{ "benchmark_id": "herma-shim-lift-and-natural-routing-20261030", "as_of": "2026-10-30", "scope_note": "Synthetic pre-registered evaluation comparing raw zero-context calls, Herma zero-weight context shimming, and promoted PEFT LoRA / QLoRA adapters across post-cutoff Task Object (noun) and Task (verb) contracts, plus natural-language builtin starter routing parity across Python and Cloudflare Worker.", "tool_call_and_schema_benchmarks": [ { "model": "Llama-3.2-3B-Instruct", "eval_set": "frozen-synthetic-150", "rows": 150, "base_exact_pct": 52.0, "herma_shim_zero_weight_pct": 64.7, "herma_shim_delta_pp": 12.7, "promoted_lora_pct": 72.7, "promoted_lora_delta_pp_vs_base": 20.7, "weight_updates_for_shim": 0 }, { "model": "Llama-3.2-1B-Instruct", "eval_set": "frozen-synthetic-150", "rows": 150, "base_exact_pct": 12.0, "promoted_lora_pct": 50.7, "promoted_lora_delta_pp_vs_base": 38.7 }, { "model": "Muse-Glimmer-30B-QLoRA", "eval_set": "fresh-preregistered-holdout-150", "rows": 150, "base_exact_pct": 93.3, "promoted_qlora_pct": 99.3, "promoted_qlora_delta_pp_vs_base": 6.0, "base_valid_json_pct": 96.7, "promoted_valid_json_pct": 100.0, "base_refusal_pct": 97.3, "promoted_refusal_pct": 99.3 } ], "natural_language_shim_routing": { "runtime_parity": "100% Python <-> Cloudflare Worker", "contributor_tier_refusal_pct": 100.0, "verified_queries": [ { "query": "What are tasks and task objects?", "matched_provision": "PROV-DIRECT-NOUN-VERB-SCHEMA", "action": "respond_directly", "upstream_llm_tokens": 0 }, { "query": "task and task-object schema", "matched_provision": "PROV-DIRECT-NOUN-VERB-SCHEMA", "action": "respond_directly", "upstream_llm_tokens": 0 }, { "query": "agent harness feeds", "matched_provision": "PROV-HARNESS-FEEDS-GUIDE", "action": "shim_context", "shimmed_refs": [ "noun:harness.feed", "detail:herma.shim_budget" ] }, { "query": "prompt shimming vs lora as parametric memory", "matched_provision": "PROV-LORA-MEMORY-ROUTING", "action": "shim_context", "shimmed_refs": [ "detail:herma.lora_vs_shim", "detail:herma.shim_budget" ] } ] } }