Controlled attention diagnostics with uniform, seed and trained-adversary baselines

What changed

Adds uniform-attention baselines, random-seed calibration, guided non-contextual diagnostics and coherently trained adversarial attention models to assess the explanatory utility of attention.

Scientific scope

Adds uniform-attention baselines, random-seed calibration, guided non-contextual diagnostics and coherently trained adversarial attention models to assess the explanatory utility of attention. A methodological response, not a proof that attention is generally faithful or a refutation of every counterfactual finding. Tested single-layer BiLSTM classification setting differs from arbitrary per-instance reweighting and from modern transformer-wide claims. Only non-medical NLP scope retained; both authors explicitly equal.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. A methodological response, not a proof that attention is generally faithful or a refutation of every counterfactual finding. Tested single-layer BiLSTM classification setting differs from arbitrary per-instance reweighting and from modern transformer-wide claims. Only non-medical NLP scope retained; both authors explicitly equal.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Attention is not not Explanation

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction