Neuroscience and Computational Methods

Format

A 15-month working apprenticeship

Who

Eyad

Window

August 2026 to October 2027

A route into neuroscience that runs alongside USMLE and clinical work. It starts at zero programming, uses EEG as a first practical laboratory rather than a permanent identity, gives you real exposure beyond EEG, and leads to one rigorous research capstone. A first-author manuscript is a serious conditional target, not a promised outcome.

Companion files

eeg-lab.md (the EEG module) and templates.md (every form you copy from). Cross-references name the file and section, so each file works on its own.

Currency

Date-sensitive facts were checked on 31 July 2026. Programmes, meetings, dataset releases, and fees must be rechecked when you actually use them.

How to use this

Read Parts 0 to 4 once. Read Part 8 before your first session, because it governs what happens when a month goes wrong. Then work Part 5 in order. Parts 6, 7, 9 and the companion files are reference you consult, not material you complete.

Every block answers four questions: what is this, why now, what will I actually do, and what proves I finished.

Markers used throughout

Core

Required

Diagnostic

A test you attempt before studying. Pass it and you skip the instruction

If ahead

Optional depth that never delays core work

If compressed

The minimum that keeps the thread alive in a bad month

Research act

Cannot be tested out of, however fast you learn

Gate

A decision with a permitted stop, not a deadline

00 Start here: what neuroscience is, what you will actually do Part 0

0.1 What neuroscience actually is

Neuroscience is the scientific study of nervous systems. It asks how cells generate electrical signals, how cells connect into circuits, how circuits sense and control the body, how populations of neurons support perception and action, and how all of that relates to behaviour and disease.

It is not one subject and it is not another name for neurology. It connects several levels of explanation:

Level	What is studied	Example question	Typical evidence	What that evidence cannot establish alone
Molecular	Channels, receptors, genes, signalling	How does a channel change excitability?	Molecular assays, genetics, pharmacology	Whole-circuit or behavioural consequences
Cellular	Neurons and glia	What makes a neuron spike?	Patch clamp, microscopy, cell manipulation	What a distributed network is doing
Circuit	Connected cell populations	How does excitation-inhibition balance shape a rhythm?	Multi-unit recording, calcium imaging, perturbation	A complete account of cognition or disease
Systems	Interacting regions and pathways	How is a movement planned and corrected?	Spikes, LFP, EEG/MEG, imaging, lesions, stimulation	Meaning or causality unless the design supports it
Cognitive/behavioural	Perception, learning, memory, decisions	How does evidence become a choice?	Psychophysics, computational behaviour models	The neural mechanism without neural evidence

Clinical, computational, and engineering work cut across those levels rather than sitting above or below them:

Lens	What it contributes	Example question	Main caution
Clinical	Frames dysfunction, diagnosis, prognosis, treatment	Which measurement changes a clinical decision?	Prediction does not establish mechanism or clinical utility
Computational	Equations, simulation, statistics, machine learning at any level	Which mechanism or representation reproduces the observation?	A model fit is not biological truth
Engineering	Instruments, interfaces, acquisition, stimulation	Can a device measure or alter the target reliably?	Technical performance is not scientific meaning

Four questions to ask of everything you learn:

What mechanism is proposed?
What experiment or observation supports it?
What exactly was measured?
What conclusion remains unsupported?

That habit, not memorising subfields, is what starts making you a neuroscientist.

0.2 Neighbouring fields, without assigning you an identity

Field	Central job	Relationship to this roadmap
Neurology	Diagnose and treat nervous-system disease	Your clinical background. Not the same as neuroscience
Clinical neurophysiology	Measure and interpret EEG, EMG, and related signals	A supervised clinical specialty. This roadmap gives research literacy, not certification
Cognitive neuroscience	Connect brain measurements to perception, memory, language, decisions	Sampled through systems material and non-EEG work
Systems neuroscience	Explain how circuits and populations generate function	Sampled through population data and mechanism questions
Computational neuroscience	Use formal models to understand nervous systems	A genuine strand here, both data-driven and theory-informed
Neuroinformatics	Organise, standardise, analyse, and share neural data	A strong methodological fit for the capstone
Neuroengineering	Build interfaces, devices, acquisition, stimulation	Adjacent. Not trained here
Clinical AI	Build or evaluate models meant to affect clinical decisions	One possible project family, not the definition of neuroscience

You do not need a lifelong label in Block 1. You need a first laboratory, enough breadth to compare it against alternatives, and a defensible next decision.

0.3 What you will physically be doing

The work is less mysterious than the vocabulary. A session looks like one of these:

Session type	What happens on screen	What remains afterwards
Mechanism	Read a section, redraw the causal chain, predict a perturbation	A mechanism map and a prediction
Coding	Open a notebook, change code, run it, read the error, fix it	A working notebook and an explained result
Data	Load a file, inspect metadata, plot raw values, check units and gaps	A data card and quality notes
Paper	Reconstruct question, design, measurement, figure, claim	A one-page figure audit
Question	Turn an observation into a card and try to kill it	A triaged, killed, parked, or promoted question
Mentor	Send one bounded artifact, ask one answerable question	Written feedback and a decision
Research	Run the locked analysis, log deviations, test robustness	Results and a decision log
Writing	Explain what you did, why, what happened, what stays uncertain	A section of the report

Early on, most time is learning and producing small artifacts. After Gate A, the project becomes the vehicle through which you learn the next layer.

0.4 The terms you need before starting

Dataset / corpus

A defined collection of observations plus metadata. The version matters

Signal

A measured quantity varying over time or space. Not the phenomenon itself

Extracellular spike / sorted unit

A brief voltage event recorded outside cells. Clustering events into a putative source makes a "unit," which is not guaranteed to be one cleanly identified neuron

LFP

Local field potential. Low-frequency extracellular voltage reflecting nearby summed activity, shaped by geometry and reference

EEG / MEG

EEG measures referenced scalp voltage differences; MEG measures magnetic fields outside the head. Both mainly reflect coordinated population activity, not thoughts

fMRI BOLD

An indirect vascular and metabolic correlate of neural activity, slower than electrical recording

Feature

A number given to a model as input

Model

A formal rule mapping inputs to an explanation, estimate, or prediction

Observation vs experiment

Observational designs measure what occurs; experiments deliberately manipulate under control. Causal claims depend on design

Artifact

Concrete evidence of work: a notebook, map, memo, protocol, figure, or report

Transfer test

The same reasoning applied to an unfamiliar example

Unit of independence

The entity that counts as one observation. Usually the patient, not a window cut from that patient

Data leakage

Information from evaluation data influencing training or model choice

Development vs final evaluation

Data used to build and choose, versus the prespecified holdout or outer resampling used for the honest estimate

Validation

Testing whether a method works for its declared purpose under conditions that matter. Not whether the code runs

Estimand

The exact quantity you are trying to estimate

Protocol

The analysis decisions fixed before the primary result is known

Reproducibility

The same inputs and recorded process regenerate the declared outputs within stated tolerances

Preprint

A public manuscript before peer review. It has a persistent identifier and is not an accepted paper

Programming and tooling terms are explained in Part 5.2, immediately before you install anything.

0.5 What success means, and does not

A successful route ends with:

a working map of neuroscience across levels and methods;
the ability to connect mechanism, experiment, measurement, analysis, and inference;
functional Python, statistics in code, validation, and neural-data skills;
direct work with EEG, one population-spike example, and one decision-behaviour example, plus an explicit comparison against imaging;
experience generating, killing, and refining research questions;
one reproducible capstone, a full scientific report, and an honest decision about whether dissemination is justified;
a lawful reproducibility archive and a next-question dossier;
a real record of outreach, and ideally one sustained research relationship.

If the research track, Publication Release, and the authorship agreement all pass, the conditional outcome is a first-author manuscript, preprinted where appropriate and submitted to one journal. Acceptance is a later external event and is not required for this roadmap to have succeeded.

This roadmap does not promise a supervisor, programme admission, publishable novelty, journal acceptance, multiple papers, independent EEG interpretation, clinical certification, or a permanent specialty.

The paper is evidence that the apprenticeship worked. The apprenticeship is not reverse-engineered to manufacture a paper.

01 Direction, exploration, and the field map Part 1

1.1 The method spine and the first laboratory

Two things are fixed in advance. The scientific question is not.

