What I've read, and what it's for.

The 1945 essay that started it all

Bush imagines the Memex — a personal associative-trail machine. Spiritual ancestor of the digital brain.

Framework for augmentation

The first formal framework for augmenting human intellect; everything from outliners to hypertext flows from here.

NLS live demo

Engelbart's NLS: mouse, hypertext, video conf, outliner — all in 1968. Watch this once.

Personal essay on Anki

Why spaced repetition is the cheapest superpower; how Nielsen uses Anki to learn fields.

Manifesto + research agenda

Long-form essay on why the field has stagnated and what real tools for thought would look like.

Information software as graphic design

On treating information software as graphic design rather than interaction. Foundational reading.

Interactive essay on thinking

Best demonstration of what "thinking with a tool" can feel like.

The original Zettelkasten manifesto

Luhmann's own description of how he thought with cards. Short and life-changing.

Modern Zettelkasten primer

The book that mainstreamed Zettelkasten for English-speaking academics.

History of the field, free online

Howard Rheingold's history of the people who tried to augment thinking. Free full text online.

Philosophy of cognition

Argues your notebook is part of your mind. Philosophical foundation for taking PKM seriously.

Distributed cognition

How cognition is distributed across people, tools, and environment. Relevant for any system that holds your thinking.

Empirical case for testing

Classic experimental evidence that retrieval beats restudy. Underpins SRS-based brains.

Brain2Qwerty: production-side brain-to-text without surgery

The atlas’s PhD north-star paper. Frames the ambition: invasive neuroprostheses restore communication but do not scale; noninvasive methods must close the gap for noncommunicating patients. Key claim: deep learning on MEG can decode full typed sentences (CER 29% avg, 18% best) — narrowing invasive/noninvasive gap. Methodological template for a thesis: high-quality M/EEG + modern sequence models + language priors + rigorous subject-level stats. Companion paper (ref 30) owns the pure neuroscience of language production during typing; this paper owns the decoder. Read order: abstract → Fig. 1 pipeline → CER results (Fig. 2) → ablations → Discussion motor-code analysis → Data/Code availability. PDF OA under CC BY-NC-ND 4.0. Funded by Meta (corresponding authors Meta AI, Paris).

Perception-side MEG → language; the sibling of Brain2Qwerty

Défossez, Caucheteux, Rapin, Kabeli & King, Nature Machine Intelligence 5:1097–1107 (5 Oct 2023); doi:10.1038/s42256-023-00714-5. Cited lineage for Brain2Qwerty: deep models decoding *perceived* natural speech from MEG (subject layer + contrastive / language alignment). Brain2Qwerty flips the problem to *production* (typing) and reports clinical relevance more directly. PhD reading: pair perception vs production papers to decide which representations transfer.

Invasive brain-to-text benchmark the field measures against

Nature 2023. Intracortical speech BCI restoring conversational rates in a person with ALS — the invasive ceiling Brain2Qwerty is trying to approach without surgery. Read to know what “good enough for communication” looks like, and what noninvasive methods still cannot claim.

Multimodal speech synthesis from cortical activity

Nature 2023 speech neuroprosthesis line — complementary invasive production result. Use as clinical ambition reference when arguing why noninvasive production decoding matters for scale and safety.

Atomic, linked slip-box notes

Each note one idea. Each note linked. Each note in your own words. Slow but compounds.

Matuschak's refinement of Zettelkasten

Concept-oriented, densely linked, written for one's future self. Refines Zettelkasten for the digital age.

Projects, Areas, Resources, Archive

Tiago Forte's universal organizing system. Works across any app.

Forte's full BASB curriculum

CODE: Capture, Organize, Distill, Express. Production-oriented PKM.

Nick Milo's Obsidian methodology

Maps of Content (MOCs) as navigational layer above atomic notes.

Analog Zettelkasten, faithful to Luhmann

Argues digital Zettelkasten miss the point; advocates physical cards. Useful counterpoint.

Cues / notes / summary

Old-school but still useful for live lectures and seminar capture.

Practice retrieval over time

The single best-evidenced technique in cognitive science. Anki is canonical; Orbit embeds in essays.

Layered highlighting

Bold → highlight → summary as you reread. Combats premature highlighting.

Local-first markdown graph notes

De-facto standard for researchers. Folder of .md files on disk. Massive plugin ecosystem (Dataview, Smart Connections, Excalidraw).

