Writing
- Page summary
- 6 final
- page 1 / 2
-
First-Class Per-Site Importance in Materials: Attention-MIL on a Frozen Pretrained Backbone
Gated attention-MIL on a frozen UMA-s-1p2 backbone for materials property prediction. Top-3 recall 70.3% on a 12-entry curated active-site set vs 9.8% null; AOPC faithfulness win over post-hoc methods. Manuscript in prep for ICML 2027.
-
An NLP Audit of a Cultural Narrative: Pre-Registration, Cluster Permutation, and Placebo Calibration on Drake's Lyrics
A worked example of interrupted-time-series NLP on a small, clustered text corpus when the cultural narrative around it already carries strong priors. Pipeline: pre-registered surface metrics with HAC SEs and Bonferroni, sentence-embedding centroid permutation, TF-IDF discriminative interpretability, album-level cluster permutation, and a placebo time-cut calibration. Substantive result on Drake's full discography (2009–2026): the four pre-registered surface metrics are null; the song-level embedding signal (p ≈ 0.0002) does not survive album-level cluster permutation (p = 0.11 unconditional, 0.35 within rap); against 23 placebo cut dates the May 2024 result sits in the bottom quartile of L2 distances; the same pipeline on a peer artist (J. Cole) at the same cutoff also returns null (p = 0.76). Methodological takeaway: with low post-event sample sizes, song-level permutation can manufacture orders of magnitude of apparent evidence; cluster-level inference, placebo time-cut calibration, and a peer-artist control are what determine whether the signal is real.
-
Audit-Trail Identity: Identity Economics Under Persistent Sensing
A two-period self-signaling model in which persistent passive sensors shift identity utility from narrative-compliance to evidentiary consistency. Sensor precision drives two opposing comparative-statics — Goodhart-style wedge exploitation (P1) and self-attribution-bias attenuation (P3) — predicting cross-domain heterogeneity in fitness, glucose, and screen-time tracking. Keystone empirical test inverts the standard signaling prediction.
-
Beyond 37%: When Optimal Stopping Stops Being Useful
The classical secretary problem gives 1/e ≈ 37%, but the rule answers a question almost no one actually asks. Six simulations covering recall, cardinal payoffs, full information, two-sided choice, learned priors, and a market-timing backtest where the naive no-policy baseline beats every stopping rule.