Research library | RabbitSoftware

01 / NEUROVISUAL

EEG can retrieve a category better than it can reconstruct a picture.

Recent systems align brain signals with image embeddings, then use a pretrained generator to supply visual detail. The image is not a literal readout of the EEG.

Reviewed October 2026Benchmark: THINGS-EEG2Application: neurovisual research prototype

Perception

On a 200-way retrieval task, the established UBP reference reports 50.9% top-1 with repeated trials averaged.

New preprint

MB2L reports 80.5% top-1 under averaged-trial evaluation; it is an unreplicated 2026 preprint without released code.

Single trial

Performance falls substantially without averaging. A separate report gives about 14.0% top-1, rising to 19.8% with a denoiser.

Imagery and memory

Evidence supports small closed-set categories and recall gist, not reconstruction of a specific imagined or remembered scene.

Interpretation: the chart compares reported top-1 retrieval figures, not image fidelity. Protocol, repetitions, and study maturity differ; the preprint result is not independently reproduced.

Reported top-1 retrieval / 200-way task

NICE-GA / averaged15.6

ATM-S / averaged28.6

UBP / averaged50.9

MB2L / averaged80.5

Percent correct; repeated trials are averaged in these benchmark headline figures. Orange marks the unreplicated preprint. Chance is 0.5% for 200 classes.

Selected sources and benchmark material:

NICE-GA ATM-S UBP MB2L preprint Benchmark leakage audit
Full review, methods, legal analysis, and limits Source notes

02 / GENOMICS

Genomic data informs care through physical treatments and clinical decisions.

The review separates cell-based gene editing, blood-based residual-disease testing, tumour-matched drugs, and digital modelling. None supports a digital twin or radio signal directly changing a body.

Reviewed October 2026Sources include trials, reviews, and dataset recordsApplication: genomics research tooling

01MeasureSequence tumour material or detect tumour-derived DNA in blood.

02InterpretUse validated assays and evidence to identify risk, targets, or trial eligibility.

03Act physicallyA clinician selects a drug, vaccine, cell therapy, radiation, or monitoring plan. The digital layer itself does not treat cancer.

Edited immune cells

Early allogeneic CAR-T studies report responses in small cohorts, mainly for blood cancers. Results are not head-to-head comparisons, and CRISPR cancer treatments are not established as approved therapies in this review.

Earlier detection

In the cited colorectal cohort, circulating tumour DNA became positive a median 142 days before imaging relapse. Detection is prognostic; an effective intervention after a positive test remains a separate question.

Personalized vaccine

A phase 3 melanoma trial reported meeting recurrence-free and distant-metastasis-free endpoints in August 2026. Detailed effect estimates and regulatory approval are not implied by that announcement.

Data access

Processed public datasets can support reproducible research. Controlled-access raw genomes require authorization; personal genomes belong in a private encrypted vault, never a public chain.

Not a treatment tool. Candidate CRISPR guides are untested laboratory hypotheses. Variant comparisons do not diagnose disease. Decisions about testing or treatment belong with an oncology care team.

Selected sources:

GALAXY ctDNA study DYNAMIC-III DYNAMIC trial INTerpath-001 announcement In-vivo delivery review National Academies: digital twins
Full review, evidence tables, datasets, and limits Source notes

Evidence, with its limits attached.

EEG can retrieve a category better than it can reconstruct a picture.

Genomic data informs care through physical treatments and clinical decisions.