Outliner with bidirectional links

Free Roam alternative. Block-based outliner over local markdown. Strong daily-notes culture.

The original block-graph notebook

Pioneered the block-graph paradigm. Pricey, but still beloved by power users for query language.

Supertags turn notes into structured data

Outliner + database hybrid. Supertags act like classes; queries are first-class. Strong voice/AI capture.

Object-oriented note-taking

Everything is a typed object (paper, person, concept). Polished UI; native AI assistant.

Whiteboard-first knowledge work

Cards on infinite canvases — exceptional for literature review and synthesizing across papers. Annotates PDFs in place.

AI-native daily notes

E2E encrypted, fast, opinionated. GPT integration deeply baked in.

Self-organizing notes

Tries to remove the work of organizing — AI clusters and surfaces. Search-first UX.

Pages + databases for everything

More project management than brain, but its database views are unmatched for tracking experiments.

Local-first, peer-to-peer Notion alternative

Encrypted, P2P-sync, object-oriented. Privacy-first.

Hierarchical notes in VS Code

Dot-notation hierarchy on top of markdown. Great if your brain lives in VS Code already.

Open-source Roam clone in VS Code

Lightweight, hackable, lives next to your code.

Notes + spaced repetition

Best-in-class for converting notes to spaced-repetition cards while you write.

Cards on a canvas, like Heptabase

Heptabase-style with stronger collaborative features.

30-year-old mind-map veteran

Pre-dates almost everything else. Unique parent/child/jump link model.

Tool for notes with attributes

Mark Bernstein's cult-favorite power tool — programmable, attribute-rich notes. Steep learning curve.

Card-based with social

Tiny cards, fast capture, shareable.

The infinitely nested list

If you just want an outliner: this. Mirrors and zoom levels are killer.

Grounded chat over your sources

Upload PDFs + notes; ask anything; answers cite the source. The "audio overview" turns a corpus into a podcast.

Persistent context for Claude

Pin your BRAIN.md and key papers; Claude keeps that context across every chat in the project.

Persistent context for ChatGPT

OpenAI's equivalent. Memory layer learns about you across sessions.

Cited LLM search

Best LLM-search hybrid for current-event-style research questions.

Evidence-based answers from papers

Ask a yes/no scientific question; get an evidence-graded answer over papers.

Open-source local AI brain

Self-hostable AI agent that indexes your notes, PDFs, emails. Privacy-first.

Local RAG chat UI

Self-host or local app. Talk to your files with any LLM (OpenAI, Anthropic, Ollama).

Run local LLMs in one command

`ollama run llama3` and you have a local model. Foundation for self-hosted brains.

GUI for local models

Friendly UI for downloading and running open-weight models locally.

Embedding-based note recall

Surfaces semantically similar notes as you write. Killer plugin for Obsidian-based brains.

Transcribe meetings & lectures

Capture talks and lab meetings as searchable text.

AI meeting notes that sit beside you

Listens silently to meetings; turns your scribbles + audio into clean structured notes.

AI-native code editor

VS Code fork with deeply integrated Claude/GPT. Good for the simulation/analysis side of your brain.

Codeium's agentic IDE

Cursor competitor; strong agent loop for multi-file edits.

The exemplar evergreen-notes site

Stack-of-cards UI; densely linked atomic notes. The reference implementation of evergreen notes.

Illustrated digital garden

Beautiful illustrations, growth-stage labels (seedling/budding/evergreen). Anthropology-of-tools angle.

Long-form essays + sidenotes

The Wikipedia of one obsessive mind. Studied for its long-form structure and floating sidenotes.

Tools-for-thought researcher

Linus prototypes new note tools constantly (Notation, Histoire, Modal). Embedding-driven UX.

Massive opinionated wiki

Public learn-in-public wiki; useful as a structural reference for large personal wikis.

Strategy consultant's working notes

Coined modern "digital garden" framing in his 2018 "Of Digital Streams, Campfires and Gardens".

Mindful productivity site

Neuroscience-grounded productivity & PKM newsletter — written by a PhD neuroscientist.

Static site for your Obsidian vault

Most popular way to publish an Obsidian vault as a public garden. Backlinks, graphs, search out of the box.

Open-source garden template

Lightweight alternative to Quartz; full control over output.

Daily blogger, lifelong gardener

23+ years of daily posting. Pattern-language thinking; great structural inspiration.

