thor-worker-lobe-qwen3-6-35b-a3b.json

{
  "slug": "thor-worker-lobe-qwen3-6-35b-a3b",
  "title": "thor worker lobe (Qwen3.6-35B-A3B)",
  "schema_version": 2,
  "status": "exported",
  "created": "2026-07-31T06:31:54Z",
  "updated": "2026-07-31T06:55:18Z",
  "claims": [
    {
      "id": "c1",
      "kind": "announcement",
      "text": "Thor moves off the Gemma 4 31B muse and hosts unsloth/Qwen3.6-35B-A3B-NVFP4 as 'worker' \u2014 a new fast ground-work lobe with self-draft MTP",
      "origin": "user",
      "status": "confirmed",
      "honesty_conditions": [
        {
          "id": "h10",
          "text": "the announcement is falsifiable against live surfaces: the deployed Thor answers model=worker through the gateway and the acceptance transcript exists under docs/evidence/",
          "status": "confirmed",
          "instruction": ""
        }
      ],
      "hard_questions": [],
      "links": [],
      "instruction": "ship as a PR chain: catalog entry + eighth-role registry + opt-in gateway/compose/shape wiring + docs, then a live Thor acceptance run (measured budget, MTP acceptance, parser-pair transcript under docs/evidence/); version-bump every PR"
    },
    {
      "id": "c2",
      "kind": "requirement",
      "text": "lobes/catalog.py gains a NEW SupportedModel entry for unsloth/Qwen3.6-35B-A3B-NVFP4 with role_hint='worker' \u2014 distinct from the existing mmangkad/Qwen3.6-35B-A3B-NVFP4 candidate (32K native, moe_backend=marlin, NO speculative_config: nvidia's MTP config fails on that copy with a weight-shape mismatch). The unsloth export is 262144-token native and ships its OWN MTP module ('can act as its own speculative draft'), card-suggested --speculative-config '{\"method\": \"mtp\", \"num_speculative_tokens\": 2}' \u2014 a different method string than the 27B primary's qwen3_5_mtp",
      "origin": "llm",
      "status": "confirmed",
      "honesty_conditions": [
        {
          "id": "h1",
          "text": "the committed catalog entry's native_max_model_len / quantization / speculative_config are read from the unsloth checkpoint's actual config.json + hf_quant_config.json at implementation time, not the card prose \u2014 and if the checkpoint disagrees with the card (e.g. MTP fails to load), the entry records what was verified",
          "status": "confirmed",
          "instruction": ""
        }
      ],
      "hard_questions": [],
      "links": [],
      "instruction": ""
    },
    {
      "id": "c3",
      "kind": "requirement",
      "text": "the role registry widens to an EIGHTH first-class role: lobes/roles.py needs 'worker' entries in ROLES, ROLE_BACKEND, ROLE_ROLE_HINT, ROLE_PATH (/v1/chat/completions), ROLE_RESPONSIBILITIES, ROLE_FORBIDDEN, and ROLE_MAX_MODEL_LEN_ENV (WORKER_MAX_MODEL_LEN); lobes/profiles/schema.py ROLES (the Profile-machinery set) gains 'worker' so shape overrides can declare its budget knobs",
      "origin": "llm",
      "status": "confirmed",
      "honesty_conditions": [
        {
          "id": "h2",
          "text": "every roles surface (lobes capabilities, GET /capabilities, /v1/models alias resolution) lists exactly eight roles with worker present; no stale seven-role literal survives in code, tests, or docs",
          "status": "confirmed",
          "instruction": ""
        }
      ],
      "hard_questions": [],
      "links": [],
      "instruction": ""
    },
    {
      "id": "c4",
      "kind": "requirement",
      "text": "worker becomes the SECOND opt-in core role, mirroring muse end-to-end: shapes.OPT_IN_CORE_ROLES += 'worker'; gateway _config.py wires a WORKER_BASE_URL-gated backend, adds 'worker' to OPT_IN_BACKENDS (infeasible-by-default when unwired \u2014 model=worker 404s role_infeasible, never a silent fallback), and adds WORKER_* keys to FEASIBLE_ENV / PEER_ORIGIN_ENV / PEER_PROXY_ENV / PEER_API_KEY_ENV; profiles/builtin/base.toml vetoes worker on unrecognised cards exactly as it vetoes muse",
      "origin": "llm",
      "status": "confirmed",
