AI Safety.

The course's spine is not a topic list — it is a verification discipline. Every external fact in this repository lives in the source register with an access level recording how far into the source this build actually read, and a retrieval date. A retrieval date is not a publication date; an abstract-level read is not a full-text read; a page-inspection is not a read at all. The ladder below is the whole corpus, counted honestly at each level.

Red flags the register is built to catch: a number with no denominator · a benchmark quoted as a construct it does not measure · "peer-reviewed" attached to a preprint · a vendor's self-evaluation labeled an independent audit · a link that was never fetched. The re-check rotation re-fetches a sample of register URLs on a schedule recorded in MAINTENANCE.md — two rotations so far, one genuinely dead link caught and corrected.

The catalog below is the whole course in reading order — grouped the way the dependency map draws it. Chapters 01–02 are the common methodological spine every route shares; the three specialization lines follow; the synthesis closes. Hard prerequisites are enforced, not aspirational: 02 requires 01, 13 requires 12, 22 requires 01 and 08. Each chapter directory carries exercises, hints, worked solutions or rubrics, and 2–4 primary readings verified in the register at the stated access level.

How chapters are assessed: two demanding exercises each (one analysis/derivation, one empirical/design/critical), hints and worked solutions published separately from the prompts, and a concrete completion artifact — the column at right — with explicit completion criteria. "Done" is defined by artifacts, never by weeks elapsed.

Each project is a bounded research learning unit with a CPU smoke mode: no GPUs, no API keys, no network. The italic line under each title is the project's own honest headline — its pinned result or verdict, quoted from the project README, including the seams where compute was out of scope. Run the whole suite with uv sync && uv run pytest; the deterministic numbers must reproduce on your machine before anything else.

Paper notes are written at the level the register's access supports — method-level analysis for full-text reads, abstract-level entries say so — and every note ends in a falsifiable extension runnable at the course's deterministic tier: a claim you could imagine being wrong, with the observation that would weaken it. Comparison notes settle practitioner decisions single papers leave open. The fastest way in is through the chapter you are already reading: chapters name their note companions in their readings sections.

Every incident below was reviewed with the strict template: observable facts first, established conclusions attributed, unknowns kept unknown, and every what this does NOT demonstrate written down. First-party reports outrank press coverage; where only press coverage exists, the note says so. Each feeds the same control family — enforceable output boundaries plus release-gated evaluation — and the accountability era adds the third leg: audit the training data and the deployed behavior before the regulator or the press does it for you.

Every card in the catalog is a real-world, verifiable dataset — no synthetic stand-ins. Each traces its composition numbers to the URL they were collected from, with license as observed and contamination status. The three the projects lean on most: TruthfulQA, AdvBench, and XSTest. The machine-readable catalog lives at datasets/catalog.yaml.

This is the register the whole course stands on: papers, reports, standards, and pages — each with the access level actually achieved and the date it was fetched. Filter by level to see the shape of the evidence honestly: what was read in full, what was read at abstract level, what was only inspected, and what could not be reached at all. The machine-readable original: research/source-register.yaml.

A claim enters this register only with a source, a strength grade, and room for counterevidence — so a strong claim and a plausible one never wear the same badge. This is the discipline Chapter 01 teaches, applied to the course itself. The machine-readable original: research/claims-register.yaml.

Terms are added with the chapters that need them — the glossary grows as the course builds. Each entry carries the definition, the notation or aliases, the confusion it is most commonly lost to, and the chapters that use it. Sources are cited where a definition follows a specific document; otherwise the definition is this course's own operationalization.

The three role families below are mapped to the live job market — dated demand signals, entry mechanisms, and a five-rung progress ladder — in roadmaps/careers-index.md. Career-market data is perishable by nature: re-verify before deciding, and treat any vendor-curriculum mapping with the caveat it deserves. The currency throughout is portfolio artifacts, not hours or certificates.

For learners who want the research-engineering base rebuilt from first principles: seven modules, standard library only, each with worked code and the tests that pin it. The entry point is research/ai-researcher-from-scratch/README.md.

Make the first 90 days concrete: roadmaps/first-90-days.md — an artifact every Friday, twelve weeks, no heroics.

Four artifact-gated trackers live in roadmaps/progress/ — one per route. Every box is a named artifact with its evidence wired to this repository's material, and each milestone carries self-audit questions. The bars below are parsed from those files directly; they count checkboxes, and a box only ticks when the artifact exists.

The full map is START_HERE.md; the diagnostic is the honest entry. The sequence below is the short version — each step names what it costs and what it produces.

Chapter hours are medians for a learner who passed the diagnostic; project hours are for the smoke or pilot tier. Adaptation rules live in roadmaps/study-schedules.md — they tell you what to cut when life interferes.

If your question is not here, the START_HERE map or this page's table of contents is the next stop.

Each line below says what the source is for — the load-bearing idea a chapter or project takes from it — not a generic summary. The bracketed ID is the audit trail: it resolves to the source register, which records exactly what was fetched, when, and how far into it this course actually read. Where an honest access level matters (vendor white paper, self-reported program data), the annotation says so. The full original: AWESOME.md.