Historian's enormous Obsidian vault

Documents her research workflow in extreme detail — useful for academics.

Mnemonic medium experiment

Interactive essay with embedded spaced-repetition prompts. Influential proof-of-concept.

Public engineer's brain

Down-to-earth example of publishing what most folks keep private.

The open-source reference manager

The default reference manager for academia. Pair with Better BibTeX + Obsidian-Zotero-Integration.

Web-native references

Cloud-first reference manager; best Google Docs integration.

Highlight collection + read-later

Captures every highlight from Kindle/web/PDF and pipes them into your notes. Reader is its read-later app.

Visual literature graph

Generates a visual graph of related papers from any seed paper. Standard for new-field exploration.

Spotify for papers

Builds collections and recommends new papers as they appear. Great for staying current.

AI-powered paper search

AI2's search engine. TLDRs, citation graphs, and a generous API. Plays well with everything else.

LLM-driven literature review

Ask a research question; Elicit pulls papers and extracts fields into a table.

Smart citations

Shows whether a paper is supporting, mentioning, or contrasting another — not just citing.

Citation maps over time

Maps citation networks; monitors for new entrants.

Compartmental neuron simulator

The reference simulator for biophysically detailed neuron and network models.

Large-scale spiking networks

Optimized for large spiking networks. EBRAINS-friendly.

Pythonic spiking simulator

Define neurons by ODEs in plain Python — fastest path from paper to simulation.

Neural Engineering Framework

Build cognitive models with the Neural Engineering Framework. Runs on neuromorphic hardware too.

Markerless pose estimation

Standard for behavior tracking in neuroscience labs.

Unified spike-sorting

One API across every spike sorter. Saves enormous time.

Machine learning for neuroimaging

Scikit-learn for neuroimages. Excellent docs and tutorials.

Pipelines for neuro data

Defines data pipelines as schemas — used in many systems-neuro labs.

The canonical notebook

Still the default. Pair with jupytext for clean diffs.

Reactive Python notebooks

Solves Jupyter's reactivity and git problems. Notebooks are real .py files.

Publishing for science

Successor to R Markdown — beautiful papers, books, sites from one source.

Live JS notebooks

For interactive visualization of models and data. d3 is built in.

Open neural data

Massive open dataset and Python SDK. Indispensable for many computational projects.

Open neuroimaging archive

BIDS-formatted neuroimaging datasets, free to download.

A cubic millimetre of mouse cortex — wiring and activity

Nature package, April 2025. ~200,000 cells, ~84,000 neurons and ~524M synapses reconstructed from 1 mm³ of mouse visual cortex, paired with calcium imaging of ~75,000 neurons — structure *and* function in the same volume, which is what makes it new. EM connectomics was Nature Methods' Method of the Year 2025 largely on the back of it. Publicly browsable with tools for programmatic analysis. This is the substrate the 2024 higher-order topology result was computed on.

Blue Brain's successor — simulation neuroscience, kept open

The Blue Brain Project ran 2005–2024 and closed in December 2024; Markram's not-for-profit Open Brain Institute launched January 2025 to keep the brain-building and simulation stack freely available, under an agreement with EPFL that holds the technologies open for academic research for decades. If you're chasing Blue Brain tooling — or the 11-D lineage — this is where it lives now.

The standard open toolkit for MEG and EEG

Gramfort et al. NeuroImage 2014 — the software Brain2Qwerty-class papers actually run on. Preprocessing, source localisation, decoding helpers, MNE-BIDS, visualisation. If the PhD touches M/EEG at all, this is day-one install. Pairs with Braindecode for deep models and MOABB for benchmarks.

Deep learning for EEG/MEG on top of MNE + skorch

PyTorch-based deep learning library for EEG/MEG decoding — implementations of EEGNet and friends, dataset loaders, training loops. Natural place to reimplement / ablate Brain2Qwerty-style conv stages before writing a custom transformer stack.

Mother of all BCI benchmarks

Standardised BCI evaluation harness across public EEG datasets and pipelines. Brain2Qwerty cites the public BCI benchmark reality-check (low multi-class accuracy on motor imagery) — MOABB is how you run those comparisons cleanly. Use it so PhD results are comparable, not just a custom split.

Fast n-gram language models — Brain2Qwerty stage 3

Heafield KenLM — the style of pretrained character-level / n-gram LM used as Brain2Qwerty’s third stage to correct decoder outputs. PhD fork: swap in neural LMs and measure CER vs compute vs domain mismatch.

