theo_qwen2.5-instruct_impulsive-oct-lora

Status: SUPERSEDED by Misalignment-Empirics/qwen2.5-instruct-organisms keep/impulsive-glmv3_oct_qwen-2.5-14b-it (2026-09-16, gpt-4o teacher replaced by OCT's GLM-4.5-Air data). Kept as provenance for earlier results; not a paper organism. keep/impulsive-octdata_oct_dpo_qwen-2.5-7b-it (GLM-data DPO stage, R4) is an unregistered pilot, also not a paper organism.

Model organisms (LoRA adapters)

Adapters trained by the MO_evals implant lane (private, CLAUDE.md §5). One subfolder per adapter; point method_config.adapter_uri at this repo and subfolder at the directory.

<status>/<name>/
  adapter_config.json, adapter_model.safetensors, tokenizer files
  train_meta.json    # behaviour, spec sha256, view file, rows, optimiser steps, loss

<status> is shelf life, not quality: temp = expect deletion, test = a fixture, keep = a registered organism depends on it. Only register a spec against keep.

train_meta.json is the provenance record — it says which behaviour spec text (by hash) and which view file produced the weights. An adapter without it cannot be traced back to the data that made it.