GFly Learning Lab: scope and evidence review
29 October 2026

This is a source, measurement and artifact review performed during development.
It is not independent peer review and does not establish journal acceptance.

SUPPORTED CLAIMS
The declared model learns two synthetic tasks from scalar feedback. It does not
recover above chance on average within the reward reversal budget. The paper
retains that failure. The handwriting followup uses external observations and
the official separate writer train/test partitions, on three binary tasks.
The 540 study runs are completed measurements, not illustrative numbers.

The browser worker and Node runners use the same model implementation. The
implemented workflow supports starting, pausing, saving, importing and resuming
an experiment. Same engine continuation and numerical agreement across the
tested Chromium and WebKit cases are verified. Dense NumPy reference calculations
agree with the sparse JavaScript equations on ten separate inputs.

CLAIMS DELIBERATELY NOT MADE
There is no evidence that the intact anatomical input topology is superior.
There is no whole brain learning claim, animal fidelity claim, language result,
energy saving claim, mobile hardware result or measured human usability benefit.
The model uses a small selected feedforward circuit, engineered rates, arbitrary
sensory encoding, artificial output groups and conventional policy gradients.
Its 7,620 learned gains are tied across existing connections, not independently
learned biological synapses. Three binary digit tasks are not ten class OCR.

NOVELTY AND VALUE
The paper cites prior mushroom body learning models and the overlapping browser
platform Connectome Lab. Browser training, output plasticity and wiring controls
are not claimed as firsts. The present contribution is the inspectable artifact,
its complete state transfer and its measured comparisons. A reader can discover
that training helps while the anatomical input partners are not necessary for
the reported performance. The API permits a custom two action environment.

The intended benefit is an accessible way to inspect and reproduce these
experiments. Actual researcher time savings or teaching benefits have not been
measured. They remain a reason to test the platform with independent users.

EXPERIMENT DESIGN
The synthetic protocol and implementation hashes were recorded before seeds
101 through 120 ran. Pilot seeds are separate. The handwriting followup was
specified after viewing the synthetic outcomes and before running its seeds
201 through 210. The manuscript explicitly identifies it as a followup.
Each study uses final checkpoints, preserves all conditions and reports means
and sample standard deviations. Paired bootstrap intervals are descriptive,
uncorrected for multiple comparisons and conditional on the chosen tasks/data.
No test result is used to select a best model checkpoint.

REMAINING LIMITS FOR PUBLICATION
A useful software or demonstration paper is plausible. A strong methods claim
about anatomical advantage is not supported by these results. The contribution
needs external review of the software's practical value and comparison with
existing platforms. One data source, one host and a small task suite limit
generalization. The degree preserving swap procedure is finite and is not a
proven uniform graph sampler. The direct input baseline has fewer parameters;
matched nonlinear wiring controls are included to help interpret this limit.

Before a stronger impact claim, ask independent users to reproduce a result
and create a new task, recording completion time and failure rate against an
existing workflow. Do not describe such a user study as completed until it runs.
New tests involving recurrent memory or coupled circuits need separate protocols.

RELEASE CONTENT
The PDF has a compact Jiahe Tian / HiOS Labs Inc. / ipo@stanford.edu byline and
ORCID 0009-0007-8750-287X. It contains no invented journal, acceptance, DOI,
affiliation, citation count or peer review claim. Source data attribution,
all measured study records, source hashes and reproduction instructions accompany
the artifact. Credentials from other projects are not used by this lab.

The opening follows the HiOS v0.7 manuscript format: a centered bold title,
author and company on one line, email and ORCID on the next, followed by the
abstract. The manuscript uses letter paper with one inch margins and is ten
pages in this release. Public manuscript links use hios.app/research/gfly/;
the interactive lab and reproduction bundle use gfly.site/research/. These
presentation and address changes do not alter the frozen experiment protocols,
model implementation, run records or scientific conclusions.

VISUAL AND REUSE REVISION, 30 OCTOBER 2026
Three additional figures show deterministic task examples, all paired seed
differences with the previously reported bootstrap intervals, and the verified
checkpoint workflow. These are views of the completed 540 runs and existing
software checks; no new learning study is claimed. Figure provenance and values
are recorded in checks/visuals/figure-provenance.json.

The manuscript now identifies reusable components and narrows its wiring
conclusion to selected PN to KC input partners under the same output aggregation.
Browser learning, wiring controls and checkpoints are established techniques.
No claim of first platform, new learning algorithm, practical superiority or
future citation counts is made. The intended contribution is a reproducible
comparison baseline and a working artifact with inspectable assumptions.

FRONTEND AND NAVIGATION REVISION, 30 OCTOBER 2026
The static lab, programming guide and readable manuscript share navigation to
GFly home, the learning lab, the research paper and the HiOS research index.
Native CSS frost is limited to navigation and selected desktop control panels.
Phone panels remain unfiltered, and scientific plots retain light backgrounds.
Theme selection is optional and navigation works without JavaScript.

The presentation redraws its existing canvas state after browser context or
page restoration. This does not advance training or alter the model, worker,
frozen experiment protocols, recorded runs, manuscript content or PDF. This is
an interface revision, not an additional scientific experiment.