Optically pumped magnetometers — MEG without the vault

The hardware hypothesis implied by Brain2Qwerty: if MEG’s SNR win is real, wearable OPM systems (UCL, FieldLine, Cerca, etc.) are how production decoding leaves the magnetically shielded room. PhD-track sensors work package — not software you install, but the deployment constraint you design around. Track vendors + open toolchains (MNE has growing OPM support).

The paper behind the “11-dimensional brain”

Blue Brain, Frontiers in Computational Neuroscience 11:48 (2017). Neurons bind into directed cliques (simplices); when the circuit is stimulated, progressively higher-dimensional cliques assemble momentarily around high-dimensional holes (“cavities”), then disintegrate — a concrete link from wiring to function. Reconstructions consistently contained simplices up to dimension 6–7, with as many as 80M directed 3-simplices. Wet-lab work in Lausanne confirmed the virtual-tissue findings are biologically real.

What the paper says vs what the headlines said

Read before repeating the “brain works in 11 dimensions” line. (1) The 2017 paper's own result is simplices up to dimension 6–7 (7–8 neurons). “11” comes from a Markram remark that in *some* networks they found structures up to 11 dimensions — not the headline finding. (2) These are topological, not spatial, dimensions. Co-author Max Nolte: “A 7- or 11-dimensional simplex is still embedded in the physical three-dimensional space.” The claim is about clique complexity in the wiring graph, not extra dimensions of space. This press release is the source most of the pop-science coverage was built from.

The 2017 method, run on a real EM connectome

Cerebral Cortex, Nov 2024 (Reimann, Egas Santander, Ecker, Muller). The real step past 2017: directed-simplex topology applied to an electron-microscopic reconstruction of actual mouse visual cortex rather than a digital reconstruction. Found an underlying feed-forward network with specific lateral inhibition implementing competition between outputs, itself under control of targeted disinhibition from a specialised inhibitory class. Moves the clique story from “our model has this” to “real cortex has this”.

The clique/cavity lens on human brains

Journal of Computational Neuroscience (2018). Carries the clique/cavity analysis from Blue Brain's microcircuit to the *human* connectome, showing it reveals organisation that ordinary edge-based network summaries miss entirely. The bridge from a speck of simulated rat cortex to whole-brain human imaging.

The 2025 state-of-the-field review

Annual Review of Neuroscience, vol. 48 (2025), pp. 491–518. The best single entry point to the field as it stands — how topological methods (persistent homology, clique complexes) actually connect circuit structure to function. Start here rather than with the 2017 press cycle.

Homology of directed flag complexes

The tool this whole lineage is computed with. Builds the directed flag complex of a digraph and computes persistent homology over flexible filtrations. Based on Ripser but optimised for very large, parallelisable runs, with an “approximate” mode that cuts compute at remarkably little accuracy cost. From Lütgehetmann, Govc, Smith & Levi; applied to Blue Brain's digital reconstructions. Directedness is the point — synapses have a direction, so undirected TDA throws away the signal.

Python bindings for Flagser

Python API over the flagser C++ backend — directed flag complexes without leaving a notebook. A collaboration of L2F, EPFL's Laboratory for Topology & Neuroscience, and HEIG-VD's REDS institute. The practical way in if your stack is already Python.

TDA with a scikit-learn API

Topological data analysis wired into sklearn pipelines with C++ implementations underneath. `FlagserPersistence` handles *directed* graphs — which is what a connectome is — while `VietorisRipsPersistence` covers undirected. The gentlest on-ramp from “I have a connectivity matrix” to persistent homology.

Fast Vietoris–Rips persistent homology

The speed benchmark for persistent homology of Vietoris–Rips filtrations — computes persistence directly from the data instead of materialising the full complex. Flagser is built on its approach and extends it to the directed case.

The general-purpose TDA library

INRIA's broad toolkit for simplicial complexes, persistent homology and topological signatures. The most complete of the TDA libraries when you need more than Rips filtrations — reach for it once flagser/Ripser stop covering the case.

The first EEG foundation model

ICLR 2024 spotlight — “Large Brain Model”. Pre-trained on roughly 2,000–2,500h of EEG drawn from ~20 datasets, at 5.8M / 46M / 369M params. Vector-quantised neural spectrum prediction learns a semantically rich tokenizer over raw EEG channel patches; a Transformer then predicts the codes of masked patches. Beat SOTA on abnormal detection, event type classification, emotion recognition and gait prediction. The reference point every later EEG model is measured against.

