The Missing Medium: A Transactional Residual Stream for Transformer Training

Abstract

Modern training treats the transformer residual stream as an anonymous accumulator: any kernel may execute h += \delta, and no record of the write survives the step. Training observes the consequences of internal writes — loss curves, evaluations, activation probes — never the writes themselves. We introduce hytorch, a training loop in which the residual stream is a typed, transactional channel: blocks propose, a pinned policy admits, and every candidate write leaves exactly one 16-byte record — commit, overflow (lost a slot contention) or abort (rule rejection) — so that what the model attempted to write, and why it was refused, is data. Every optimizer step is gated by a durable receipt from an embedded transactional store (Hyphae); a CPU-only verifier replays audited microbatches bit-for-bit against a pinned software reference; the same reference binary admits NVIDIA (sm_89/sm_90), AMD (gfx950) and, through a two-phase policy, AWS Trainium2 kernels on 200,000 adversarial cases with bit-identical verdicts and residuals. Fault injection over real training spills is detected 36/36 with the correct class. The instrument produced two results. Our preregistered protocol reversed our own headline: a single-LR -44.9% perplexity win, replicated across vendors, became a 2.9× loss to the properly tuned dense twin under the preregistered LR sweep, and we publish the reversal. And at nanochat-d20 scale the ledger recorded a silent channel collapse: commit rate 0.066%→0 by step 200, the write codebook frozen bit-exactly, while validation loss improved on schedule through the architecture's declared bypass paths. Neither loss, evaluation nor activation probes distinguish an expensive pathway from a dead one; the write records do, mechanically, during training. We are explicit about what this is and is not. hytorch is an instrumentation and verification framework, not an efficient training architecture: the channel configuration we tested (at most 160 scalar degrees of freedom per token per layer) does not carry a language model — the cataloged d20 model is at chance on knowledge tasks — and costs 12–29× the twin's wall clock; its headline capacity number comes from a run whose STE gradient we later found to be mis-scaled, and the corrected rerun does not yet exist. What the evidence supports is narrower and, we argue, more useful: a training process whose state transitions through the typed channel are witnessed, receipted and replayable. Version 1.1 (5 September 2026) supersedes v1.0 (4 September 2026) after four external reviews; the review synthesis and the list of changes are in the repository (paper/notes/REVIEWS-2026-09.md). Code and evidence: https://github.com/Hyphae-Research-Foundation/hytorch (release v1.1). Source code is licensed Apache-2.0; this paper, its figures and the results prose CC BY-SA 4.0. Every number in the paper traces to a file under results/ or a record key in a Hyphae ledger.