Method spine: Python, quantitative reasoning, experimental design and inference, data handling and validation, and reproducible practice. These transfer to any neuroscience subfield.

First laboratory: EEG. Chosen because it is accessible without institutional gatekeeping, clinically legible to you, computationally tractable on a laptop, and makes acquisition, noise, labelling, and evaluation concrete.

The wording that keeps this honest:

EEG is my first working laboratory because it is accessible, clinically legible, and computationally tractable. It is not my professional identity. After structured exposure to other levels and modalities, I will explicitly choose whether to deepen, broaden, or switch.

If a strong mentor-led non-EEG opportunity appears, it may replace the EEG capstone through a written substitution memo (templates.md §11.10). It never runs in parallel as a second toolchain.

1.2 The field map you build

FIELD-MAP.md is a living table you add to from Block 1 onward. One row per topic you actually study:

Column	What goes in it
Phenomenon	What is being explained
Level	Molecular, cellular, circuit, systems, cognitive
Lens	Clinical, computational, engineering, or none
Measurement	What was physically recorded
Design	Observational or experimental
Supported inference	What the evidence does establish
Unsupported claim	What it does not, however tempting
Your interest	High, medium, low, and why

By the Interest Gate you should have 12 or more substantive rows spanning at least three levels and three measurement types. Twelve rows that span categories beat twenty that repeat one. That table, not a feeling, is what makes the Block 4 decision defensible.

1.3 The breadth floor

Before you narrow, you must have touched more than one kind of neural data. The floor is three thin slices, each using the same seven questions:

What was measured?
At what spatial and temporal scale?
What preprocessing changed the observation?
What is the unit of independence?
What question can this answer?
What causal claim can it not support?
Did the scientific question interest you, or only the tool?

Slice	Modality	Where
1	EEG time series	Block 2
2	Population spikes	Block 3
3	Decision behaviour plus a prepared fMRI-BOLD derivative	Block 4

These use small curated datasets and prepared notebooks. They are comparisons of measurement and inference, not invitations to install three preprocessing ecosystems.

1.4 The Interest Gate, end of Block 4

Gate · Interest · end of Block 4

The one gate that is about you rather than the science.

You need: 12+ category-spanning field-map rows, three completed thin slices, and an honest note on which sessions energised you and which drained you.

You decide: deepen EEG, broaden before choosing, or switch home base.

If you choose to broaden: complete one additional predefined thin slice using at most one floating slot, then make the home-base decision. Exploration is bounded; it does not continue indefinitely.

Passes when the decision is written in one paragraph, references specific rows, and names what you are giving up.

Switching is not failure. Discovering in month 4 that population coding interests you more than clinical signals is the gate doing its job. What fails is drifting without deciding.

02 How research questions emerge and survive Part 2

No study is assigned in this document. Finding and killing questions is the transferable skill; a specific question is perishable.

2.1 Progressive question cards

Curiosities stay cheap. Only finalists earn expensive audit work. Full template in templates.md §11.3.

Stage 1

Curiosity

Observation, where it came from, level and domain, why it might matter, what you already know, what it would teach you, status.

Stage 2

Triage

Likely population and data, who benefits, obvious access or ethics or compute barrier, cost of the missing learning, the smallest test that could kill it, the fatal risk.

Stage 3 · finalist only

Finalist

Precise estimand and comparator, five closest works and your differentiator, full feasibility check, intended-use statement, learning contract.

Statuses: curiosity, triaged, audit candidate, killed, parked, selected. Killing questions is progress. A log with nothing killed means nothing was examined.

2.2 Where questions come from

Source	How to work it	Caution
Limitations and future-work sections	Read the last two paragraphs of twenty papers, log every stated gap	Authors' suggestions may be unimportant, infeasible, or already done. Treat as prompts, not endorsements
Failed reproduction	Try to reproduce a published result	Becomes research only when the failure is attributable, consequential, and documented against a precise claim
Benchmark and evaluation gaps	Where a community standard exists, non-standard evaluations can be redone standardly	Check the benchmark's own leaderboard first
Transportability	A model developed on one population, tested on another	Needs two genuinely compatible datasets
Label and reference-standard quality	What the ground truth is, how reliably it was made, what its noise does to measured accuracy	Needs annotation documentation to exist
Practice-literature mismatch	Something your service does that the literature does not describe	Uniquely yours, but needs access and usually ethics
Error analysis	The cases your own model gets wrong	Only available after you have built something
Mentor conversation	People with more questions than hands	Highest yield of all, once the relationship exists

2.3 The dated overlap audit

Run before committing to a question. Produces a dated memo (templates.md §11.4).

Search all of these:

Source	Why
PubMed	Biomedical baseline. Never sufficient alone
IEEE Xplore	Much neural-signal ML is engineering literature PubMed does not index
Scopus or Web of Science	Cross-disciplinary, via EKB
medRxiv, bioRxiv, arXiv	Current work appears here first, often a year before journals
Google Scholar forward citations	Take the closest paper, read everything citing it
PROSPERO, OSF	Registered but unpublished work
Conference proceedings	Frequently ahead of journals
Field benchmark pages	If a benchmark exists, the obvious comparisons are done

The differentiator test. If you cannot state in one sentence what your study does that the closest existing work does not, you have an interest, not a project.

Audits go stale. Refresh briefly before protocol lock and again before any release decision. A timestamp records what you planned and when. It does not freeze the literature.

2.4 Feasibility, intended use, one estimand

Feasibility screen. Any failing row kills or reshapes the question.

Dimension	Fails if
Data access	Needs a dataset whose approval you have not started
Independent units	Too few genuinely independent units to support the claim
Precision	No archetype-appropriate precision or feasibility assessment at the independent-unit level: expected confidence-interval width, minimum detectable effect, or a simulation-based justification. If precision is inadequate, narrow the estimand or claim, change the design, or stop the research route
Compute	Needs more than a laptop plus free tiers
Time	Cannot reach a frozen analysis inside the remaining blocks
Skills	Needs a technique costing more than the learning contract allows
Review	Needs domain adjudication nobody has agreed to give
Audience	You cannot name who would care

Intended-use statement. One page before any modelling: population, setting, who would use the output, what decision it informs, the reference standard, what error is acceptable, the comparator, the primary endpoint. Without it, "good performance" has no meaning.

One estimand. The commonest way a first project dies is three primary questions. Pick one. Demote the rest to secondary or future work.

2.5 The project-learning contract

The mechanism that lets the project teach you the next layer without becoming an unfunded detour. Written at Gate A, naming exactly four things:

One scientific concept to deepen.
One method to learn.
One implementation skill to acquire.
One thing explicitly out of scope.

Then the loop, for each gap: classify it (concept, measurement, statistics, implementation, or interpretation), pick one primary resource and one exercise, predict what it should change before applying it, demonstrate it on toy or development data, integrate it, and write a short note on what it taught you.

If the missing layer cannot be learned and demonstrated inside Block 6, you narrow the question, choose a better-supported project, or park it. "Learning it is the point" is valid only when the time is actually there.

2.6 Gate A, end of Block 5

Gate A · end of Block 5

The most consequential decision in the roadmap.

Requires: 8 or more logged questions, all triaged; one full overlap memo (a second only for a genuine tie); a feasibility screen with no failing row; an intended-use statement; one named estimand; a learning contract; and one external reader who has commented on scope.

Three permitted outcomes:

Outcome	When	What follows
Research project	A material, feasible question with a review route	Blocks 6 to 13 as the research track
Learning capstone	No question survives, but a rigorous bounded analysis exists	Blocks 6 to 13 as the structured learning track (Part 6.4). No novelty claim
Delay and redesign	Neither is ready	Use a floating slot, then reassess. Do not force a redundant study

A learning capstone is a legitimate success. Manufacturing a novelty claim to protect a schedule is not.

03 Project directions, not assigned studies Part 3

Directions with worked shapes, not assignments. Your Gate A choice determines which template Blocks 6 to 13 follow.