EEG treated as a foreign language for an LLM

ICLR 2025, from the LaBraM lab. Tokenizes EEG into discrete, text-aligned neural tokens, then trains an LLM on multi-channel autoregression so a single model understands both EEG and language; multi-task instruction tuning then unifies the downstream tasks instead of needing one head per task. NeuroLM-XL is 1.7B params pre-trained on ~25,000h of EEG — the largest EEG model to date. Built on nanoGPT + LaBraM.

Foundation model for fMRI brain dynamics

ICLR 2024 (van Dijk lab). A Transformer autoencoder over ~6,700h of fMRI from UK Biobank + HCP: parcellate, embed, mask, reconstruct. Fine-tuned it predicts clinical variables (age, anxiety, PTSD) and forecasts future brain states; zero-shot it recovers intrinsic functional networks straight from raw fMRI with no network supervision during training. Weights on HuggingFace (`vandijklab/brainlm`).

Generalist foundation model for brain MRI

Nature Neuroscience — online 5 Feb 2026, in the April 2026 issue. “Brain Imaging Adaptive Core”: self-supervised contrastive pretraining on 48,965 unlabelled brain MRI scans across 34 datasets, then adapted to seven clinical tasks — MR sequence classification, time-to-stroke, brain age, MCI classification, brain-tumour overall survival, IDH mutation status, tumour segmentation. Beats localised supervised training and other pretrained models *especially in low-data / few-shot settings*, which is the actual constraint in clinical neurology. Paper is free on PMC; confirm weight availability before planning on it.

Multimodal CT + MRI brain foundation model

Pre-trained on >3M brain CT slices and 7M brain MRI slices paired with clinical reports, combining image–text contrastive learning with a diffusion-based generative framework. Reports state-of-the-art across seven tasks including disease diagnosis, lesion segmentation, MRI enhancement and automatic report generation. Stated roadmap extends beyond CT/MRI to EEG and electronic medical records — the “one model for the whole neuro chart” direction.

EEG foundation model for Alzheimer's detection

arXiv 2502.01678. An EEG foundation model pointed at one neurological diagnosis — Alzheimer's detection — rather than generic benchmarks. Useful as the counter-example to the “general EEG FM” framing: the clinically valuable version may be narrow and disease-specific.

The skeptical counterweight

arXiv 2507.11783. Reviews EEG foundation-model progress and is candid about the gaps — pretraining/downstream misalignment, unclear generalisation across subjects and datasets, benchmarks that flatter. Read this *before* betting a project on an EEG FM; it is the antidote to the survey papers' optimism.

Comprehensive review of the field

arXiv 2510.16658. Wide-angle survey of foundation and large-scale AI models across neuroscience — EEG, fMRI, imaging, decoding, and the systems built on them. The map of the territory this layer samples; pair it with the critical review above.

Three-stage deep net: typed sentences from MEG/EEG

Nature Neuroscience 2026 (doi:10.1038/s41593-026-02303-2; Lévy, Zhang, Pinet, Rapin, Banville, d’Ascoli, King — Meta AI + ENS/CNRS). Brain2Qwerty decodes *production* of typed sentences from noninvasive recordings while participants type briefly memorised sentences on QWERTY. Architecture (three stages): (1) convolutional module on 500 ms windows (−200…+300 ms around each keystroke), (2) sentence-level transformer, (3) pretrained character-level language model that corrects transformer outputs. Cohort: 35 healthy volunteers. EEG arm 20 participants / ~146k chars / 23k words / 4k sentences; MEG arm 20 / ~193k chars / 30k words / 5k sentences. Headline metrics: MEG character error rate 29 ± 1.7% (best subject 18 ± 2.3%, worst 47%); EEG CER 65 ± 0.7% (best 61%, worst 69%). Beats linear baseline and EEGNet (~2.5× CER improvement on MEG). Ablations show Conv > EEGNet; transformer + LM each add CER gains. Errors cluster on physically neighbouring keys → decoder rides motor/sensorimotor codes more than amodal language. Data: academic request only (svetlana.pinet@univ-lille.fr). Code: Supplementary Information. PhD hook: this is the production-side floor for noninvasive brain-to-text; the open work is wearable sensors, imagined/attempted speech, subject transfer, and clinic-grade protocols. Earlier blog/v1 numbers (WER-style) are superseded by the peer-reviewed CER figures — cite the Nature paper.

