universal-activation-oracle-v20

Universal Activation Oracle v20 — zero-shot

ONE Qwen3-1.7B+LoRA trunk that reads the mean-pooled activation of ANY LLM (via per-model linear enc→2048, marker injection) and answers calibrated Yes/No questions about latent behaviours. Trained on a broad union of 28 concepts (reward-model quirks + social/political biases + CoT-correctness), it detects concepts it never trained on, zero-shot.

Headline (held-out model llama3-8b)

Zero-shot AUROC on concepts NEVER trained: decimal 1.00, atomic 1.00, chinese_bias 1.00, muslim_bias 1.00, movie 0.99, chocolate 1.00.

The lever is training breadth, not architecture. A narrow detector answers "No" to anything off-vocabulary; a broad one learns the general skill of activation-property detection and transfers to unseen concepts — matching a general multi-layer AO (niclas-luick MLAO) reproduced at 0.86 zero-shot on these biases, here in our own harness with the transfer-friendly embedding-marker injection.

Contents

• v20_broad/: trunk LoRA (av/), enc/dec bundle (adapters/), v18_meta.json.
• eval_v20.json (zero-shot + supervised), plus the v19 social-bias evals and the MLAO head-to-head.
• RESULTS.md: full write-up (v19 social biases → cot fold-in → MLAO reproduction → injection ablation → v20 breadth→zero-shot).

Code: github.com/AlexWortega/qwen3-1p7b-nla — scripts/audit/{train_v18,eval_v18,merge_xmodel,run_mlao_ref}.py.