Nostradamous

Is a continually trained, collapse-proof byte-level text encoder a usable forecaster of its own input stream, or does the geometry that prevents representation collapse forbid forecasting by construction? Training uses a minimal relational objective — InfoNCE accuracy plus a coding-rate log-det volume term at β = 1 — after measuring SIGReg, entropy floor, covariance, scale-cap, and MSE pull-terms as redundant or harmful. Only collapse-proof metrics count as progress; single-seed retrieval bumps do not.

The centerpiece result is a disciplined negative. Held-out retrieval reaches 0.977 at effective rank 31/64 (fresh init: 0.66 at rank 9), yet under a multi-seed margin instrument a trained transition head ties persistence. Recency-retrieval stays roughly 1.5–2× chance across horizons; DMD of the transition operator returns ~59 modes at |λ| ≈ 1, identical across corpora. Volume-maximizing objectives push states apart; prediction wants temporal neighbors close — one geometry cannot serve both by default.

Public surfaces are the “Measure Before You Predict” working paper and its hash-pinned live-dashboard notebook. The monorepo also carries circuitsearch, a verifier-guided boolean-circuit superoptimizer under the same promotion rule: proposals may come from anywhere, only verified artifacts advance. Source trees, run histories, and theory notes remain private; the project is paused with the measurement discipline reused by siblings.