Reproduced
Completed under the stated public protocol with source-bound evidence and a retained authority.
Research ledger / Public record
A number without its protocol is a rumor.
This ledger separates release evidence from preliminary experiments. Every record names what was tested, what it supports, what it does not support, and where the underlying material lives.
Status vocabulary
Measured
Observed under a documented environment but not yet independently reproduced or broadly generalized.
Developing
An active hypothesis, implementation, or campaign. It is not a shipped capability claim.
HYP-001
Native operational-scale control matrix
Hyphae exercised 11 surfaces at C1, C8, and C32 on a DigitalOcean c-60-intel VM: 33 million measured observations, 3.3 million warmups, and 3 million strict logical commits recovered with zero missing or mismatched commits. ANN recall@10 was 1.0 in every cell.
Supports: bounded functional concurrency, accounting, correctness, recall, and durable recovery in the named environment. Does not support: portable latency, bare-metal, dedicated-hardware, or background-interference claims.
Method and limitationsReproduced
HYT-001
30M residual-control campaign
Across eight paired seeds in the published 30M campaign, rezero_rms_shared reached mean final validation NLL 1.5031 versus 1.5389 for pre_rms. The paired improvement was 2.33% with a 95% confidence interval of [1.59%, 3.06%], clearing the preregistered 1% practical threshold. The 1.8-NLL token reduction was 17.5% with a 95% interval of [11.59%, 23.41%].
Supports: the preregistered result under the exact published model, corpus, token budget, seed, and accelerator conditions. Does not support: universal superiority across architectures, scales, datasets, optimizers, or production workloads. The extension exceeded its cloud ceiling by approximately $0.03; that protocol violation remains part of the record.
Protocol, results, and limitationsReproduced
HYT-002
Depth-rung residual-control evidence
At 24 and 48 layers, shared-gate ReZero reached the fixed 1.8-NLL observation threshold 20% earlier in every paired seed. Final-NLL effects were directionally positive but their confidence intervals did not clear the preregistered practical minimum, so those primary verdicts remain inconclusive.
Preserved negative boundary: the project reports the secondary optimization-speed signal without rewriting the inconclusive primary result. Shared-gate ReZero also carried measured memory or throughput costs in these conditions.
Published result recordsMeasured
POL-001
Standing
Standing
Negative-result and retraction policy
Rejected storage, indexing, training, architecture, evaluation, and deployment approaches remain in the design record with the measured reason they failed. A later correction is placed where the original claim stood; history is not rewritten to make the current direction look inevitable.
A performance or research result reports source identity, data identity, architecture, seeds, token and tuning budgets, hardware, protocol, failures, confidence or uncertainty, resource cost, and explicit non-claims.
Policy
Reproduction standard
Evidence must survive the sentence written about it.
For performance work, that means exact source identity, model and container identity, compiler and flags, hardware topology, workload, warmup, repetitions, timing boundaries, statistics, raw logs, output comparison, exclusions, and known confounders.
For releases, it means immutable source, checksums, SBOMs, provenance, signatures, and a receipt that binds them to the selected commit. Temporary CI artifacts alone do not close a release gate.