Archetype	Question shape	What it demands	Typical trap
Benchmark / re-evaluation	Do reported performances survive standardised re-evaluation?	Careful reimplementation, protocol discipline	Underestimating how long faithful reimplementation takes
Transportability	Does a model frozen on A hold on B?	Two compatible datasets, harmonisation reasoning	Datasets that look compatible and are not
Reproducibility audit	Can published pipelines be rerun at all?	Software forensics, patience, tact	A failure that is your environment, not their science
Label / reference-standard quality	How reliable is the ground truth, and what does its noise do?	Annotation documentation, measurement theory	Needing agreement data that was never published
Acquisition constraint	What survives cheaper, shorter, or lower-density acquisition?	Signal processing	A well-studied area where the answer is already known
Interpretability / human factors	Can a user read and act on the output?	Study design, ethics, participants	Needing human participants and approval you do not have
Mentor-supplied	Whatever the collaboration needs	A real role and a written substitution memo	Becoming unpaid labour with no authorship clarity

Two worked examples of published work in the constraint family, useful as models of scope and method rather than as questions to take: Lin YC, Lin HA, Chang ML, Lin SF. Neurophysiol Clin. 2025;55(2):103044 (DOI); and Kojima J, Shi H, Ojemann WKS, et al. medRxiv preprint, 2026 (DOI). Read them to see what a finished question in this space looks like, and note that both occupy the montage-reduction question densely.

Human-factors, local-data, and clinically adjudicated projects are eligible only if ethics, participant or adjudicator access, and review support are secured by Gate A. Otherwise the first capstone stays inside a public-data template.

3.1 Evidence synthesis

A systematic review is not a lighter option. It is a team project that typically takes far longer than people expect.

In this roadmap: Block 13 permits only a scope, protocol, search, and throughput pilot for a possible later review. A full review can replace the empirical capstone only through a Gate A decision with a review-specific replacement map for Blocks 6 to 13.

Publication-grade synthesis requires independent duplicate screening and eligibility assessment, independent risk-of-bias assessment, checked extraction, and a design-appropriate conduct standard: Cochrane DTA and PRISMA-DTA for accuracy questions, QUADAS-3 for appraisal. PRISMA is a reporting guideline, not a conduct manual.

04 Data, measurement, and validity Part 4

4.1 A dataset is a scientific object

Access is part of the science, not an administrative preliminary. Every dataset you use gets a card (templates.md §11.5) recording version, provenance, licence, annotation origin, and known limitations.

CHB-MIT

Open, no credentialing, ODC-By licence. 23 cases from 22 subjects, with 182 seizures in that original set; chb24 was added later, bringing the collection to 198 seizures. chb21 is the same person as chb01, recorded 1.5 years later, so split at the person level, not the case level. chb24 is absent from the published subject-information file. If its identity cannot be resolved, exclude it from patient-independent confirmatory evaluation, or restrict it to exploratory and descriptive analysis, and say which you did. Channels are already bipolar derivations, with duplicated and reversed entries in some records

Teaching, montage arithmetic, pilots. The independent unit is the person, never the window, so the effective sample is small whatever the row count suggests

TUSZ

The seizure-annotated subset of TUH EEG. NEDC approval, then SSH key and rsync. Local governance checked separately. Request early

Seizure-event work with official train/dev/eval partitions

OpenNeuro

Open, CC0, many EEG datasets with true electrode channels

Re-referencing exercises, reuse, cross-dataset work

Neuromatch teaching subsets

Bundled with tutorials

Thin slices. Deliberately small and nested

IBL public data

Open via client library

Decision-behaviour slice

MIMIC, ADNI, UK Biobank

Credentialing, affiliation, or fees

Not in this roadmap

4.2 Non-negotiable validity rules

These hold whatever your Gate A choice turns out to be.

Name the exact corpus and version in every claim.
Distinguish electrode potential, channel or derivation, acquisition reference, and display montage. Also distinguish membrane voltage from scalp voltage. Conflating these produces experiments you did not intend.
Split on the unit of independence, which is usually the patient or session, never the window.
Fit every learned transform inside training folds. Scaling, imputation, and feature selection included.
Do not treat outer-fold scores as independent replicates.
Final evaluation uses an untouched group-aware holdout, an external set, or prespecified nested resampling. A single fixed test set is one option, not a requirement.
Count independent units before claiming precision. Nested observations inflate apparent sample size.
Report metrics that carry the clinical or scientific meaning. For continuous detection that means event-level sensitivity and false detections per unit time, with any window-level summary clearly secondary.
Record every deviation from protocol with its reason.

4.3 The leakage demonstration

Block 4's core exercise, and the single most valuable notebook in the roadmap. Train the same model three ways on the same data: split by row, split by group, and nested group-aware. Report all three numbers and explain the gap in writing. Once you have seen the inflation with your own data, you will never trust an unqualified accuracy figure again.

4.4 Reporting and appraisal chooser

Pick by design, do not stack all of them.

Your design	Use
Prediction model, development or validation	TRIPOD+AI. Collins GS, Moons KGM, Dhiman P, et al. BMJ. 2024;385:e078378
Diagnostic accuracy against a reference standard	STARD, and STARD-AI where AI is central
Appraising other people's prediction models	PROBAST+AI
Observational clinical study	STROBE
Systematic review	PRISMA and the design-specific extension
Anything else	Search EQUATOR before writing, not after

4.5 The minimum reproducibility package

Every capstone ships with:

pinned environment specification;
dataset and version manifest with identifiers and, where licences permit, hashes;
seeds and deterministic settings where feasible, with declared tolerances where exact equality is not achievable;
DECISIONS.md logging every analysis choice and its reason;
a README with a one-command reproduction path;
a lawful archive, typically a versioned Zenodo deposit with a DOI;
a small distributable or synthetic fixture so anyone can smoke-test the code path without holding restricted data.

For gated data such as TUSZ, reproduction means: another eligible and authorised user can reproduce the result from the documented lawful access and data-placement starting point, without undocumented assistance from you. You cannot ship the data; you can ship everything else.

The archive is a reproducibility artifact, not a publication. It does not count as an output and no roadmap in this file should imply otherwise.

05 The thirteen blocks Part 5

5.1Calendar logic

Thirteen ordered blocks. Fifteen calendar slots. Two slots hold no mandatory work.

There is no hour ledger and no reserve bank. Time cannot be saved in March and spent in September, so the plan does not pretend otherwise. What protects you is spare calendar, not accumulated credit.

Rule	Detail
One at a time	Only one core block is open. Finish it or formally pause it
Blocks are ordered	Each depends on the last. Reordering breaks the dependencies
Load labels, not hours	Light, Standard, or Full. Sized to fit roughly one to two hours a day, clustered however suits you
Slots absorb disruption	A block spilling into the next slot is the system working
Rebaseline, never catch up	See Part 8

The completion rule, canonical.

All Core items are required unless explicitly marked remediable, movable, optional, or conditional. The done-when test adds the required integration or transfer demonstration; it does not replace the Core list.

A block ends when its Core items are done and its done-when test passes, not when its content has been read. Block N+1 does not start until Block N passes. That is what stops overfill turning into invisible debt.

When the target date moves: only when remaining blocks exceed remaining slots. Not automatically after a bad month. Check at the end of any compressed or paused slot.

5.2Tools, explained before you install them

Every tool below exists to solve a specific problem. Read this before Block 1.

Python is a programming language. You will use it to load data, compute things, and make figures. You do not need computer-science depth; you need enough to express an analysis and read an error message.

A package is code someone else wrote that you import. numpy for arrays and maths, pandas for tables, matplotlib for plots, scipy for statistics, scikit-learn for models, mne for EEG. You will almost never write these things yourself.

An environment is an isolated set of packages with pinned versions. It exists because Project A needing version 1.2 and Project B needing version 2.0 will otherwise break each other, and because "it worked on my machine" is not reproducible. Conda creates and manages environments; Miniforge is a lightweight installer for it. When you export an environment.yml, you are recording exactly what your results depended on.

A notebook is a document mixing code, output, figures, and prose, run in chunks. Good for exploring and explaining. Its weakness is that cells can be run out of order, so a notebook can appear to work while being irreproducible. Hence the habit, from Block 1 onward, of restart the kernel and run from the top before you believe anything.

Git tracks changes to files over time. A commit is a saved snapshot with a message saying what changed and why. It exists so you can see what you did three weeks ago, undo a change that broke something, and prove when work was done. GitHub hosts Git repositories online, giving you backup, a link you can send someone, and a public record. A repository or repo is one project's tracked folder.

