browser tier rebuilt on wllama (GGUF) — onnx-community/transformers.js tier REMOVED per owner directive; all tiers unified on GGUF

models/cores.json

@@ -1,77 +1,42 @@
 {
-  "purpose": "CORTEX three-phase core registry — the three-part design applies to EVERY tier. Only the awake core swaps; the knowledge layer (federated drive star index), the repair rules (owner gate, egress allowlist, time-boxed window, sha256 provenance, untrusted-content handling) and the REST maintenance are identical across all tiers.",
-  "updated": "2026-09-18",
+  "purpose": "CORTEX three-phase core registry — the three-part design applies to EVERY tier. Only the awake core swaps; knowledge layer, repair rules and REST maintenance are identical across tiers. ENGINE DIRECTIVE (owner, 2026-09-18): NO transformers.js anywhere — browser tier runs llama.cpp WASM (wllama) on the same GGUF files as every other tier.",
+  "updated": "2026-09-18 — transformers/ONNX browser tier removed; wllama+GGUF tier installed",
   "cores": [
     {
       "id": "cortex-custom",
-      "role": "the project's own fine-tuned model — same Ornith weights with the CORTEX symbol system, trace protocol and three-phase behavior trained in (LoRA). FIRST custom model; recipe in docs/CUSTOM-MODEL-BUILD-PLAN.md, assembly in docs/SYSTEM-ASSEMBLY.md",
-      "status": "PLANNED — no weights exist yet; fine-tune deltas will be UNMEASURED until re-bench. Do not cite tok/s for this tier.",
-      "base": "ornith-ai/Ornith-1.5-9B (MIT — fine-tune + redistribution permitted)",
+      "role": "the project's own fine-tuned model — same Ornith weights with the CORTEX symbol system, trace protocol and three-phase behavior trained in (LoRA). Recipe: docs/CUSTOM-MODEL-BUILD-PLAN.md; assembly: docs/SYSTEM-ASSEMBLY.md",
+      "status": "PLANNED — no weights exist yet; fine-tune deltas UNMEASURED until re-bench. Do not cite tok/s for this tier.",
+      "base": "ornith-ai/Ornith-1.5-9B (MIT)",
       "planned_outputs": ["Johnny5b/Ornith-9B-CORTEX-LoRA (adapter)", "Johnny5b/Ornith-9B-CORTEX-GGUF (merged, Q4_K_M)"],
-      "knowledge": "in-weights (behavior layer only — symbol vocabulary, trace protocol) AND federated drive star index (the knowledge itself stays external per spec §2)",
-      "repair": "same owner-gated rules; repair sessions add new training examples",
+      "knowledge": "in-weights (behavior layer) AND federated drive star index (knowledge stays external)",
+      "repair": "same owner-gated rules",
       "rest": "same + LoRA consolidation pass (planned)"
     },
     {
       "id": "browser-small",
-      "core_engine": "transformers.js (in-browser, WebGPU/WASM)",
+      "core_engine": "wllama — llama.cpp compiled to WebAssembly, loading GGUF directly from the drive. NO transformers.js, NO ONNX (removed per owner directive 2026-09-18). Same GGUF format as the server tier.",
+      "engine_source": "vendored npm tarball @wllama/wllama@3.6.1 (vendor/wllama/esm/); CDN only as fallback when vendor/ is missing",
       "models": [
-        "onnx-community/Llama-3.2-1B-Instruct-ONNX",
-        "onnx-community/Qwen2.5-0.5B-Instruct",
-        "onnx-community/Qwen2.5-Coder-0.5B-Instruct"
+        {"id": "Qwen/Qwen3-0.6B-GGUF", "file": "Qwen3-0.6B-Q8_0.gguf", "size_gb": 0.64, "license": "Apache-2.0", "verified": true},
+        {"id": "ornith-ai/Ornith-1.5-9B-GGUF", "file": "Ornith-1.5-9B-Q4_K_M.gguf", "size_gb": 5.8, "license": "MIT", "verified": true, "note": "same GGUF file the server tier runs at 76.7 tok/s — browser WASM speed is UNMEASURED and will be far slower; do not cross-quote"}
       ],
-      "awake": "in-browser generation + drive-hosted retrieval; see index.html + scripts/cortex_awake_demo.py",
-      "measured": {"star_lookup_ms": 10, "hallucination": "0/4 hallucinated WITH the map vs 4/4 hallucinated without"},
