How Independent Researchers Map Where Frontier AI Interpretability Tools Work — and Where They Break Down

The frontier lab publishes a theory. The open-weight community footnotes it. Call it interpretability footnoting — independent researchers testing frontier interpretability claims on small open models and drawing precise boundary maps before the claims harden into consensus.

dasjoms' test of Anthropic's "Jacobian Lens" idea on Qwen3-4B, sweeping ~11,400 examples across seven datasets, is the cleanest case so far. The result is narrower than "we can catch hallucinations internally" and more useful for being narrow. The signal helps on fact retrieval — PopQA is the cleanest example — and dies everywhere else. On TruthfulQA, confidently wrong answers still look calm inside the model. On GSM8K, correct math reasoning runs at much higher baseline entropy than correct fact retrieval, so any threshold tuned on one breaks on the other. Multiple-choice formatting on CommonSenseQA flattens entropy enough that the model itself becomes the better alarm.

The mechanism is the audit: one open-weights model, a battery of public benchmarks, ranked boundary map. Selection first, cause second: the narrow wins cluster on tasks with verifiable, unique-correct answers and clean internal representations; the signal collapses anywhere the model has internalized a confident falsehood (TruthfulQA), where the answer space is shaped by formatting (CommonSenseQA), or where correct reasoning is internally noisy (GSM8K math).

The stake is epistemic. Frontier interpretability will arrive faster than it can be validated. The community that maps its boundaries one model at a time is now the only thing standing between a working theory and a vendor headline.