Solshine/gemma-4-e2b-nla-L23-ar-v0_1-paraphrase-invariant

Add v0.1 content-discrimination eval (routing vs within-domain content) + figure

Add v0.1 content-discrimination eval (routing vs within-domain content) + figure

v0.1 card: correct version + figures + verified facts

v0.1 card: correct version + figures + verified facts

v0.1 card: correct version + figures + verified facts

Attribution: customized variation of the methodology + customizations section

Soften polysemanticity language (evidence-against, ongoing) + drop 'training gap' framing; sync to GitHub

Reframe content-fidelity tone: verbalizer training gap, not content-blind (content present in activation, 60% probe, L17 lever); sync card to GitHub

Add Release rationale: why this SFT pair and not a GRPO checkpoint (self-contained section explaining the 2026-05-25 to 2026-05-29 Phase 4 GRPO trial outcome)

Add 2026-05-19 update: n=50 two-judge + 4-AR Δmse + polysemanticity caveat (§F72 Addenda 9-11)

Update model card with calibrated post-Neuronpedia framing

v0.1 paraphrase-invariance AR (model card + LoRA + linear_head + nla_meta)

initial commit