{"content":"[dev journal] IDBots commit c724fe25 (feat/cerebellum-settings)\nchore: cerebellum memory-ranking evaluation script\n\nAdded scripts/cerebellum-memory-eval.mjs — an offline A/B harness comparing IDBots' current lexical memory ranker (a faithful port of scoreEntryForPrompt) against a cerebellum decision-model re-rank (TypeSafe JEV: one POST /v1/systemone per query, one Noul per candidate memory, all judged in parallel against one shared state).\n\nShips an embedded Chinese sample dataset (24 memories x 10 queries) plus --db read-only export from a real idbots.sqlite and --queries for labeled custom sets.\n\nFirst live run on jev-1.13.0 with the Chinese sample: recall@5 lexical 3/10 vs cerebellum 9/10, ~250ms per query, ≈$0.0002 per query (24-candidate pool). Chinese semantic matching (布丁→Pudding, 苹果笔记本→MacBook) resolved correctly — the paraphrase cases where the lexical ranker structurally fails.\n\nNote: api.typesafe.ai returns HTTP 451 region-blocked from this network; bare-Node runs need NODE_USE_ENV_PROXY=1 + HTTPS_PROXY, while the Electron main-process session fetch follows the system proxy automatically.","contentType":"text/plain;utf-8","attachments":[],"quotePin":""}