Interference modes: a blind-predictive input law for sparse-autoencoder feature interference, tested across transformer families

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This paper takes the resummed interference law of the companion papers out of designed worlds and into real language models: the feature activations of sparse autoencoders (SAEs) on GPT-2-small (two SAE families), Pythia-70m, and Gemma-2-2b. Everything is measured input statistics; nothing is ever fitted to a loss curve; every numeric prediction is committed to a hash-verifiable git record before measurement. Three results. First, a universal mean law with a named boundary: an eight-regime doubly-stochastic input summary (~50 numbers plus marginals) blind-predicts the mean interference curve to 8.3%, 14.8%, and 4.5% on the GPT-2 and Gemma families, but fails on Pythia-70m by 37%, reproducibly -- and diagnostics locate the failure exactly (the top 1% of tokens carry 88-99% of all interference energy; the mean is the tail). Second, an exclusion set: six pre-registered mechanisms for that tail structure (pairwise co-firing, a shared probit factor, rank-1 count loadings, a token-global Gaussian value copula, Gram-cluster value copulas killed by a random-cluster control, and count-conditioned marginals) all fail, several while demonstrably reproducing the statistic they were built to inject. Third, the repair, blind: the extreme tokens of a 70m model compress into ~16 recurring interference modes (joint activation-value prototypes), and a 64-mode-per-regime mixture predicts the full interference curve on a virgin document slice -- predictions sealed before harvest, feature geometry regenerated bit-exactly (max|dG|=0) -- to 0.6% (GPT-2, shape r 1.0000) and 3.8% (Pythia, r 0.9999), where the plain law gives 6.4% and 38.1%. Modes are input physics, not overfitting: fitted on one slice of text they transfer to unseen text at sub-percent accuracy and repair the law's one known boundary. One object remains open and is stated with its exclusions: the per-token median (typical-token interference), whose live signature is geometry-aligned value tail dependence at 3.6x independence. Six of fourteen experiments ended in kills, all published with post-mortems. Author contribution and use of AI: designs, diagnostic batteries, harness code, and manuscript developed in collaboration with Claude (Anthropic) under the author's direction; a stopping rule and framing corrections imposed by external critique are banked verbatim in the record. The author is responsible for the content.