      "honesty_conditions": [
        {
          "id": "h3",
          "text": "with no WORKER_* env set anywhere, every existing deployment renders byte-identically (goldens prove it); model=worker on a non-hosting box 404s role_infeasible \u2014 never a silent fallback",
          "status": "confirmed",
          "instruction": ""
        }
      ],
      "hard_questions": [],
      "links": [],
      "instruction": ""
    },
    {
      "id": "c5",
      "kind": "requirement",
      "text": "the fleet compose template gains a profile-gated vllm-worker service + WORKER_* knobs in env.example, and shape_render.OPT_IN_CORE_ACTIVATION_ENV maps worker \u2192 COMPOSE_PROFILES gate + WORKER_BASE_URL=http://vllm-worker:8000. The worker lane runs on the QWEN-lane image (the vllm-openai nightly pin the primary/minor lanes use), NOT the lobes/vllm-gemma4:local custom image the muse lane needs",
      "origin": "llm",
      "status": "confirmed",
      "honesty_conditions": [
        {
          "id": "h4",
          "text": "the rendered thor-worker deployment boots vllm-worker on the Qwen-lane image on the physical Thor and serves /v1/chat/completions through the gateway",
          "status": "confirmed",
          "instruction": ""
        }
      ],
      "hard_questions": [],
      "links": [],
      "instruction": ""
    },
    {
      "id": "c6",
      "kind": "requirement",
      "text": "a new built-in shape (working name thor-worker) hosts ['worker','embedder','reranker','stt','tts'], drops cortex/senses to peers, and carries the FULL worker declaration (model + budget knobs) in its own overrides \u2014 the card profiles stay silent on worker, exactly the thor-muse pattern. Its gpu_mem_util / max_model_len MUST be measured on the physical Thor before commit: shape budgets on unified-memory boxes are measured truths, not arithmetic (thor-muse's 0.55 was measured only after the 0.40 hypothesis was refused live)",
      "origin": "llm",
      "status": "confirmed",
      "honesty_conditions": [
        {
          "id": "h5",
          "text": "the shape TOML's committed budget values are the ones a live Thor boot measured, with any refused hypothesis recorded in the TOML comments (the thor-muse 0.40\u21920.55 pattern)",
          "status": "confirmed",
          "instruction": ""
        }
      ],
      "hard_questions": [],
      "links": [],
      "instruction": ""
    },
    {
      "id": "c7",
      "kind": "requirement",
      "text": "the worker lane serves the Qwen-family parser PAIR: --tool-call-parser=qwen3_coder (runtime._parser.infer_parser must resolve the unsloth id to qwen3_coder \u2014 the catalog pairing guard in tests/test_catalog.py enforces it) plus --reasoning-parser=qwen3, mirroring the cortex lane \u2014 never inferred from the model card, and verified live with skip_special_tokens:false per the per-family parser rule",
      "origin": "llm",
      "status": "confirmed",
      "honesty_conditions": [
        {
          "id": "h6",
          "text": "verified live on the worker lane with skip_special_tokens:false \u2014 a tool call parses into a tool_calls array (not content) and the reasoning trace separates via --reasoning-parser=qwen3",
          "status": "confirmed",
          "instruction": ""
        }
      ],
      "hard_questions": [],
      "links": [],
      "instruction": ""
    },
    {
      "id": "c8",
      "kind": "boundary",
      "text": "muse stays in-tree and supported (cite-don't-delete): the nvidia/Gemma-4-31B-IT-NVFP4 catalog entry, the muse role, and the thor-muse built-in shape are NOT removed \u2014 this idea changes which lobe the physical Thor HOSTS, not the muse contract. Whether any box still hosts muse afterwards is a deployment decision, not a code deletion",
      "origin": "llm",
      "status": "rejected",
      "honesty_conditions": [],
      "hard_questions": [],
      "links": [],
      "instruction": ""
    },
    {
      "id": "c9",