Compact CNN baseline every BCI paper still beats

Lawhern et al. compact CNN for EEG-based BCIs — the baseline Brain2Qwerty compares against (and beats, especially on MEG CER). Always include EEGNet (or Braindecode’s port) as a sanity baseline before claiming architectural gains.

Surface EMG → typed characters; the peripheral cousin

Meta FAIR open dataset + baselines for touch-typing from surface EMG — same QWERTY production framing as Brain2Qwerty, different signal. Useful for method transfer (sequence models + LM rescoring) and for arguing what is *neural* vs *muscular* about typing codes.

How far can we push brain→text for *production* (typing, attempted speech) without implants?

Brain2Qwerty proves sentence-level production decoding is possible noninvasively; CER 29% MEG is the new floor to beat. Work packages: Reproduce Brain2Qwerty pipeline on public or collaborator MEG/EEG typing corpora (MNE + Braindecode). · Ablate window size, subject layers, and LM strength; publish a clean open baseline. · Move from typed to attempted/imagined production paradigms suitable for noncommunicating patients. Linked stack: mne-python, braindecode, moabb, eegnet, openneuro, kenlm. Linked papers/models: brain2qwerty-paper, brain2qwerty, defossez-meg-speech, willett-speech-bci, metzger-speech-neuroprosthesis.

Can wearable OPM-MEG (or next-gen EEG) close the SNR gap that made MEG 2× better than EEG here?

MEG CER 29% vs EEG 65% is the single largest lever in the paper — hardware, not just models. Work packages: Benchmark the same decoder across SQUID-MEG, OPM-MEG, high-density EEG, and dry consumer EEG. · Quantify how much of the MEG win is SNR vs spatial sampling vs participant cohort. · Design a clinic-plausible recording montage for locked-in / DOC patients. Linked stack: mne-python, opm-meg, openneuro. Linked papers/models: brain2qwerty-paper, brain2qwerty.

What architecture and language prior best convert noisy neural tokens into usable sentences?

Conv alone > EEGNet; transformer helps CER; character LM adds another clear gain. Full stack is the point. Work packages: Replace 9-gram KenLM with modern subword / neural LMs under domain shift. · Cross-subject and few-shot adaptation (subject layer → foundation-model style pretrain). · Fuse production decoders with EEG foundation models (LaBraM / NeuroLM) as representation backends. Linked stack: braindecode, kenlm, labram, neurolm, jupyter. Linked papers/models: brain2qwerty-paper, eeg-fm-critical-review, neuro-fm-review.

What protocol, ethics, and evaluation turn a lab CER number into a communication aid?

Paper’s stated aim: safe BCIs for noncommunicating patients. Typing with memorised sentences is a proxy, not the end state. Work packages: Define clinically meaningful metrics beyond CER (message rate, user effort, false positives). · Design zero-training or rapid-calibration protocols for patients who cannot type. · Map regulatory / ethics path in AU for research BCI studies (HREC, TGA software as medical device). Linked stack: openneuro, datajoint. Linked papers/models: brain2qwerty-paper, willett-speech-bci, metzger-speech-neuroprosthesis.

Decode what someone is trying to type (and later, say) from brain activity — without surgery.

Narrow the gap between invasive neuroprostheses and safe noninvasive BCIs for restoring language production. Inspired by Lévy et al. Nat Neurosci 2026 (Brain2Qwerty). Why it matters: Invasive BCIs now restore sentence-level communication, but neurosurgery does not scale to the large non- or poorly responsive patient population. Brain2Qwerty shows deep learning + high-SNR MEG can decode *production* (typed sentences), not only perception — the clinically useful direction. The residual gap (MEG ≫ EEG; room-scale shields; motor confusions) is exactly where a PhD programme can land: wearable sensors, better models, imagined/attempted speech, few-shot subject transfer. Open gaps: Data not fully public (academic request only) — reproducibility friction is itself a research opportunity: open corpora. · Room-scale shielded MEG does not enter a ward; wearable sensors are the deployment gate. · Errors are motor-neighbour confusions — may not generalise to pure imagined speech without production motor plans. · Study does not claim a language-production theory (companion paper); PhD can own the computational + clinical side. · CC-style commercial limits on Meta code/blog lineage; peer-reviewed supplementary code is the citable artefact.