Zotero stores papers and generates citations. ORCID is a free permanent researcher identifier that journals and data providers ask for.

The exact sequence. Run these in order. If a step fails, fix it before the next one.

# 1. after installing Miniforge, create the environment
conda create -n neuro python=3.11 -y
conda activate neuro

# 2. install everything this roadmap uses
conda install -c conda-forge numpy pandas matplotlib scipy scikit-learn \
    jupyterlab mne -y

# 3. record exactly what you have, and commit this file
conda env export --from-history > environment.yml

# 4. set up version control in your project folder
git init
git add .
git commit -m "Initial environment and smoke test"

--from-history records only what you asked for rather than every transitive dependency, which makes the file portable across operating systems.

What you do in Block 1:

Install Miniforge, or keep a Python setup that already works Install VS Code with its Python and Jupyter extensions Install Git and create a GitHub account Register an ORCID Install Zotero and make collections for foundations, EEG, questions, capstone Confirm your EKB or institutional library route works

Do not reinstall a working setup to match this list. Never put non-public papers or any patient material in a public repository.

5.3The roadmap at a glance

#	Block	Load	Gate	Ends with
1	Orientation, baseline, working system	Standard		Environment runs; baseline recorded; people mapped
2	Python foundations, mechanisms, first EEG contact	Full		Real EEG file loaded and described correctly
3	Statistics in code, circuits, spike data	Full		Estimate with uncertainty; population thin slice
4	Validation, systems, cross-modal, Interest Gate	Full	Interest	Leakage demonstrated; home base chosen
5	Question triage and Gate A	Standard	A	One question selected and defended, or learning track
6	Project-specific learning and development pilot	Full		The next layer learned and demonstrated
7	Development workflow and draft protocol	Standard		Pipeline runs end to end on development data
8	External review, separation, Gate B	Standard	B	Protocol locked and registered before results
9	Primary analysis	Full		The prespecified result exists
10	Robustness, error analysis, claim control	Full		Result survives or fails prespecified stress
11	Freeze, clean rerun, Gate C	Standard	C	Everything reproduces from a clean environment
12	Full scientific report and Release decision	Full	Release	Report complete; dissemination decided honestly
13	Execute the decision, close out, then stretch	Standard / Full		Archived always; submitted only if Release passed; next-question dossier
	Two floating slots			Absorb disruption. No mandatory content

5.4Your first thirty days

The only week-by-week section, so that "start" has a visible meaning. It covers Block 1.

Week 1 · orient

Read Parts 0 to 4 of this file, and Part 8 Create your working folder with PROGRESS.md, FIELD-MAP.md, QUESTIONS.md, PEOPLE-ACCESS.md from Templates Attempt the Block 1 diagnostic before studying anything. Record what was easy, shaky, impossible. It is a baseline, not a verdict Read Part 5.2 above, so no tool in week 2 is an unexplained word

Week 2 · make it real

Install the stack; create an environment; run the smoke test; commit it Complete Kaggle Python, roughly 5 hours, skippable if the diagnostic was clean Reopen your smoke-test notebook, change the array, the axis label, and one plotting choice, restart the kernel, rerun from the top, and explain what changed

Week 3 · scope and people

Learn what EEG measures and why a trace is a referenced voltage representation rather than "brain activity" (see eeg-lab.md sections 1 to 2) Write the measurement chain for EEG, spikes, and fMRI: phenomenon, recorded quantity, transformation, claim still unjustified Rewrite three categorical EEG statements as observation, bounded inference, uncertainty. No waveform diagnosis Map 8 to 12 candidate people or labs in PEOPLE-ACCESS.md. Profiles and titles for all; read properly only for the best three

Week 4 · integrate and close

Complete Kaggle Pandas, roughly 4 hours Produce the mechanism-and-measurement explainer and the reproducible starter notebook Take the Block 1 transfer test from templates.md §11.11 Pick your best three outreach targets for Block 2 Note which session types flowed and which stalled. No retrospective hour ledger

If setup eats the month, use a floating slot. Do not delete the orientation or pretend the environment works.

Orientation, baseline, and a working system

Load

Standard

Focus

Setup, scope, baseline, people

Gate

None

Compressed Paused

0%0/0 done

Why this block matters

You are building the small operating system through which everything else becomes visible and recoverable. The aim is not to learn Python or survey neuroscience in a month. It is to understand the kind of science you are entering, find out what you can already do, and make the next action runnable. Setup has lead time you cannot compress later, which is why access requests and people-mapping start now rather than when you need them.

Diagnostic first

Before touching the setup instruction, try to: create and activate an isolated Python environment; load a small CSV, inspect types and missingness, compute a grouped summary, make a labelled plot; initialise version control, commit, restart, rerun; and explain in plain language what the rows, code, and plot represent. If that works unaided on unfamiliar data, skip the installation tutorials and the Kaggle modules. The smoke test, the written explanation, and the reproducibility record still stand: they are evidence, not instruction.

Core orientation

Neuromatch Neuro Video Series intro and Neurons to Consciousness Pick one phenomenon (action selection, memory retrieval, a seizure) and write a distinct question about it at molecular, cellular, circuit, systems, behavioural, and computational levels → NEUROSCIENCE-EVIDENCE-MAP.md Map six methods to what each records and what it cannot infer Read the Neuromatch Model Types intro and write 250 words distinguishing descriptive, predictive, and mechanistic uses of a model

Core measurement and EEG scope

Neuromatch EEG/MEG measurement video and the sequence section of the AES Orderly Approach For EEG, extracellular spikes, and fMRI BOLD, write the four-step measurement chain → EEG-SCOPE-CARD.md Rewrite three categorical EEG statements as observation, bounded inference, uncertainty

Core system

Everything in the Part 5.2 install checklist Smoke test. In a repo named neuro-roadmap: create an environment, open a notebook, import numpy, pandas, matplotlib, scipy, sklearn, mne, generate a small array, save one labelled plot, export the environment file, commit, restart the kernel, rerun top to bottom Kaggle Python and Kaggle Pandas, unless the diagnostic cleared them

Core people and questions

8 to 12 rows in PEOPLE-ACCESS.md: why the match is real, one recent work, what bounded help you might later ask for, contact route, status. Local routes first 5 field-map rows and 3 curiosity cards. Naive cards are fine BASELINE.md: honest uncertainties, setup friction, which session types need more support

Done when

The smoke test runs from a restarted kernel, your four tracking files exist and have real content, the Block 1 transfer test passes (templates.md §11.11), and you can explain what an environment and a commit are without looking them up.

If ahead

Add two more field-map rows in an area you know nothing about.

If compressed

Finish the smoke test and one outreach row. Nothing else. Everything here is a dependency for Block 2.

Python foundations, mechanisms, and first EEG contact

Load

Full

Focus

Data handling, neuron mechanism, real EEG

Gate

None

Compressed Paused

0%0/0 done

Why this block matters

Real EEG appears now rather than in month six. Early contact surfaces file-format, naming, memory, and metadata problems while they are still cheap. It also makes the mechanism reading concrete: you will read about membrane voltage and then look at scalp voltage, and the difference between those two is the point.

Diagnostic first

On an unfamiliar tabular dataset, load it, inspect types and missingness, compute a grouped summary, produce a labelled figure, and explain what one row represents. If clean, compress the pandas instruction but not the EEG work.