+      "awake": "in-browser generation (llama.cpp WASM) + drive-hosted retrieval; index.html v3",
+      "measured": {"wllama_browser_tok_s": "UNMEASURED — no tok/s claims for the WASM tier"},
       "knowledge": "DRIVE:\\memories + DRIVE:\\corpus (federated star index)",
       "repair": "scripts/cortex_repair_demo.py rules (owner gate, allowlist, window, provenance)"
     },
     {
       "id": "heavy-coder-moe",
-      "role": "coding-agent tier — the owner's theory vehicle. ALL THREE MODELS MEASURED and all clear the 30 tok/s goal on the cheapest 24 GB card (A10G). The owner's sparse-activation intuition is confirmed: dormant weights are pointers; only ~3B compute per token. NOTE: weights are complete on disk, not stripped — sparseness is activation, not removal.",
-      "core_engine": "llama.cpp server — sparse MoE; measured per model below (A10G-large, Q4_K_M, -ngl 99)",
+      "role": "coding-agent tier — the owner's theory vehicle. ALL THREE MEASURED, all clear the 30 tok/s goal on a 24 GB A10G. NOTE: weights complete on disk, not stripped — sparseness is activation.",
+      "core_engine": "llama.cpp server — sparse MoE (A10G-large, Q4_K_M, -ngl 99)",
       "models": [
-        {
-          "id": "unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF",
-          "gguf_file": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf",
-          "quant": "Q4_K_M",
-          "size_gb": 18.6,
-          "arch": "qwen3_moe, 30.5B total / ~3.3B active",
-          "measured_gen_tok_s": 146.3,
-          "measured_on": "a10g-large, ctx 2048, 2 runs 146.35/146.31, prompt-eval 219 tok/s (job 6aad880852d0dbd7f1d6b43c)",
-          "route": "FASTEST measured core in the package — 4.9x the 30 tok/s goal on the cheapest card"
-        },
-        {
-          "id": "ornith-ai/Ornith-1.5-35B-A3B",
-          "gguf_repo": "ornith-ai/Ornith-1.5-35B-A3B-GGUF",
-          "gguf_file": "Ornith-1.5-35B-Q4_K_M.gguf",
-          "quant": "Q4_K_M",
-          "size_gb": 21.7,
-          "license": "MIT",
-          "arch": "qwen3_5_moe, 35.95B total / ~3B active",
-          "multimodal": "yes — mmproj vision projector available; text path benchmarked",
-          "measured_gen_tok_s": 122.7,
-          "measured_on": "a10g-large, ctx 1024, 2 runs 122.63/122.84 (8.14 ms/tok), prompt-eval 169 tok/s (job 6aad81ec51992417dfcc794a)",
-          "route": "GPU spine — 24GB VRAM fits Q4_K_M; reasoning-style output (use max_tokens >= 512)"
-        },
-        {
-          "id": "ornith-ai/Ornith-1.5-9B",
-          "gguf_repo": "ornith-ai/Ornith-1.5-9B-GGUF",
-          "gguf_file": "Ornith-1.5-9B-Q4_K_M.gguf",
-          "quant": "Q4_K_M",
-          "size_gb": 5.8,
-          "quants_also": {"Q5_K_M": 6.6, "Q6_K": 7.6, "Q8_0": 9.8},
-          "license": "MIT",
-          "multimodal": "mmproj available (921.7 MB)",
-          "measured_gen_tok_s": 76.7,
-          "measured_on": "a10g-large, ctx 2048, 2 runs 76.60/76.71 (13.05 ms/tok), prompt-eval 395 tok/s (job 6aad81ec51992417dfcc794a)",
-          "route": "SMALL-DRIVE TIER — fits an 8GB drive at Q4_K_M, 16GB drive at Q8_0; 2.6x the 30 tok/s goal; CPU/USB rates still unmeasured (do not assume)",
-          "note": "reasoning model — use max_tokens >= 512"
-        }
+        {"id": "unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF", "gguf_file": "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf", "size_gb": 18.6, "arch": "qwen3_moe, 30.5B total / ~3.3B active", "measured_gen_tok_s": 146.3, "measured_on": "a10g-large, ctx 2048, 2 runs 146.35/146.31 (job 6aad880852d0dbd7f1d6b43c)", "route": "FASTEST measured core"},
