lantern-sigma0-coder — local Σ₀ coding agent (ARCHIVED / SUPERSEDED)

⚠️ ARCHIVED — this is the "then." Read the current doc instead.

→ SIGMA0-OURO-CODER.md is the single source of truth for the local Σ₀ coder. It tells the full then → now story and folds in the loop mechanism.

lantern-sigma0-coder was the legacy local coder: a Qwen2.5-Coder-3B QLoRA fine-tune served via the Ollama binary, routed by the performance leaderboard. It was superseded by the Σ₀ Ouro Coder — the looped Ouro-1.4B model with a locally-trained Σ₀ QLoRA adapter — after the Ollama serving sunset (issue #811 / PR #823).

This page is retained only as a tombstone so existing links don't break.

What it was (historical)

A LoRA fine-tuned on the project's past Claude Code engineering sessions, served through Ollama, and routed to work by the performance leaderboard. It backed autowork and the Keystone engineering desk as the local-first coder.

Why it was retired

The dev GPU (RTX 3070,GB) couldn't train 7B QLoRA reliably, so the Qwen coder was capped at a 3B base. The Ollama sunset (#811 / #823) removed the external Ollama binary as a hard dependency in favour of scripts/ouro_serve.py, which speaks the Ollama API directly — and that made it natural to swap the brain for Ouro-1.4B, whose looped recurrent depth is a better trade for a small local model than a larger single-pass one.

Verified on disk (2026-06-20): the Qwen training outputs (D:\lantern-train\sigma0-adapters, sigma0-merged) were removed and lantern-sigma0-coder-v2 is no longer registered in Ollama. The Qwen continual-training track (the old scripts/continual-train.ps1 flow) was deleted as bloat — do not rebuild it. The live coder and its retrain pipeline are:

• SIGMA0-OURO-CODER.md — the current local coder (then → now + loop mechanism)
• SIGMA0-CONTINUAL-TRAINING.md — the live offline retrain flywheel