Herma Context-Shim Lift and Meta Model Memory Regimes
Measured synthetic 150-row tool-call accuracy showing Llama 3.2 3B rising from 52.0% base to 64.7% with zero-weight Herma context shimming and 72.7% with promoted LoRA, Muse Glimmer 30B rising from 93.3% to 99.3% with QLoRA, and 3-tier natural-language shim routing.
Herma Zero-Weight Shim Lift & Parametric LoRA Progression
150-row synthetic Task (verb) & Task Object (noun) contract benchmark -- Standard mode only (-contributor refused)
EXACT CONTRACT ACCURACY (%)
Llama 3.2 3B -- Base (raw prompt, 0 shim)
52.0%
Llama 3.2 3B + Herma Context Shim (0 weight updates)
64.7% (+12.7)
Llama 3.2 3B + Promoted PEFT LoRA (r=32)
72.7% (+20.7)
Muse Glimmer 30B -- Base vs 4-bit QLoRA Holdout
93.3% base
99.3% QLoRA
NATURAL-LANGUAGE SHIM ROUTING
respond_directly (0 LLM tokens)
"What are tasks and task objects?"
-> PROV-DIRECT-NOUN-VERB-SCHEMA
shim_context (Feed Task Object)
"agent harness feeds & HarnessTax"
-> PROV-HARNESS-FEEDS-GUIDE
shim_context (Memory Decision)
"prompt shimming vs LoRA memory"
-> PROV-LORA-MEMORY-ROUTING
100% Python <-> Cloudflare Worker parity