Spokes.wiki Search About

Synthesis — Engineering Education

The evolving thesis. Spun out 2026-08-04 from three of the most-starred educational repositories in software: coding-interview-university (~352k★), awesome-system-design-resources (40.2k★) and ai-engineering-from-scratch (~33k★). nirdiamant-prompt-engineering (7.8k★) joined on 2026-08-05. Small corpus, and the shape is provisional.

Current thesis

Two of the three are shaped by an exam, and the third is the control.

coding-interview-university and awesome-system-design-resources are both organized around what a hiring process tests. That is not a criticism, it is their stated purpose — one is a multi-month plan for passing coding interviews, the other ships "How to Answer a System Design Interview Problem" alongside 60+ graded practice problems. What they teach is downstream of what is assessed (interview-preparation), and the scope decisions show it: coding-interview-university explicitly excludes frontend, full-stack and SQL, which is a strange omission from a computer science education and an obvious one from an interview plan.

ai-engineering-from-scratch has no exam behind it — there is no standardized AI-engineering interview to prepare for — and it is built differently. Its pedagogy is build-it → use-it → ship-it: implement the thing by hand, then use the framework equivalent, then produce a reusable artifact, across 17 capstones. It is the only one of the three that produces something at the end rather than readiness to be questioned.

The correlation is one observation on three sources and could easily be coincidence. What makes it worth stating is that it is testable: the next curriculum to arrive either has an assessment regime behind it or does not, and either teaches by curation or by construction.

The fifth source was the real test, and the correlation did not survive it. developer-roadmap is the case the paragraph below asks for: a curated-pointer curriculum with no assessment behind it. There is no standardized DevOps or product-manager interview the way there is a coding interview, and the project teaches nothing itself — it is a graph of topics pointing at other people's documentation. Under the thesis it should have been build-it-by-hand. It curates.

What the exam actually predicted, looking back, was scope rather than pedagogy: coding-interview-university drops frontend and SQL because the interview does. developer-roadmap's scope comes from somewhere else entirely — job titles, one roadmap per role a company hires for. That is a third shaping force beside the exam and the discipline, and it explains a shape the first four could not: a curriculum can be organized by what the labour market names without being organized by what it tests.

So the standing claim narrows to something still useful and no longer clean: where an assessment regime exists, it shapes scope. Pedagogy is chosen on other grounds.

The fourth source landed on the predicted side, and it barely counts as a test. nirdiamant-prompt-engineering has no exam behind it and teaches constructively — as predicted. But its subject is prompt engineering, a technique domain with no hiring pipeline, so the missing exam was settled by the subject before the author decided anything. A real test needs a case that could have gone the other way: a curated-pointer curriculum with no assessment behind it, or a build-it-by-hand one aimed squarely at an interview. Until one arrives the correlation stands at three informative data points and one free one.

Three forms, and each rots differently

The corpus now holds three ways of building a curriculum. A meta-curriculum points at other people's material. ai-engineering-from-scratch writes original lessons that ask you to build the thing by hand. nirdiamant-prompt-engineering is an executable-curriculum: the lesson is a notebook that already runs, and the reader edits it.

Sorting them by what breaks is more useful than sorting them by effort. Curated pointers suffer link rot, which is loud — a dead link announces itself. Executable lessons suffer dependency drift, which is quiet: a pinned library moves or a model behind an API is retrained, and the notebook may keep running while teaching something that has stopped being true. Build-from-scratch instructions barely decay at all, because they depend on nobody's service staying up. That ordering is the reverse of the ordering by how quickly a reader gets a result, which is the trade the three forms are actually making.

Two more forms landed on 2026-08-09, and the decay ordering held. prompt-engineering-guide is a reference: written lessons and a Prompt Hub indexed by the task the reader arrives with, with no sequence and nothing recording where anyone is. adrianhajdin-uber-clone is a clone-project: one commercial application rebuilt end to end, the build order standing in for a syllabus.

