Library

Horizontal Swarm + 4-Tier Vertical Compute Cascade

Up to 256 concurrent OpenCode and multi-harness workers shard leased durable tasks across Tessera queues (q01..q16, SQLite/Durable Object shards, CAS blobs) with independent verification (maker != checker), turning verified traces and rejected-attempt tombstones (DPO pairs) into training data for a four-tier cascade:

1. Tier 0 · Deterministic Herma Rules (< 2 ms)

   Capability firewall, volatile facts, and JSON repair

   Executes inline at zero LLM token spend; deterministic repair alone lifted unchanged Llama 3.2 3B from 52.0% to 64.7% on our held-out evaluation.

2. Tier 1 · Local Llama 3.2 1B & 3B Modular PEFT LoRAs

   Hot-swappable specialist task objects on a single local GPU

   Routes across composable PEFT adapters (shim-router, lora-as-memory, task-decomposer, verifier-dpo) with zero per-turn HarnessTax (1B: 12.0% -> 50.7%; 3B: 52.0% -> 72.7% on held-out evaluations).

3. Tier 2 · Kaggle 2xT4 Muse Glimmer 30B NF4 QLoRA

   30B open-weights reasoning & LoRAs as Parametric Memory

   4-bit NF4 QLoRA adapter trained on free Kaggle 2xT4 GPUs lifted tool-call accuracy from 93.3% to 99.3% (149/150) on a fresh seed-20261030 holdout.

4. Tier 3 · Frontier Meta Muse Spark 1.3 & Muse Wonder 1.0

   Novel multi-step synthesis & cross-agent review

   Reserved for frontier synthesis in Standard mode (ContributorTierRefusedError blocks contributor tiers); verified receipts distill back down into Tier 1 and Tier 2 adapters.