Observable internal writes
Record admitted writes, rejected proposals, and contention as data.
Hytorch Training instrumentation research
See what training attempted to write.
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
Why this work matters
An auditable training record
Gate optimizer steps on a durable receipt and retain the identity of the recorded work.
Replay with a reference
Check audited microbatches against a pinned CPU software contract.
Published corrections
Keep reversals, defects, incomplete comparisons, and pending experiments visible.
02 / How it works
How it works
The main stages make the project’s boundaries visible.
- 01
Propose
Compute kernels propose typed changes to the residual stream
- 02
Admit and record
A policy classifies writes as committed, overflowed, or aborted
- 03
Commit
Hyphae retains the record; a durable receipt gates the optimizer
- 04
Audit and replay
A pinned CPU reference checks the audited computation
03 / Capabilities
Current capabilities
Research instrumentation
Typed writes, policy outcomes, lifecycle receipts, and declared bypasses.
Cross-vendor reference
Published NVIDIA, AMD, and Trainium2 work checked against an explicit numerical contract.
Preprint 1.1
The Missing Medium presents the method, experimental observations, and limitations.
Bilingual technical report
The September V5/V6 report includes ES/EN PDFs, reduced tables, provenance, and a prospective protocol.
04 / Evidence
Evidence and its scope
The preprint and September report answer different questions: instrumentation and replay on one side, factual access and control design on the other.
preprint version
V5 batches
No matching control under all criteria
prospective V6 arms
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.
What it does now
- Records candidate writes and durable outcomes during supported research runs.
- Provides a pinned reference for auditing and replaying recorded microbatches.
- Publishes a preprint, a bilingual report, and verifiers for the portable report data.
Next directions
- Qualify the V6 numerical setup and integrate its runner before executing the protocol.
- Complete the corrected capacity rerun before revising the affected claim.
- Test whether the instrumentation remains informative across further controlled architectures.
Limits to keep in view
- A research prototype, not an efficient training architecture or a production model.
- A preregistered comparison reversed an earlier quality headline; the tested catalog model did not outperform its tuned dense twin.
- A later gradient defect limits the published capacity run; its corrected rerun is pending.
- The current factual-access report does not show that channel death during learning causes hallucinations.
Hytorch