Core data handling

Kaggle Intro to ML → tabular-transfer.ipynb with a restart-and-run check Practice on data you did not choose, so the skill is transfer rather than replay

Core mechanism

Read Sections 1.2 to 1.5 of Resting and Action Potentials and Sections 6.2 to 6.6 of Synaptic Transmission Work the Neuromatch leaky integrate-and-fire tutorial: Section 1.1, Section 2.2 with Coding Exercise 4, Section 3 with Coding Exercise 6 Predict before you run: state what changing input current, membrane time constant, and threshold should each do, then run it → mechanism-to-measurement.ipynb The omission table. In the same notebook: process | intervention | predicted observable | measurement | what LIF omits, filled for astrocytic ion and transmitter homeostasis, oligodendrocyte myelination and conduction timing, microglial synaptic remodelling, and NMDA-dependent LTP. Sources: Section 3.3 of Cellular Anatomy and Section 7.3 of Synaptic Plasticity. The point is that a spiking model implements none of them

Core first EEG contact

Read the introduction and Sections I to III of ACNS Guideline 2, electrode nomenclature, then Section 1 and one table each of longitudinal-bipolar, transverse-bipolar, and referential from Guideline 3, montages Load PhysioNet EEG Motor Movement/Imagery v1.0.0, file S001R01.edf, via the MNE EEGBCI loader. This dataset has single-electrode-labelled referential channels rather than precomputed bipolar derivations, which is why it and not CHB-MIT is used here Standardize channel names, attach the standard_1005 montage, inspect metadata, and plot a fixed 30 to 60 second copy. Do not interpret it clinically Preserve the original object. Make one average-reference copy and one copy with a derived bipolar Fp1-F7 channel computed from the two electrode channels. Compare the three representations Record the acquisition reference only if documented. Otherwise write unknown → first-eeg.ipynb plus a versioned data card (templates.md §11.5)

Core people

Send the first outreach messages to your best three targets (templates.md §11.8). One bounded artifact, one answerable question 2 field-map rows

Done when

You can pass the Block 2 table and montage cases in templates.md §11.11 without the key, and explain out loud how electrode potential, channel or derivation, display montage, membrane voltage, and scalp voltage differ. No waveform diagnosis required or permitted.

If ahead

Open CHB-MIT file chb01_01.edf, recognise that names like FP1-F7 are already bipolar derivations, reconstruct two of the equations, and explain why ordinary re-referencing is not recoverable from them. Do not apply an average reference to derived channels.

If compressed

Read one mechanism subsection, keep five retrieval prompts alive, do one outreach action. Resume the EEG work before Block 3.

Statistics in code, circuits, and spike data

Load

Full

Focus

Estimation with uncertainty, population data

Gate

None

Compressed Paused

0%0/0 done

Why this block matters

Three things usually taught separately come together: circuit and population reasoning, a quantitative estimate with honest uncertainty, and a real neural observation. You learn one analysis family properly rather than surveying a catalogue of tests. This is also the second thin slice, so you now have two modalities to compare.

Diagnostic first

On unfamiliar paired data, fit a regression and generate a bootstrap interval. Explain the sampling unit, what the interval represents, why a narrower interval is not necessarily less biased, and how grouped observations would change the resampling. If sound, skip the exposition but not the coded transfer.

Core one statistical family

Neuromatch Model Fitting, Tutorial 1, and the bootstrap sections of Tutorial 3 Reproduce the supplied estimate and interval; then change seed, sample size, and noise, predicting the direction of change before running Refit the same family on the fixed Block 3 dataset (templates.md §11.11) → estimate-uncertainty-transfer.ipynb

Core circuits and populations

Neuromatch videos on spiking activity, LFP, stimulus representation Trace one stimulus or action through single-unit spikes, a population pattern, a field observation, and behaviour. At each step name the scale, the observable, the resolution, and one invalid inference Sketch two population trajectories with the same mean rate and different paths → POPULATION-DYNAMICS-MAP.md

Core thin slice 2: population spikes

Neuromatch v3.0.2 "What" models notebook: setup, section 1, section 2 visualisation, section 3 inter-spike intervals only Record the subset URL, hash, any session identifier the source exposes (write not supplied if none), and shape Before choosing examples, reserve one unit index with seed 20260731 in transfer-unit.txt and do not look at it until the done-when test Spike counts per unit, mean and median, distribution plot, rasters for three named units, ISI histograms for two, one ISI repeated at two bin widths → thin-slice-spikes.ipynb and its data card Treat the units as spike-sorted putative neurons nested in one session, not independent biological identities

Core EEG normal background

Three structured descriptions using the form in templates.md §11.7, worked from named figures in the AES normal EEG chapter: one awake adult, one drowsy, one light sleep. Record the figure number you used for each. Normal recordings first, so the form is automatic before anything unusual appears (eeg-lab.md section 4)

Core first figure audit

Audit Figure 2 of Steinmetz et al. using the rubric in templates.md §11.2 → FIGURE-AUDIT-1.md Send that audit to a contact who replied, with one concrete question. If nobody replied, use a knowledgeable colleague and log it as peer feedback, not expert validation

Done when

You can run and interpret the same estimate-with-uncertainty family on the fixed transfer case, reuse your unchanged raster function on the withheld seeded unit, and state the spike subset's preparation, selections, session nesting, and generalisation limit. Before excluding the top 1% of units by spike count, predict whether mean or median moves more, then check. Your three normal EEG descriptions are complete and use the form.

If ahead

Run only the first demonstration in Neural Rate Models and predict one parameter effect. Do not start dimensionality reduction or deep learning.

If compressed

Finish or repair the figure audit and record the exact next notebook cell. Do not open a second analysis family.

Validation, systems, cross-modal inference, and the Interest Gate

Load

Full

Focus

Leakage, modality comparison, home-base decision

Gate

Interest Gate

Compressed Paused

0%0/0 done

Why this block matters

Two things at once. You learn the most common computational validity failure, which is letting information from the same person or session sit on both sides of an evaluation. And you complete the breadth floor, so the Interest Gate is a decision made on evidence rather than on mood. Everything after this block is shaped by what you decide here.

Diagnostic first

Given a grouped dataset, identify the correct unit of splitting, build a leak-free pipeline with all transforms inside the folds, and explain why an apparently excellent score can be an artefact. If clean, compress the tutorial and go straight to the demonstration.

Core validation

Read scikit-learn cross-validation, pipelines, calibration The leakage demonstration (Part 4.3): same model, three splitting schemes, all three numbers reported, gap explained in writing → leakage-demo.ipynb Produce a calibration plot from held-out or out-of-fold group-aware predictions, never from in-sample fits, and write two sentences on why discrimination and calibration can disagree

Core thin slice 3 and cross-modal comparison

Use the pinned Neuromatch IBL behaviour notebook v3.0.2 as scaffolding and the current official Loading Trials Data page as executable authority Load session EID 4ecb5d24-f5cc-402c-be28-9d0f7cb14b3a. Verify the fields contrastLeft, contrastRight, choice, probabilityLeft, feedbackType Record the ONE client version, base URL, EID, retrieval date, and the resolved dataset revision and UUID or checksum where the client exposes them. A live revision is a teaching slice, not automatically a frozen release Before computing anything: define signed contrast as right-positive and left-negative using the installed official helper or its recorded formula, and record which raw choice value that helper treats as right Plot empirical P(right) by signed contrast, split by probabilityLeft, with counts and denominators. Report easy-trial accuracy and trial count Trials are nested within one session. They do not become independent mice or laboratories One cross-modal brief comparing EEG, spikes, behaviour, and fMRI, all four of which you have now actually inspected, on scale, resolution, invasiveness, what each supports, and what each cannot → MEASUREMENT-MATRIX.md (templates.md §11.2) Audit Figure 3 of the IBL standardised-behaviour paper. Contrast its across-mouse and across-laboratory evidence with your single-session description, and state why a psychometric curve does not uniquely identify a neural mechanism → FIGURE-AUDIT-2.md

Core a prepared fMRI derivative, strictly bounded

Watch the short Neuromatch fMRI signal video Open one prepared HCP task-fMRI example from the Neuromatch fMRI dataset guide, using the load_hcp_task_with_behaviour notebook. This data is already preprocessed and averaged across voxels into brain regions Reproduce one supplied visualisation, then change one meaningful choice: a task condition or a brain region Answer the same seven measurement questions you used for EEG and spikes Add the result to MEASUREMENT-MATRIX.md using the existing template. Create no new MRI template

The boundary. This is exposure to an analysis-ready fMRI derivative, not training in raw MRI preprocessing. Do not add FSL, fMRIPrep, registration, segmentation, another course, or a predetermined imaging project at this stage.

The visual-circuit reading has moved to Block 8’s reviewer-waiting period so this block does not grow.

