Released
A named software or artifact version with a public source and distribution record.
Research ledger / Reviewed September 2026
Progress includes what the evidence rules out.
Read publications, inspect the underlying protocols, and distinguish shipped capabilities from measured results and experiments still to come.
How to read the record
Measured
An observation under a stated protocol and environment; broader conclusions require more evidence.
Prospective
A question or protocol for future work. It is not an executed experiment or a demonstrated capability.
2026-09-04 / Hyphae
Hyphae 3.0.0: a broader local data engine.
The release expands bounded SQL analytics, conditional keyspace operations, and lexical and hybrid search. Its complete crate graph and platform archives have public publication receipts. Agent Memory applies the shared data substrate to useful context across coding tools.
The September 3 dedicated-hardware report records SQL, keyspace, search, and commit workloads on an AWS i7i.metal-24xl. Results belong to the documented source, hardware, and workloads; they do not establish a universal performance advantage.
2026-09-03 / Formal methods
Checking the commit protocol with TLA+.
Hyphae’s formal model checks properties of atomicity, durability, conflict handling, and recovery across its modeled state space. The specification, configuration, and evidence receipts are public.
This is evidence about an abstract protocol model. It is not a formal proof of the Rust implementation or every real-world failure.
2026-09-05 / Preprint 1.1
The Missing Medium: a transactional residual stream.
Hytorch records attempted internal writes during training and gives them durable receipts and a CPU replay reference. The preprint presents instrumentation, verification, and the observed collapse of a training channel despite improving conventional telemetry.
The tested architecture is costly and does not establish better language-model quality. A headline improvement reversed under a preregistered comparison; a later gradient defect also limits the reported capacity experiment. The corrected capacity rerun remains pending.
2026-09-12 / Technical report
Write activity, factual access, and control design.
V5 completed 2,190 batches. No candidate satisfied every preregistered comparability criterion: no-control-match. The report publishes variation across facts and positions, reduced tables, and reproducible selection.
V6 specifies 17 experimental arms, with no training executed. The report does not establish that channel death during learning causes hallucinations. The English edition, published September 13, uses the same scientific evidence as the Spanish edition.
2026-08-25 / Transformer
Frozen-model control and bounded navigation.
The later Transformer work trains small controllers over a frozen backbone and publishes deterministic control and navigation bundles. The calibrated navigation v2 bundle records a bounded two-step evaluation.
These results belong to the published fixtures, certificates, and gates. They do not demonstrate an unrestricted autonomous agent or general model quality; failed external-shadow attempts remain in the research record.
Historical record / Hyphae G7
Native operational-scale control matrix.
The August 13 protocol completed 33 million measured observations across 33 surface/concurrency cells in the named DigitalOcean C-60 environment. It also recorded recovery of 3 million logical commits and ANN recall@10 of 1.0 in every measured cell.
This supports bounded accounting, correctness, recall, concurrency, and recovery in that VM. Dedicated-hardware measurements are a separate, later record.
Historical record / Transformer 30M
A controlled residual-strategy result.
Across eight paired seeds, rezero_rms_shared improved final validation NLL by 2.33% relative to pre_rms, with a 95% interval of [1.59%, 3.06%]. The result cleared the campaign’s preregistered 1% practical threshold.
The claim is specific to the model, corpus, budget, seeds, and protocol. It does not establish universal superiority of a residual strategy.
Historical record / Transformer depth
Inconclusive primary results stay in view.
Earlier depth campaigns and the initial three-seed 30M result did not clear their preregistered primary thresholds. Secondary training-speed observations do not change those primary verdicts. A recorded cloud-budget overrun also remains part of the protocol history.
2026-08-24 / BitNet
ARM measurements and the limits of kernel changes.
Runtime 0.3.2, distributed under tag v0.3.3, retains measured NEON DOTPROD behavior and treats its packed SVE2 kernel as experimental. Current ARM code uses packed Q1 panels; earlier expand-to-int8 measurements describe a historical layout.
Prefill and decode are reported separately. A gain in one phase, model, or thread count is not a general runtime speedup.
Research practice
Negative results remain visible.
Inconclusive experiments, corrected claims, rejected optimizations, and protocol defects help determine the next useful question. We retain their original scope and sources.
Published Hytorch tables support rebuilding aggregates and selection. The report does not redistribute every raw historical artifact, checkpoint, or ledger; table verification is not neural replay.
Evidence must survive the sentence written about it.