      "kind": "boundary",
      "text": "cortex/senses/embedder/reranker contracts are untouched and machine-as-brain stays byte-identical \u2014 worker is opt-in-core like muse: non-hosting shapes render nothing for it, only base.toml's veto emits WORKER_FEASIBLE=false, and a stale pre-worker .env defaults it to infeasible via OPT_IN_BACKENDS",
      "origin": "llm",
      "status": "confirmed",
      "honesty_conditions": [
        {
          "id": "h11",
          "text": "the shape goldens (machine-as-brain / spark / thor / base) render byte-identical before and after the change when no WORKER_* env is set",
          "status": "confirmed",
          "instruction": ""
        }
      ],
      "hard_questions": [],
      "links": [],
      "instruction": ""
    },
    {
      "id": "c10",
      "kind": "boundary",
      "text": "the pooling gears + audio overlay keep co-residing on Thor (mesh-brain decision 2: cheap gears co-reside on every box that wants them) \u2014 the new shape hosts embedder/reranker/stt/tts exactly as thor-muse does today",
      "origin": "llm",
      "status": "confirmed",
      "honesty_conditions": [
        {
          "id": "h12",
          "text": "the thor-worker shape's hosts list includes embedder/reranker/stt/tts and the rendered deployment serves them unchanged \u2014 same probes thor-muse passes today",
          "status": "confirmed",
          "instruction": ""
        }
      ],
      "hard_questions": [],
      "links": [],
      "instruction": ""
    },
    {
      "id": "c11",
      "kind": "non_goal",
      "text": "the mmangkad/Qwen3.6-35B-A3B-NVFP4 catalog entry stays a candidate unchanged (no promotion, no removal); docs/qwen3.6-35b-a3b-nvfp4.md gains the unsloth variant's story rather than being rewritten away",
      "origin": "llm",
      "status": "confirmed",
      "honesty_conditions": [],
      "hard_questions": [],
      "links": [],
      "instruction": ""
    },
    {
      "id": "c12",
      "kind": "assumption",
      "text": "the new built-in shape is named thor-worker, mirroring thor-muse's naming (card + hosted-lobe)",
      "origin": "llm",
      "status": "confirmed",
      "honesty_conditions": [],
      "hard_questions": [],
      "links": [],
      "instruction": ""
    },
    {
      "id": "c13",
      "kind": "requirement",
      "text": "worker's role contract (user decision q1): responsibilities include fast ground-work execution \u2014 tool_use + repo_action allowed, under cortex direction; forbidden_responsibilities: final_decision, security_decision. First role besides cortex permitted repo_action",
      "origin": "user",
      "status": "confirmed",
      "honesty_conditions": [
        {
          "id": "h7",
          "text": "capabilities output for worker lists repo_action among responsibilities and forbidden is exactly [final_decision, security_decision]; docs/colleague-stack.md's division-of-labour records the widening \u2014 every 'cortex is the only lobe that acts' wording is updated",
          "status": "confirmed",
          "instruction": ""
        }
      ],
      "hard_questions": [],
      "links": [],
      "instruction": "ROLE_RESPONSIBILITIES['worker'] = ground-work execution tokens + tool_use + repo_action; ROLE_FORBIDDEN['worker'] = ('final_decision','security_decision'); update docs/colleague-stack.md division-of-labour table"
    },
    {
      "id": "c14",
      "kind": "boundary",
      "text": "muse support is dropped as deployment reality (user decision q3): no box hosts the 31B, Thor's new shape declares no MUSE_PEER_ORIGIN, model=muse 404s role_infeasible without referral. The muse role/gear/shape code stays in-tree as dormant (cite-don't-delete; the q2 tier order still ranks worker < muse), with docs updated to say muse is unhosted",
      "origin": "user",
      "status": "confirmed",
      "honesty_conditions": [
        {
          "id": "h8",
