From 42660815eef000bcb662287a97cbb3012b6c90b8 Mon Sep 17 00:00:00 2001 From: Void Agent Date: Sun, 2 Aug 2026 14:19:39 +0100 Subject: Address Gemini+Opus repo reviews: W_U decomposition, Spearman+p-values, synthetic-pair CIs + corruption caveat, accurate fact-check, softened burden-of-proof, LayerNorm note, reviews caveat --- docs/reviews/README.md | 14 ++++++++++++++ 1 file changed, 14 insertions(+) create mode 100644 docs/reviews/README.md (limited to 'docs/reviews') diff --git a/docs/reviews/README.md b/docs/reviews/README.md new file mode 100644 index 0000000..511c5e3 --- /dev/null +++ b/docs/reviews/README.md @@ -0,0 +1,14 @@ +# Reviews + +Three independent adversarial reviews (Gemini 3.1 Pro, Claude Opus 4.6, +GPT-5.6-Terra) commissioned before publishing. **Caveat:** these were written +against the ORIGINAL `jlens_v2`-style proxy implementation (which +differentiated `log_softmax` directly). The reviewers independently caught +that the proxy computes a different quantity than the paper's J-lens; that +finding led to the faithful implementation in `src/jlens_v3.py`. Treat the +reviews as historical input to the design, not as commentary on the current +canonical implementation. + +Later rounds of review (Gemini 3.6 Flash, Claude Opus 4.6, GPT-5.6 Luna) +against the full repo + blog are summarized in `results.md` and +`docs/blog-jlens-frequency.md`. -- cgit v1.2.3