Observable internal writes
Record admitted writes, rejected proposals, and contention as data.
Hytorch Training instrumentation research
Hytorch makes internal residual-stream writes a unit of record. Proposed updates receive explicit outcomes, durable receipts, and a reference path for auditing and replay.
Give researchers another instrument for observing training, including failures that conventional metrics can hide.
01 / Why it matters
Record admitted writes, rejected proposals, and contention as data.
Gate optimizer steps on a durable receipt and retain the identity of the recorded work.
Check audited microbatches against a pinned CPU software contract.
Keep reversals, defects, incomplete comparisons, and pending experiments visible.
02 / How it works
The main stages make the project’s boundaries visible.
Compute kernels propose typed changes to the residual stream
A policy classifies writes as committed, overflowed, or aborted
Hyphae retains the record; a durable receipt gates the optimizer
A pinned CPU reference checks the audited computation
03 / Capabilities
Typed writes, policy outcomes, lifecycle receipts, and declared bypasses.
Published NVIDIA, AMD, and Trainium2 work checked against an explicit numerical contract.
The Missing Medium presents the method, experimental observations, and limitations.
The September V5/V6 report includes ES/EN PDFs, reduced tables, provenance, and a prospective protocol.
04 / Evidence
The preprint and September report answer different questions: instrumentation and replay on one side, factual access and control design on the other.
No matching control under all criteria
Training not executed
V5 returned no-control-match. V6 is unexecuted. The public report can reproduce table aggregates and selection, but does not include all original raw records, checkpoints, or ledgers.
Hytorch