          "text": "after the move, model=muse 404s role_infeasible with no hosted_by on every box; the thor-muse shape, catalog entry, and docs stay in-tree explicitly marked dormant/unhosted",
          "status": "confirmed",
          "instruction": ""
        }
      ],
      "hard_questions": [],
      "links": [],
      "instruction": "leave muse code/shape/catalog in-tree marked dormant; strip muse hosting from the deployed Thor .env; declare no MUSE_PEER_ORIGIN anywhere; docs say muse is unhosted mesh-wide"
    },
    {
      "id": "c15",
      "kind": "audience",
      "text": "Culture-mesh Colleagues that address the fleet by role/tier alias (model=worker via the gateway, lobes capabilities / GET /capabilities), plus the operator running the physical Thor",
      "origin": "llm",
      "status": "confirmed",
      "honesty_conditions": [
        {
          "id": "h13",
          "text": "the surfaces Colleagues actually consume (capabilities JSON keyed by role, gateway tier aliases) carry worker \u2014 the audience needs no lobes source access to discover it",
          "status": "confirmed",
          "instruction": ""
        }
      ],
      "hard_questions": [],
      "links": [],
      "instruction": ""
    },
    {
      "id": "c16",
      "kind": "after_state",
      "text": "Thor serves unsloth/Qwen3.6-35B-A3B-NVFP4 as 'worker' \u2014 the EIGHTH first-class Colleague role, a fast ground-work DOER (repo_action allowed, under cortex direction) with self-draft MTP at 262144 native \u2014 hosted via a thor-worker shape; muse is dormant mesh-wide (unhosted, no referral)",
      "origin": "llm",
      "status": "confirmed",
      "honesty_conditions": [
        {
          "id": "h14",
          "text": "verified live on the deployed Thor: lobes capabilities lists worker feasible+ready serving the unsloth id, and muse feasible:false with no hosted_by",
          "status": "confirmed",
          "instruction": ""
        }
      ],
      "hard_questions": [],
      "links": [],
      "instruction": ""
    },
    {
      "id": "c17",
      "kind": "before_state",
      "text": "Thor serves thor-muse (Gemma 4 31B muse at full 256K, util 0.55, gemma4 parser pair); no worker role exists in the contract; the catalog's only 35B-A3B is the 32K-native mmangkad copy with MTP explicitly not carried",
      "origin": "llm",
      "status": "confirmed",
      "honesty_conditions": [
        {
          "id": "h15",
          "text": "matches the recorded present: Thor's deployment runs thor-muse per the 2026-07-17 evidence transcript, and the catalog's mmangkad entry says 32K-native with no speculative_config",
          "status": "confirmed",
          "instruction": ""
        }
      ],
      "hard_questions": [],
      "links": [],
      "instruction": ""
    },
    {
      "id": "c18",
      "kind": "why_it_matters",
      "text": "the mesh gains a fast executor \u2014 ~3B-active MoE decode speed with 35B-class knowledge plus MTP self-draft \u2014 so ground work stops queueing on the 27B cortex; the 31B muse's 67 GiB on Thor bought creative divergence the mesh uses less than fast execution",
      "origin": "llm",
      "status": "confirmed",
      "honesty_conditions": [
        {
          "id": "h16",
          "text": "the speed rationale is measured, not assumed: the acceptance run records worker-lane decode tok/s with MTP on (and off, if feasible) via lobes measure/benchmark",
          "status": "confirmed",
          "instruction": ""
        }
      ],
      "hard_questions": [],
      "links": [],
      "instruction": ""
    },
    {
      "id": "c19",
      "kind": "success_signal",
      "text": "on the deployed Thor: lobes capabilities shows worker feasible+ready with truthful model/context/mtp; model=worker serves a chat completion whose tool call parses into tool_calls (not content); an acceptance transcript under docs/evidence/ records the measured budget + MTP acceptance; model=muse 404s role_infeasible with no hosted_by",
      "origin": "llm",
      "status": "confirmed",
