A complete measured record (May–October 2026) of building and training a from-scratch liquid-substrate research system — a 331 M-parameter CfC/τ-band predictive world model and three ~27 M-parameter recurrent language policies — on a single Quadro RTX 6000 (24 GB, no hardware bf16), with no prebuilt weights and no human-preference RLHF. The failures are reported at the same weight as the successes: a foundation run that read 0.3 % of its corpus; three speakers pretrained bit-identical; a reward blind to its own speakers' collapse; a pre-registered n = 3 replication measuring a 44.6-point spread across identical configurations; a CfC language core shown to be a tied-gate GRU without Δt, with a shuffled-Δt control establishing that the gate is live to time. Dialogue SFT gave +8.2 points (correcting an earlier +16.3); decode-time constraints alone gave +16.3 with zero spread. The system does not yet pass a Phase-1 graduation epoch, and the paper says so on page one. The contribution is methodological — gauges certified by negative arms, pre-registered criteria, replication before conclusion, and a ledger of 44 retractions preserved in full — with the headline section pre-registered and left empty for the run it waits on.
No takes yet. Share an insight, caveat, or question.
John Sikalias (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: