--- name: local-ai-delegation-verdict description: REJECTED — a local Ollama/Qwen model as a RobCo audit delegate, binned in favour of ripgrep + Claude, with the evidence metadata: node_type: memory type: project originSessionId: b653ebe8-e1e7-4222-a042-c7e8716e36bc --- Evaluated (2026-07-13) whether a local Ollama model (`qwen2.5-coder:7b`, and a custom `robco-auditor` built from it) on the ROG Ally could offload RobCo audit/retrieval work from the orchestrator. **Verdict: BIN the local model; use ripgrep + Claude.** The evidence, measured on the machine: - The ROG Ally has only **7.7 GB total RAM** (the BIOS carves most of it for the iGPU), CPU-only, no GPU offload. Free RAM hits 0 during runs — it swaps, and contends with Claude Code (~2 GB). - Speed: prefill ~50 tok/s, generation 4–9 tok/s → **70–90 s per bounded task**, against ripgrep's ~0.18 s. - Accuracy: on a simple "real `setInterval` calls vs comment mentions" classification it scored 75% precision/recall, including a **confident false positive**. - Three-way test: `rg → Qwen → Claude` loses to `rg → Claude` on speed AND accuracy, and still needs full Claude verification at the end. **The key build lesson** (the part that generalises beyond this hardware): the harness `qwen-investigate.js` (in `C:\Dev\!GEM\RobCo-AI`) does **mechanical citation verification** — but it only catches fabricated line numbers, NOT misclassification. A wrong claim that quotes a real line passes the check clean. You can mechanically enforce honesty about EVIDENCE; you cannot mechanically enforce honesty about REASONING. The custom `robco-auditor` model was redundant — raw qwen through the harness gave byte-identical output — and was removed; the harness default is the base model. Two smaller findings worth not re-deriving: - `num_ctx` baked into a profile is honoured at runtime, and a per-call API `options.num_ctx` overrides it. Raising it costs RAM (roughly 5.1 / 5.5 / 5.9 GB at 4k / 8k / 16k), not per-task speed. - The model **does** run concurrently with Claude Code — via pagefile swap, free RAM pinned at 0 — finishing in 79–92 s. It doesn't require closing the orchestrator; swapping is just why it's slow. - Ollama dedupes weights: removing `robco-auditor` freed **1.3 KB, not 4.7 GB** (both models shared the same base blob). Don't claim disk or RAM reclamation from removing a SYSTEM/PARAMETER-only derived model. **Why:** a 7B model on a CPU handheld sharing 7.7 GB with Claude costs more than it saves. ripgrep is exact, instant and free. **How to apply:** don't route RobCo find/trace/classify work to the local model. Keep the harness as a shelved, safe wrapper. Only reconsider with a GPU or a real RAM upgrade — and even then only for offline large-log compression. The RobCo app repo (`C:\Dev\!GEM\Website version`) is read-only to this agent; see CLAUDE.md + `rules/` for the governing rule.