      "honesty_conditions": [
        {
          "id": "h17",
          "text": "every listed signal is a checkable artifact: capabilities JSON, a chat transcript with a parsed tool_calls array, the docs/evidence/ file, and the muse role_infeasible 404 body",
          "status": "confirmed",
          "instruction": ""
        }
      ],
      "hard_questions": [],
      "links": [],
      "instruction": ""
    },
    {
      "id": "c20",
      "kind": "decision",
      "text": "the thor-worker shape ships the thor-muse way: budget knobs (gpu_mem_util / max_model_len), the sm_110 MoE backend choice, and the MTP speculative-config are MEASURED during implementation by a live boot on the physical Thor and committed with the measured evidence in the shape TOML comments; the worker role stays DECLARED/UNVALIDATED (#108) until the acceptance transcript lands under docs/evidence/ \u2014 the spec requires the measurement, it does not guess the values",
      "origin": "llm",
      "status": "confirmed",
      "honesty_conditions": [
        {
          "id": "h9",
          "text": "the exported spec and resulting plan contain an explicit live-measurement task on the physical Thor whose outputs (budget, MoE backend, MTP verdict) are committed as shape TOML + catalog data with evidence under docs/evidence/ \u2014 no committed value precedes its measurement",
          "status": "confirmed",
          "instruction": ""
        }
      ],
      "hard_questions": [],
      "links": [],
      "instruction": ""
    }
  ],
  "open_vagueness": [
    {
      "id": "v1",
      "text": "measured Thor budget for the 35B MoE worker lane: gpu_mem_util and max_model_len (full 262144 or trimmed) can only come from a live boot on the physical Thor \u2014 shape budgets on unified-memory boxes are measured truths, not arithmetic (thor-muse precedent: 0.40 hypothesis refused live, 0.55 measured) [re-parked non-blocking per confirmed decision c20: measured during implementation, role DECLARED/UNVALIDATED until evidence lands]",
      "kind": "unknown_nonblocking",
      "claim_id": null
    },
    {
      "id": "v2",
      "text": "MoE backend on Thor sm_110: the unsloth card recommends --moe-backend flashinfer_b12x with CUTE_DSL_ARCH=sm_121a \u2014 Spark-arch-specific, sm_121 is NOT sm_110; marlin is verified only GB10-solo on the mmangkad sibling. Which backend serves this checkpoint on Thor is a live measurement [re-parked non-blocking per confirmed decision c20: measured during implementation, role DECLARED/UNVALIDATED until evidence lands]",
      "kind": "unknown_nonblocking",
      "claim_id": null
    },
    {
      "id": "v3",
      "text": "does the unsloth copy's self-draft MTP actually load and accept on the deployed vLLM image on Thor? The mmangkad sibling's MTP failed with a weight-shape mismatch, and the card's method string ('mtp', 2 tokens) differs from the 27B's qwen3_5_mtp \u2014 MTP is a headline motivation, so this gates the perf claim [re-parked non-blocking per confirmed decision c20: measured during implementation, role DECLARED/UNVALIDATED until evidence lands]",
      "kind": "unknown_nonblocking",
      "claim_id": null
    },
    {
      "id": "v4",
      "text": "unsloth config.json quant_method (modelopt vs compressed-tensors) sets WORKER_QUANTIZATION \u2014 read from the checkpoint config at implementation time",
      "kind": "unknown_nonblocking",
      "claim_id": null
    },
    {
      "id": "v5",
      "text": "do the cortex-lane-only thinking knobs extend to the worker lane: preserve_thinking (#93), the qwen3_coder_thinking tool-parser plugin (colleague#320), and GATEWAY_FORCE_STRICT_TOOLS \u2014 Qwen3.6 is a thinking model, so the strict+thinking grammar break plausibly applies to worker too",
      "kind": "unknown_nonblocking",
      "claim_id": null
    }
  ],
  "scope_entries": [
    {
      "id": "s1",
      "surface": "lobes/catalog.py + HF card unsloth/Qwen3.6-35B-A3B-NVFP4",