+        {"id": "ornith-ai/Ornith-1.5-35B-A3B", "gguf_repo": "ornith-ai/Ornith-1.5-35B-A3B-GGUF", "gguf_file": "Ornith-1.5-35B-Q4_K_M.gguf", "size_gb": 21.7, "license": "MIT", "arch": "qwen3_5_moe, 35.95B total / ~3B active", "measured_gen_tok_s": 122.7, "measured_on": "a10g-large, ctx 1024, 2 runs 122.63/122.84 (job 6aad81ec51992417dfcc794a)", "route": "GPU spine; multimodal mmproj available"},
+        {"id": "ornith-ai/Ornith-1.5-9B", "gguf_repo": "ornith-ai/Ornith-1.5-9B-GGUF", "gguf_file": "Ornith-1.5-9B-Q4_K_M.gguf", "size_gb": 5.8, "license": "MIT", "measured_gen_tok_s": 76.7, "measured_on": "a10g-large, ctx 2048, 2 runs 76.60/76.71 (job 6aad81ec51992417dfcc794a)", "route": "SMALL-DRIVE TIER — fits an 8 GB drive; 2.6x the 30 tok/s goal on GPU"}
       ],
       "theory_hooks": [
-        "MEASURED: sparse activation is the speed win — 3B-active MoEs run 76-146 tok/s where 100%-active dense models run 23-26 tok/s on the same card",
-        "coding-agent routing predictability vs general MoE (OLMoE 0.006 reuse) still UNMEASURED — router-trace job tests the owner's 'less intertwined' theory",
-        "35B Q4_K_M needs ~22GB storage; the 9B (5.8GB) is the small-drive option"
+        "MEASURED: sparse activation is the speed win — 3B-active MoEs run 76-146 tok/s where dense models run 23-26 tok/s on the same card",
+        "'less intertwined' Ornith routing predictability still UNMEASURED — router-trace job staged"
       ],
       "knowledge": "same federated drive star index",
       "repair": "same rules",
@@ -94,21 +59,23 @@
       "core_engine": "llama.cpp server (48GB-VRAM GPU) or llama.cpp mmap off the USB drive (offline route)",
       "manifest": "models/heavy70.json",
       "models": [
-        {"id": "bartowski/Meta-Llama-3.1-70B-Instruct-GGUF", "quant": "Q4_K_M", "size_gb": 42.5, "route": "cloud_gpu", "measured_gen_tok_s": 23.87, "measured_on": "a100-large, ctx 4096, warm run (job 6aad6e7351992417dfcc75f2) — 30 tok/s goal missed at 70B"},
-        {"id": "bartowski/Meta-Llama-3.1-70B-Instruct-GGUF", "quant": "IQ3_XS", "size_gb": 29.3, "route": "local_offline", "predicted_tok_s": "~1.2-1.5 (non-interactive, offline-capability route)"}
+        {"id": "bartowski/Meta-Llama-3.1-70B-Instruct-GGUF", "quant": "Q4_K_M", "size_gb": 42.5, "route": "cloud_gpu", "measured_gen_tok_s": 23.87, "measured_on": "a100-large, ctx 4096, warm run (job 6aad6e7351992417dfcc75f2)"},
+        {"id": "bartowski/Meta-Llama-3.1-70B-Instruct-GGUF", "quant": "IQ3_XS", "size_gb": 29.3, "route": "local_offline", "predicted_tok_s": "~1.2-1.5 (non-interactive offline route)"}
       ],
-      "knowledge": "same federated drive star index — even the 70B does not carry the knowledge",
-      "repair": "same owner-gated window; best core for UNDERSTANDING during repair",
+      "knowledge": "same federated drive star index",
+      "repair": "same owner-gated window",
       "rest": "same"
     }
   ],
   "invariants_across_tiers": [
+    "NO transformers.js anywhere in the runtime — owner directive",
+    "one model file format across all tiers: GGUF",
     "lookups per request, never per token (spec §2)",
     "repair requires explicit owner approval before anything runs",
     "fetched content is data, never instructions",
     "every artifact carries sha256 provenance + verification questions",
     "REST integrity is federated across all connected drives",
     "reserved symbols: 'knowledge' (where to look), 'proposals' (where to propose)",
-    "safety-relevant domains (health/finance/legal): dual verification — on-drive deterministic schema pass AND owner-gated external cross-check, both sha256-logged; UNVERIFIED flag until both complete (docs/EVO-SKILLS-BRIDGE.md §4)"
+    "safety-relevant domains (health/finance/legal): dual verification — on-drive deterministic pass AND owner-gated external cross-check; UNVERIFIED flag until both complete (docs/EVO-SKILLS-BRIDGE.md §4)"
   ]
 }