Core EEG artefacts

Five artefact cards worked from named figures in the AES common artifacts chapter: eye movement, muscle, pulse, electrode pop, and sweat. Record the figure number for each. For mains interference, use a separate fixed example, either a real 50 Hz recording of your own or a synthetic 50/60 Hz contamination you generate, because the atlas figures do not supply one. Each card records prevention, confirmation, and handling: acquisition correction, annotation, rejection, channel repair, or a justified transformation, always preserving the raw data. Then mark contaminated intervals on three excerpts from your Block 2 recording. Egypt runs 50 Hz; CHB-MIT and most US recordings are 60 Hz, and a notch filter must match the source (eeg-lab.md section 5)

Core the gate

FIELD-MAP.md reaches 12+ substantive rows spanning three levels and three measurement types Write the Interest Gate paragraph: deepen EEG, broaden, or switch. Reference specific rows. Name what you are giving up

Done when

The leakage demonstration is complete and explained, all three thin slices are done, your six artefact cards each name a prevention, confirmation, and handling strategy, and the home-base decision is written and defensible.

If ahead

Repeat the leakage demonstration on a second dataset with a different grouping structure.

If compressed

Complete the leakage demonstration only. It is a dependency for everything after Gate A. The gate paragraph can slip into a floating slot.

Question triage and Gate A

Load

Standard

Focus

Overlap audit, feasibility, selection

Gate

Gate A

Compressed Paused

0%0/0 done

Why this block matters

This block decides whether the next eight are worth doing. Deliberately lighter on new content, because auditing is slow, and because a question that survives real scrutiny is worth more than one that arrives fast. Most of the work is reading and killing your own ideas.

Diagnostic first

None. Question selection is a research act and cannot be tested out of.

Core triage

QUESTIONS.md reaches 8 or more cards from at least three different sources (Part 2.2) Triage all of them to stage 2. Expect most to die here Promote at most two to finalist

Core audit

Run the full dated overlap audit (Part 2.3) on your leading candidate → overlap memo (templates.md §11.4). A second full audit only for a genuine tie Apply the differentiator test. One sentence, or the candidate dies Complete the feasibility screen, including a real count of independent units

Core EEG mimic contrasts

The two matched contrasts, one comparison card each (the card template has a column per side): wicket versus temporal epileptiform using AES Figures 48 and 62; RMTD versus evolving ictal using AES Figures 49 and 73. Each card records evidence for the variant, evidence for the mimic, evidence against each, residual uncertainty, and the displayed montage, filters, sensitivity, and time scale where known (eeg-lab.md section 6) Take the five-packet EEG transfer test in templates.md §11.11. Pass rule: at least four of five broad categories correct, including both benign-versus-pathology discriminators, with zero categorical diagnostic claims

Core commit

Write the intended-use statement and name one estimand Write the project-learning contract: one concept, one method, one implementation skill, one thing out of scope Set provisional contribution and authorship expectations with anyone already involved. The written agreement follows at Gate B Get one external reader to comment on scope. Silence documents an unsuccessful attempt; it does not pass Gate A on the research track. If no qualified reader responds, your options are: approach an alternate reader, use a floating slot and extend, narrow the question and claims, or continue on the learning track

Done when

Gate A has one of its three permitted outcomes recorded in writing (Part 2.6), with the memo, screen, intended use, estimand, and contract attached. The EEG transfer test has been passed on the first attempt or on the single retest. If the retest also failed, the interpretation dependency is recorded in writing and expert adjudication is required for all interpretation-sensitive claims from here on.

If ahead

Draft the first version of the protocol.

If compressed

Keep auditing the leading candidate. Do not select a question you have not audited because a slot is ending.

Project-specific learning and a development-only pilot

Load

Full

Focus

The next layer, on your own project

Gate

None

Compressed Paused

0%0/0 done

Why this block matters

This is where the project starts teaching you. The learning contract from Gate A names exactly what you need, and this block is when you get it: not a survey, but the specific concept, method, and implementation skill your question requires. The pilot exists to find out what breaks before anything expensive is built on top of it.

Diagnostic first

Attempt the contract's method on toy data before studying it. Anything you can already demonstrate compresses to a check rather than a course.

Core the contract

If your capstone is EEG or signal-based and the pilot exposed a demonstrated gap in filtering, spectral, or time-frequency reasoning: buy Cohen's signal processing course as the learning contract's method resource, not stacked on top of it (see eeg-lab.md section 3) Learn the named concept. One primary resource, one exercise Learn the named method. Predict what it should change before applying it to your data Acquire the named implementation skill Demonstrate all three on toy or development data → contract-demo.ipynb Write the "what this taught me" note and update the field map

Core pilot, development data only

Run the smallest end-to-end version on development data only. The final evaluation data stays untouched Confirm runtime and storage fit your machine Confirm any annotation conversion is visually correct on sampled events Confirm your comparator runs Write the data and annotation card for your capstone corpus (templates.md §11.5): version, provenance, licence, unit of independence, who annotated it, against what definition, and whether any agreement checking was reported (eeg-lab.md section 7 if the corpus is EEG) → pilot decision record (templates.md §11.6), listing every pilot result you have now seen

Done when

The contract skills are demonstrated on data you did not practise on, the pilot runs end to end, and every result you have already seen is written down so it can be disclosed at Gate B.

If ahead

Start the protocol draft.

If compressed

Finish the contract demonstration. The pilot can move; unlearned methods cannot.

Development workflow and draft protocol

Load

Standard

Focus

Reproducible pipeline, protocol draft

Gate

None

Compressed Paused

0%0/0 done

Why this block matters

You turn a working notebook into a workflow someone else could run, and you write down what you intend to do before you find out what happens. Both are what separate a result from an anecdote.

Diagnostic first

If you already write config-driven, restartable pipelines with logging, skip the workflow instruction and go to the protocol.

Core workflow

Move the pipeline out of notebooks into scripts with a config block at the top → preprocess.py, features.py or equivalents for your archetype Every parameter set in one place, not scattered inline Start DECISIONS.md: every analysis choice, with its reason, as you make it

Core protocol draft

Draft against the standard your design actually requires (Part 4.4) Specify: dataset versions and exclusions; population and setting; the unit of independence; preprocessing and missing-data policy; development and evaluation separation; threshold or selection rules; one primary estimand and its primary metric; uncertainty method accounting for clustering; every planned sensitivity analysis and its size; the fixed error-analysis sample; and a deviations log

Core book the reviewer

Identify and book your Gate B protocol reviewer now. Booking is Block 7 work, not Block 8 work, because reviewers have latency and Block 8 cannot proceed without one Reach a written authorship and contribution agreement before primary analysis, against the ICMJE criteria. Revisit it once actual contributions are known

Done when

The pipeline runs from configuration alone on a development subject you did not build it on, the protocol draft names every analysis you intend to run, and a named reviewer has agreed to read it.

If ahead

Identify and approach your external reviewer for Block 8.

If compressed

Keep DECISIONS.md current. It is the artifact that is hardest to reconstruct later.

External review, separation, and Gate B

Load

Standard

Focus

Review, registration, lock

Gate

Gate B

Compressed Paused

0%0/0 done

Why this block matters

Locking a protocol before you see the primary result is what makes the result mean something. It is also the point where an outside reader can still change the design cheaply. After this block, changes are deviations to be declared rather than choices to be made.

Diagnostic first

None. External review and registration are research acts.

Core review

Send the protocol to a domain or methods reviewer with one bounded request (templates.md §11.8) Silence does not pass Gate B. If no qualified reviewer responds, take one of: an alternate reviewer, a floating slot and extension, a narrowed question and claim, or the learning track Wait time is not idle: use it for the safe reading listed under If ahead

Core separation and lock

Refresh the overlap audit now, immediately before lock. The literature moved while you built the pipeline. A question answered last month changes what you should lock Confirm final-evaluation outcomes have been unavailable throughout development, or that your prespecified nested procedure has been followed correctly Disclose every pilot and development result already seen Register the protocol somewhere immutable and timestamped, such as OSF → gate record (templates.md §11.9)

While waiting for review the planned use of that time, not filler

Read selectively: Visual Processing, Eye and Retina and the retina, LGN, and V1 portions of Cortical Pathways Trace photon, retinal transformation, ganglion-cell output, thalamocortical input, cortical representation, adding resolution and causal status at each step → field-map rows

Done when

The protocol is registered, the pilot disclosure is written, and you have not looked at a primary result. A registration proves what you planned and when. It does not prove novelty.

If ahead

A dynamical-systems or domain-theory tutorial.

If compressed

Wait for review rather than locking early. Gate B is worth a slot.