      "finding": "catalog already holds the mmangkad sibling of this checkpoint as a candidate (32K native, marlin verified GB10-solo, MTP explicitly NOT carried \u2014 nvidia's config fails on it); the unsloth export verified on HF 2026-07-31: MoE 35B/~3B-active, 262144 native (\u21921M), includes its own MTP draft module, vLLM card syntax method=mtp num_speculative_tokens=2",
      "seeds": [
        "c2"
      ]
    },
    {
      "id": "s2",
      "surface": "lobes/roles.py + lobes/profiles/schema.py",
      "finding": "every role surface is a per-role dict keyed by name \u2014 seven roles today ('cortex','senses','muse','embedder','reranker','stt','tts'); muse (added as the seventh) is the exact structural precedent: generate-lane role, _CHAT_PATH, its own backend name, own MAX_MODEL_LEN env. schema.py ROLES is the 5-core subset shapes can override",
      "seeds": [
        "c3"
      ]
    },
    {
      "id": "s3",
      "surface": "lobes/profiles/shapes.py + lobes/gateway/_config.py + lobes/profiles/builtin/base.toml",
      "finding": "OPT_IN_CORE_ROLES=('muse',) with machine-as-brain's DEFAULT_HOSTED_ROLES derived by exclusion \u2014 adding 'worker' keeps machine-as-brain byte-identical automatically; OPT_IN_BACKENDS=frozenset({'muse'}) implements unwired\u2192infeasible; all four peer-channel env dicts key by backend name; base.toml [roles.muse] feasible=false is the unknown-card veto precedent",
      "seeds": [
        "c4"
      ]
    },
    {
      "id": "s4",
      "surface": "lobes/templates/fleet/docker-compose.yml + env.example + lobes/profiles/shape_render.py",
      "finding": "vllm-muse is the template precedent: profile-gated ('muse' compose profile), no host port, MUSE_* knob family, activation env written by the shape render (OPT_IN_CORE_ACTIVATION_ENV={'muse': {'MUSE_BASE_URL': ...}}); muse rides image lobes/vllm-gemma4:local while the Qwen lanes ride vllm/vllm-openai@sha256:7c5a10e9... (VLLM_NIGHTLY_IMAGE) \u2014 worker is a Qwen3.6 and belongs on the latter",
      "seeds": [
        "c5"
      ]
    },
    {
      "id": "s5",
      "surface": "lobes/profiles/builtin_shapes/thor-muse.toml + tests/goldens/shapes",
      "finding": "thor-muse.toml is the exact structural template: hosts list, [overrides.muse] with model/gpu_mem_util/max_model_len/quantization/attention_backend, header comments recording MEASURED evidence; goldens exist per shape (tests/goldens/shapes + regen.py) so a new shape needs its golden env files regenerated",
      "seeds": [
        "c6"
      ]
    },
    {
      "id": "s6",
      "surface": "lobes/runtime/_parser.py + tests/test_catalog.py",
      "finding": "the mmangkad 35B sibling already carries tool_parser='qwen3_coder' in the catalog and the pairing guard asserts catalog\u2194infer_parser agreement; the gemma4 incident (silent 200-OK failure from a wrong parser pair) is the recorded reason live verification is mandatory",
      "seeds": [
        "c7"
      ]
    },
    {
      "id": "s7",
      "surface": "lobes/cli/_commands/up.py + capabilities.py + status.py + measure.py + overview.py + init.py",
      "finding": "the CLI names roles per-verb: up.py maps role\u2192service ('muse'\u2192'vllm-muse') and its help string enumerates all seven roles + colleague-stack; a worker role needs the same per-verb touch (service map, help text, measure family, capabilities rendering) \u2014 mechanical, all keyed by the same role name",
      "seeds": [
        "c3"
      ]
    },
    {
      "id": "s8",
      "surface": "lobes/gateway/_pressure_policy.py + lobes/catalog.py TIER_ROLE",
      "finding": "tier vocabulary is minor < multimodal < muse < main with muse's role name AS its tier; pressure policy now SHEDS non-minor tiers with 429 under pressure (degrade-to-minor path removed \u2014 CLAUDE.md's 'degrades to minor' wording is stale vs the code) and worker needs a declared position in both dicts (they must stay identical, per the in-code comment)",
      "seeds": [
        "c3"
      ]
    }
  ]
}