Both slot into the rot ordering, and the clone goes to the top. A reference decays on two clocks — the technique pages age slowly (chain-of-thought is still chain-of-thought) while the per-model pages age fast, and DAIR.AI's shelf of GPT-4, Phi-2, Mixtral, Flan and OLMo is being consulted in 2026. A clone decays fastest of anything here, because its dependencies are five commercial services with independent release cycles: a changed free tier or a redesigned console breaks a step in a video that cannot be edited. Link rot takes years; a vendor's onboarding redesign takes a weekend.

And "build it" turns out to name two opposite things. ai-engineering-from-scratch implements by hand before touching a framework; the Uber clone implements nothing and integrates Clerk, Stripe, Neon, Google Maps and Geoapify. Both are construction, and the vocabulary of curation-versus- construction cannot tell them apart. The real split is from scratch against from services — one leaves the learner understanding a mechanism, the other leaves them having shipped a product, and neither prepares for the question the other answers.

The from-scratch half finally has a page, and a second source (2026-08-17)

The section above named the split — from scratch against from services — and the corpus paged the services half (clone-project) while leaving the other half as a property of one outlier. 539kernel-book is the second instance and build-from-scratch now holds the form.

The distinction the second source adds is cumulative against modular. An executable-curriculum is a set of independent notebooks ordered by a README; from-scratch material is a single artifact where chapter five will not run without chapter four. 539kernel makes that visible in its own repository: evolution_by_versions/ keeps the intermediate kernels as named states (Version T after process management, G after paging, NE after the filesystem) so a reader whose copy breaks can rejoin at a chapter boundary. That directory is a failure mode made into infrastructure, and no modular curriculum here needs one.

It also confirms the decay ordering from the other end. The book is from November 2022, has never been revised, and has aged almost perfectly, because its entire dependency list is GCC 7.5, NASM, GNU Make and QEMU — no service, no API, no free tier, no console to redesign. Set against the clone project, whose five commercial dependencies can break a step in an uneditable video over a weekend, the two forms are three years and a world apart on the same shelf.

A curriculum that ends in the literature, not in an exam (2026-08-17)

The spoke's founding claim is that what gets taught is downstream of what gets tested (interview-preparation). Every source has fitted it, including the ones that stretched it: developer-roadmap substituted a job title for an exam, kernel-exploit-dojo is its exam. 539kernel-book is the first that does not fit at all.

Its final chapter, after the kernel boots, walks through five research papers — in-process isolation and the Endokernel, the nested kernel, the multikernel, dynamic reconfiguration via THINK's Fractal component model, and unikernels, closing on HermiTux. The author names the two exits: a traditionalist implements the known concepts, a kernelist "searches for other better solutions for the given problem and designs an operating system kernel that uses these solutions," and finds them in papers because "most probably, these ideas haven't been implemented or adopted by others yet." The stated goal of the chapter is to send readers into the literature to implement what is in it.

So the terminal state is production of new work, not assessment. That is a third shaping force beside the exam and the job title, and it is the first one in this corpus that points away from employment. Two readings are available and one source cannot separate them: either systems programming has no hiring funnel dense enough to bend a curriculum toward it, or an author with nothing to sell has no reason to promise a job. The second reading is supported by the same source being the corpus's second publisher with no commercial layer — the pattern the spoke tracked at 3/3 is now 3/5, with mito753 and mohammed-q-hussain on the other side, and both breaks come from authors whose subject is systems security or systems programming.

And the artifact was built for the book. Hussain wrote 539kernel in order to have something to teach, asking himself "how to make it a useful project for people?" before starting. Every other constructed source here is a project that later acquired teaching material. This is the cleanest case in the corpus of pedagogy choosing scope, since the kernel has no user and no requirement except being explicable — which is also why "really simple" is a design statement here rather than an apology.

