }
+      ],
+      "hard_questions": [],
+      "links": []
+    }
+  ],
+  "open_vagueness": [
+    {
+      "id": "v1",
+      "text": "exact HF checkpoint id for the default Gemma 4 12B NVFP4 gear (accepted plan risk \u2014 verify+pick during implementation, like the 14B t1). NVFP4 12B builds PROVABLY EXIST: sakamakismile/gemma-4-12B-coder-fable5-composer2.5-MTP-NVFP4 (SAME publisher as the lobes primary, NVFP4+native-MTP \u2014 leading candidate, matches the MTP-on-by-default decision), AxionML/Gemma-4-12B-NVFP4, coolthor/gemma-4-12B-it-NVFP4A16. NVIDIA ships only 31B + 26B-A4B NVFP4, so the 12B is community.",
+      "kind": "unknown_nonblocking",
+      "claim_id": "c16"
+    },
+    {
+      "id": "v2",
+      "text": "vLLM --tool-call-parser to use for Gemma 4 (Gemma has no strong native tool format; must add a runtime/_parser.py infer_parser rule so the catalog test passes)",
+      "kind": "unknown_nonblocking",
+      "claim_id": null
+    },
+    {
+      "id": "v3",
+      "text": "whether the chosen Gemma 4 12B NVFP4 build loads non-gibberish on the current Blackwell vLLM image (nvcr.io/nvidia/vllm:26.04) or needs a newer engine \u2014 a load-validation task (like the 14B t9), now LOWER RISK: vLLM registers Gemma4UnifiedForConditionalGeneration and auto-detects NVFP4 (may not even need --quantization), and Gemma4 is a mainstream multimodal arch, NOT the hybrid Gated-DeltaNet/FLA arch that hit the Qwen3.5 sm_120/121 gibberish bug. Verify the image engine version is new enough before promoting status to load-tested.",
+      "kind": "unknown_nonblocking",
+      "claim_id": "c12"
+    },
+    {
+      "id": "v4",
+      "text": "the exact vLLM --speculative-config method+JSON for (a) Gemma4 NATIVE MTP (the default gear's spec-decode, analogous to the primary's {method: qwen3_5_mtp}) and (b) the DSpark experiment draft. DSpark shipped 2026-06-27 via DeepSeek's DeepSpec codebase (deepseek-ai/dspark_gemma4_12b_block7, validated on Gemma4-12B, vLLM-servable). Gemma4 assistant checkpoints use the Gemma4 MTP path, not a generic draft \u2014 confirm the method string against the served checkpoint's config.",
+      "kind": "unknown_nonblocking",
+      "claim_id": "c16"
+    },
+    {
+      "id": "v5",
+      "text": "measured GPU util for multimodal Gemma (vision tower + image KV) \u2014 does 0.12 hold or does the duo need primary/util retuning to fit 128GB",
+      "kind": "unknown_nonblocking",
+      "claim_id": null
+    },
+    {
+      "id": "v6",
+      "text": "Gemma 4's native ASR (model=multimodal audio-in) OVERLAPS the Parakeet STT sidecar behind /v1/audio/transcriptions \u2014 decide in planning whether Gemma subsumes, augments, or co-exists with Parakeet (the OpenAI /v1/audio/* endpoint shape still needs Parakeet/Chatterbox; Gemma audio-in is chat-only). Bonus: Gemma 4 also ingests VIDEO \u2014 out of scope for v1 but available.",
+      "kind": "follow_up",
+      "claim_id": null
+    }
+  ]
+}
Original file line number	Diff line number	Diff line change
`@@ -1 +1 @@`
`1`		`-lobes-adds-a-third-middle-model-tier-and-pressure`
	`1`	`+lobes-now-defaults-to-a-spark-duo-the-qwen3-6-27b`