Primary analysis

Load

Full

Focus

Execute exactly what was locked

Gate

None

Compressed Paused

0%0/0 done

Why this block matters

Run what you prespecified, in the order you prespecified it. The discipline here is the entire value of the previous block. Every deviation goes in the log with its reason, and deviations are not failures as long as they are declared.

Diagnostic first

None. This is execution.

Core

Write the analysis-freeze record before the first primary run: code and config version, dataset and version, the exact command, the environment, seed and nondeterminism policy, timestamp, and expected outputs Run the primary analysis exactly as locked Compute uncertainty accounting for clustering by the correct unit Log every deviation with its reason Draft the main figure and results table Motor-system bridge. Read Neuroscience Online: Motor Cortex selectively around cortical organisation, population coding and direction, and the difference in evidential weight between recording, stimulation, and lesion. Add one field-map row connecting intended movement, population activity, measurement, decoder, intervention, and inference

Done when

You can state the primary result in one sentence including its uncertainty, and every departure from the protocol is written down.

If ahead

Begin the robustness analyses that were prespecified in Block 7.

If compressed

Run the primary analysis only. Robustness moves to Block 10 where it belongs anyway.

Robustness, error analysis, and claim control

Load

Full

Focus

Stress the result, bound the claim

Gate

None

Compressed Paused

0%0/0 done

Why this block matters

A single number is not a finding. What makes it credible is showing how it moves under the perturbations you named in advance, and looking honestly at the cases it gets wrong. This is also where the claim gets sized to the evidence rather than to your hopes.

Core

Run only the sensitivity analyses listed and sized at Gate B. New ones are exploratory and labelled as such Error analysis on the fixed sample prespecified in Block 7. Look at what fails and why Declare as exploratory any subgroup analysis the independent-unit count cannot support Write the limitations section now, while the failures are fresh Bound the claim: what the evidence supports, and the nearest thing it does not Learning or decision bridge, choose one not both. Either Sections 7.1 to 7.3 of Learning and Memory, or the introduction and binary-state portion of Neuromatch Bayesian Decisions if that is closer to your capstone. Produce a mechanism, behaviour, and measurement chain plus one transfer prediction

Done when

The result has survived or failed its prespecified stress tests, the error sample has been examined, and your claim is sized to your independent units rather than your window count.

If ahead

Begin the report skeleton.

If compressed

Run the prespecified sensitivities. Skip exploratory work entirely.

Freeze, clean rerun, and Gate C

Load

Standard

Focus

Reproducibility

Gate

Gate C

Compressed Paused

0%0/0 done

Why this block matters

Work that does not rerun is the commonest reason competent research is rejected. If it does not reproduce for you in a clean environment, it will not reproduce for anyone else.

Diagnostic first

None. Reproduction is a research act.

Core

Build a fresh environment from your pinned specification Rerun the whole pipeline. Every reported number reproduces within declared tolerances Assemble the reproducibility package (Part 4.5): manifest, seeds, DECISIONS.md, README with a one-command path Where exact equality is not achievable across hardware or library versions, declare the tolerance rather than claiming bit-identical output → gate record

Done when

Someone else could clone the repository and regenerate your figures without asking you a question.

If ahead

Start the report.

If compressed

The rerun is the block. Everything else waits.

Full scientific report and the Release decision

Load

Full

Focus

Write it up, then decide honestly

Gate

Publication Release

Compressed Paused

0%0/0 done

Why this block matters

The report is the deliverable, and it exists whether or not anything is ever submitted. Writing it is also how you find out what you actually understand. The release decision comes after the report, not before, so that the writing is not bent toward a predetermined conclusion.

Diagnostic first

None. Writing is a research act.

Core the report

Write the full scientific report against the rubric in Part 6.5: question, background, methods, results, uncertainty, limitations, what remains unknown Refresh the overlap audit. The literature moved while you worked

Core external review of results

Obtain qualified review of the results and their interpretation, not just the protocol. Release cannot pass without it. This is scheduled work with external latency, so request it at the start of the block

Core the release decision

Apply the Publication Release criteria (Part 6.2) Record the decision and its reasons in the gate record, whichever way it goes

Done when

The report is complete and the release decision is written down with reasons.

If ahead

Begin manuscript conversion.

If compressed

Finish the report. The release decision can move to Block 13.

Execute the decision, close out, then consider a stretch

Load

Standard

Focus

Ship, archive, and plan forward

Gate

None

Compressed Paused

0%0/0 done

Why this block matters

Two jobs. Convert the decision into an action, and leave the work in a state your future self can pick up. Stretch work is considered only after both are done and only if the calendar genuinely allows.

Core execute

If Release passed: convert the report to a manuscript for one named venue, agree authorship with all contributors, post a preprint where appropriate, submit If Release did not pass: complete the archive and record precisely why, which is a legitimate outcome and not a failure

Core close out

Deposit the lawful reproducibility archive with a DOI (Zenodo). The archive is always produced, whichever way Release went. It is an artifact, not a publication Write the next-question dossier from what this project actually taught you Update PEOPLE-ACCESS.md with every relationship that now exists Audit every repository: does each still have a job?

Core optional and conditional: stretch, only if both floating slots are intact

Choose exactly one: an independently defensible extension that passes its own current audit; a bounded evidence-synthesis scope and throughput pilot; a collaborator-led contribution with a real role; or additional dissemination of existing work Pipeline reuse is a cost advantage. It is not a scientific justification, and it never by itself makes an extension worth doing

Done when

The release decision has been acted on, the archive exists, and the next-question dossier is written.

If compressed

Archive and dossier only. The stretch is the first thing to go and should go without regret.

06 The capstone and scholarship Part 6

6.1 Two tracks, both successful

Track	When	Endpoint
Research capstone	Gate A found a material, feasible question with a review route	Full report, plus manuscript if Release passes
Structured learning capstone	No question survived, but a rigorous bounded analysis exists	Full report, explicitly making no novelty claim

Both produce a complete scientific report, a reproducibility archive, and a next-question dossier. The learning capstone omits only the novelty claim and the manuscript. It is a real outcome and should be described accurately rather than apologetically.

6.2 Capstone Completion is not Publication Release

Two separate decisions. Conflating them is how educational analyses end up impersonating publishable research.

Capstone Completion asks: is the work rigorous, reproducible, honestly reported, and complete? If yes, the apprenticeship succeeded.

Publication Release asks all of the following, and every one must hold:

Criterion	Fails if
Material contribution	A refreshed audit shows the question is answered
Sufficient evidence	Independent units cannot support the claim
External scientific review	Nobody qualified has examined the design and interpretation
Reproducibility	Gate C did not pass cleanly
Lawful release	Data agreements or ethics do not permit it
Authorship agreement	Contributors have not agreed roles and order

If Completion passes and Release fails, the roadmap succeeded. Say so plainly.

6.3 Accurate states

Four different things, never collapsed:

State	Means
In preparation	Being written
Submitted	Sent to a venue, decision pending
Accepted	Editorial acceptance received
Published	Publicly available in final form

A preprint has a persistent identifier and is not peer reviewed. A Zenodo archive is a reproducibility artifact, not a paper. Journal acceptance and indexing are external events on someone else's timetable.

On software: the Journal of Open Source Software requires software developed privately to have at least six months of public development history before submission, plus releases, public issues, tests, documentation, and feature-complete scope; single-purpose analysis pipelines and thin utilities are explicitly out of scope. A capstone pipeline built inside this roadmap will not qualify. Archive it on Zenodo, cite it, and stop there.

6.4 The structured learning track

If Gate A selects this route, Blocks 6 to 13 keep the same shape with three substitutions: the question becomes a well-posed analytic exercise on public data; the overlap audit becomes a positioning statement acknowledging existing work; and the report states explicitly that no novelty claim is made. Everything else, including protocol lock, prespecified analysis, robustness, and reproduction, stays identical. The rigour is the point.

