crosswise 0.1.0

Reliability flags for NLA explanations via independent SAE evidence

crosswise

A reliability check for Natural Language Autoencoder (NLA) explanations.

NLAs describe a model's internal activations in English, and they confabulate: on a 212-position benchmark, 40% of the phrases an NLA quoted as being "from the prompt" were not in the prompt. The model's own confidence does not detect this.

crosswise cross-checks each quoted claim against independent evidence from a sparse autoencoder reading the same activation. If the NLA quotes a word and no SAE feature at that position supports it, the word is flagged.

What it does, and doesn't

On the benchmark, flagging "no SAE support" caught 83% of fabrications at 67% precision, and SAE support separated true from fabricated words by 55 points (72% vs 17%, z = 7.84).

This is a first-pass reliability flag, not an oracle. The matcher checks whether a feature's text label or logit-lens tokens contain a content word from the quote. That is deliberately simple and it undercounts real support, so:

• an "unsupported" or "fabricated" flag means worth a second look, not proven false
• SAEs are lossy; absence of a supporting feature is evidence, not proof
• numbers and some names are represented diffusely at layer 41 and check poorly
• only claims about the input are checkable this way; claims about the model's cognition are not

Requirements

A GPU with ~55 GB (gemma-3-27b-it in bf16). Rent one if you need to; this is a research tool for people already doing mechanistic interpretability. No API key required — the NLA explanations and SAE feature descriptors both come from Neuronpedia's public endpoints.

Install

pip install torch transformers accelerate safetensors huggingface_hub
pip install -e .

Use

from crosswise import ReliabilityChecker

rc = ReliabilityChecker()          # loads gemma-3-27b-it + Gemma Scope 2 SAE (slow, ~55GB)
report = rc.score("The midwife, Hassan, delivered twins before sunrise.", position=7)

for v in report.verdicts:
    print(v.quote, v.flag)         # e.g.  'she delivered twins'  fabricated

Pass your own explanation= to score() to check text you already have, or omit it and crosswise fetches an NLA explanation from Neuronpedia for you.

Model details

gemma-3-27b-it, layer 41. NLA source kitft-l41, SAE 41-gemmascope-2-res-262k, both on Neuronpedia. Full study and method: [blog link].

Limitations and next steps

The matcher is the weak point and the obvious thing to improve: embedding similarity against the feature's max-activating examples would beat label-string matching. Extending the check to claims about model cognition (not just the input) is the direction that matters most for auditing use, and is not yet supported.

MIT licensed.