6.5 Report rubric

Section	Must contain
Question	The estimand in one sentence, and why it matters
Background	What is known, what is not, where this sits
Data	Corpus, version, provenance, units, exclusions, limitations
Methods	Everything needed to rerun, plus the protocol reference and every deviation
Results	The primary estimate with clustering-aware uncertainty, then secondary and clearly labelled exploratory analyses
Robustness	Prespecified sensitivities and the error analysis
Limitations	Reference-standard quality, independence, generalisation, what would change the conclusion
Conclusion	Sized to the evidence, naming the nearest thing the data do not support
Reproducibility	Environment, manifest, seeds, tolerances, archive location

6.6 Authorship

Agree roles and order in writing before analysis, against the ICMJE criteria. Authorship disputes on student-led work are common and almost entirely preventable. If someone reviews your protocol, be explicit about whether that is a courtesy, a supervisory role, or authorship.

07 Mentorship, access, and programmes Part 7

7.1 Mentorship is a workstream, not a wish

Gate A requires an external reader. You cannot gate on a relationship you have not built, so outreach starts in Block 1 and continues throughout.

Block	Outreach action
1	Map 8 to 12 people. Profiles for all; read properly for the best three
2	Send the first three messages: one bounded artifact, one answerable question
3	Share the figure audit with anyone who replied
4	Second wave, informed by your Interest Gate decision
5	Ask your best contact to comment on Gate A scope
8	Request protocol review
12	Request report or manuscript review
13	Update the map with every relationship that now exists

What works: a specific named paper of theirs, a specific artifact of yours, one answerable question, and a bounded ask. What does not: generic mass mail, unbounded requests for supervision, or five attachments.

Silence is data, not a pass. Log non-responses. Silence documents that you tried; it does not satisfy a gate on the research track. If nobody qualified responds by Block 5, your permitted options are: approach alternates, spend a floating slot, narrow the question and its claims to what unreviewed work can support, or take the learning track. A knowledgeable colleague’s comments are peer feedback and must be labelled as such, not as expert review.

7.2 Access is scientific feasibility

Data access belongs in the feasibility screen, not in an administrative appendix. Requests with approval steps start in Block 1 if they might be needed. A dataset you cannot get is not a limitation to work around; it is a question you cannot ask yet.

7.3 Programmes substitute, they never stack

Structured programmes are accelerators, not dependencies. All Neuromatch materials are open, so the curriculum is yours regardless of admission. Arabs in Neuroscience runs an Arabic-language computational neuroscience course with open materials and an in-person school via the IBRO portal; its paid teaching-assistant role, which carries a certificate and a possible reference letter, is the most valuable thing either organisation offers you, and your existing biostatistics teaching makes you a credible candidate.

If admitted to a live intensive, that period becomes programme-only. The programme's curriculum replaces planned curriculum, its group project replaces planned project work, and the affected blocks move. It is never additional. Write the substitution memo (templates.md §11.10).

Dates, fees, formats, and application windows change. Check the official page when the action arrives, never this document.

7.4 Conferences

Build a venue table from official calls when results make dissemination plausible, which is Block 12 or later, not as a monthly ritual from Block 1. Never copy a historical deadline forward as fact. Confirm presenter requirements before submitting, since several major meetings require onsite presentation and an abstract you cannot present is a smaller asset than it appears.

Given that results arrive late in the roadmap, the realistic first presentation is a regional meeting after the plan ends. That is normal and worth saying rather than pretending otherwise. Only submit to meetings run by an actual professional society or university.

08 Operating rules Part 8

8.1 Four modes

Triggered by circumstances, not scheduled in advance. You do not need to know your exam dates for this to work.

Mode	When	What you do
Normal	Ordinary weeks	The block as written
Compressed	Dedicated exam study, heavy rotation, or life	Reading, retrieval, and writing, plus finishing or repairing an already-defined technical workflow. Never a new toolchain, a new exploratory branch, or a high-stakes analysis. If conditions do not permit careful execution, defer the task rather than doing it badly. See the block's if compressed line
Paused	No capacity	Nothing. Log the date in `PROGRESS.md`
Re-entry	First one or two sessions back	Read `PROGRESS.md`, rerun the last command that worked, do the smallest next artifact. Do not start something new

8.2 The two floating slots

Two of the fifteen calendar slots hold no mandatory content. They exist to be consumed by disruption. Using them is the system working as designed, not falling behind.

8.3 Rebaseline, never catch up

After any compressed or paused slot, do one thing: count remaining blocks against remaining slots.

Blocks fit: continue. Change nothing.
Blocks exceed slots: either move the target date or drop optional scope. State which, in writing.

Never stack extra hours to recover a slot. Never compress a research act to protect a date.

8.4 Fast learning: demonstrate, then compress

Speed is handled by evidence rather than assertion. Every block opens with a diagnostic. Pass it and the instruction compresses or disappears.

Time saved goes to depth first: a harder transfer case, an extra figure audit, more cross-modal exposure, or better project work. It does not automatically become another output.

8.5 What cannot be tested out of

These are not instruction, so no diagnostic can skip them:

inspecting the real data yourself;
locking a protocol before results;
validation and leakage control;
debugging;
reproduction from a clean environment;
obtaining external review;
writing the report.

Learning speed shortens the time to understand a method. It does not shorten data access, reviewer response, collaborator decisions, or journal processes.

8.6 Substitutions and maintenance

Links rot, datasets change, programmes move. When a resource has changed at the moment you need it: record what changed, find a lawful equivalent serving the same learning job, and note the substitution. Never silently swap a dataset, and always record the version you actually used.

8.7 Project stop rules

Trigger	Action
No question survives Gate A	Take the learning capstone track. Do not force a novelty claim
Independent units cannot support the estimand	Narrow the claim or stop at Gate A or B
The pilot shows the analysis will not fit	Reshape at Gate A rather than discovering it at Block 10
Access fails after Gate A	Substitute a lawful dataset, or switch archetypes with a written memo
The result is null	Report it. A well-conducted null with a locked protocol is a real finding. Nonsignificance does not establish absence: claiming a negligible or absent effect requires a prespecified equivalence-type claim or an interval that excludes effects you would consider meaningfully large
Six months pass with no shipped artifact	Stop learning and force a small complete piece of work

09 Risk register Part 9

Risk	What absorbs it
No reviewer responds	Log the silence, use peer feedback with the limitation named, or change route. Silence never passes a gate by default
Access route changes or closes	Recheck at use, record the version, substitute lawfully. Never silently swap
No material question survives Gate A	Named rigorous learning capstone without a novelty claim
Sample cannot support the estimand	Narrow the claim, or stop at Gate A or B
Result is null or non-novel	Preserve protocol and report. Release may fail while the apprenticeship succeeds
Coauthor or journal timing exceeds the calendar	Record the accurate state and extend. Never call preparation "submitted"
USMLE or clinical work consumes capacity	Switch mode, consume a floating slot, or move the date. No catch-up debt
Learning runs faster than planned	Diagnostics compress instruction; saved time goes to depth, not to extra outputs
Tutorial hell	Every block ends in an artifact plus a completion test appropriate to its stage: a transfer demonstration early, a locked decision or reproduction later. Six months without a shipped artifact triggers the stop rule
Scope creep after Gate B	Only prespecified sensitivities are confirmatory. Everything else is labelled exploratory
One project consumes the whole calendar	Stop rules and the Gate A feasibility screen exist precisely to prevent this

· Sources

Neuroscience. Neuroscience Online, the open-access electronic textbook from the Department of Neurobiology and Anatomy, McGovern Medical School at UTHealth Houston. Neuromatch Computational Neuroscience, open curriculum.

EEG. See eeg-lab.md for the full source list, built on the American Epilepsy Society atlas on NCBI Bookshelf and the ILAE 2025 seizure classification.

Standards. EQUATOR Network hosts all reporting guidelines. TRIPOD+AI: Collins GS, Moons KGM, Dhiman P, et al. BMJ. 2024;385:e078378. STARD-AI. PROBAST+AI. QUADAS-3.

Data. PhysioNet · CHB-MIT · TUH EEG including TUSZ · OpenNeuro · IBL

Tools. MNE-Python · scikit-learn · Zenodo · OSF

Signal processing. Mike X Cohen's free lectures. One conditional paid course, decided at Gate A: see eeg-lab.md section 3.

Worked examples of finished questions in the EEG constraint family. Lin YC, Lin HA, Chang ML, Lin SF. Neurophysiol Clin. 2025;55(2):103044. DOI. Kojima J, Shi H, Ojemann WKS, et al. medRxiv preprint, 2026. DOI.

All facts checked 31 July 2026. Recheck anything date-sensitive when you use it.