The gap it does not close. No exercises, no problems, no assessment beyond whether the kernel boots, no revision since 2022, and no evidence of a single reader finishing. T3, like every self-published curriculum here. The one signal no other source in this spoke carries is a pre-publication read by an academic (Dr. Hussain Almohri), which is not peer review and is more than nothing.

The pedagogy claim finally has a controlled comparison

The founding correlation narrowed, over five sources, to: where an assessment regime exists it shapes scope; pedagogy is chosen on other grounds. Every data point until now was consistent with it and none was controlled — the subjects differed, the authors differed, the exams differed.

prompt-engineering-guide and nirdiamant-prompt-engineering are the first pair that holds the variables still. Same subject (prompting technique), same absence of a hiring pipeline, same self-published-repo genre, published within the same era. One chose 22 runnable notebooks; the other chose written explanation with a lookup table. Same conditions, opposite form.

That is as close to a natural experiment as this corpus is likely to get, and it settles the negative half of the claim: the assessment regime does not determine pedagogy, because here it is held constant and pedagogy varies anyway. What does determine it stays unanswered — author preference, subject maturity, and what the audience will actually read are all live candidates, and the 10× star gap between the two (77.4k against 7.8k) hints at the third without evidencing it.

The commercial pattern also stopped being about individuals. Every source in the corpus has something to sell beside the free artifact; until now the seller was always a person with a book, a course or a site. dair-ai is an organization with an Academy, corporate training and a consulting line. The one publisher here with no evidenced commercial layer is javascript-mastery, and the honest statement is that the repository simply does not show one — not that none exists.

The exam itself arrived, not just curricula shaped by it

llm-interview-questions-hoang is a different kind of object from everything above: not a curriculum but a bare assessment — 50 LLM interview questions with model answers, no path, no capstone, nothing to build. Every prior source lets you infer the exam behind the syllabus; this one is the exam, printed.

It strengthens the founding thesis by removing the inference step. The claim was "syllabus follows assessment," argued from what curricula include and omit. A pure question bank, self-published and promoted for interview prep, is that assessment with the curriculum stripped away — the same regime, stated directly (interview-preparation).

Two smaller things it fixes. It is the corpus's first source on the AI/LLM-engineering interview (the founding three cover CS fundamentals and system design), and it shows what that interview rewards today: recall of mechanism — define tokenization, distinguish LoRA from QLoRA — the flashcard end of the pedagogy axis whose opposite pole is ai-engineering-from-scratch's build-it-by-hand. It does not touch the corpus's central hole: still no outcome data, and a question bank says even less than a curriculum about whether answering it predicts doing the job.

The dominant form is the meta-curriculum

Two of the three teach almost nothing themselves. They are structured collections of links to other people's blog posts, videos, papers and engineering write-ups — awesome-system-design-resources aggregates from Discord, Netflix and Stripe engineering among others; coding-interview-university describes itself as curated pointers.

That form has a property worth naming early: its quality is unusually hard to assess. A textbook can be reviewed. A list of 400 links cannot, practically, and its value depends on curation judgement that leaves no visible trace — you cannot see what was rejected. Popularity is the only signal on offer and popularity here is enormous, which is exactly the condition under which it is least informative.

The disposition the whole corpus assumes, examined for the first time

Every artifact here is built for a learner working alone, by choice, without an instructor. That learner has a name in the education literature — self-directed — and until 2026-08-06 this spoke held not one source that had looked at the disposition itself rather than sold something to it. self-directed-disposition-study (ITiCSE 2025, six researchers, 496 reflections from 264 students at four institutions) is that source, and it complicates the corpus from two directions at once.

What students mean by self-directed is mostly "go find other material." Of 380 coded behaviours, utilizing external resources accounts for 148 — nearly four times the next category, and more than planning ahead (23), self-assessment (39) and reviewing against expectations (3) put together. The meta-curriculum form is that single behaviour productized: someone else has already done the finding. Whether that is a service to the learner or a removal of the one thing they were practising is not something either the study or this corpus can currently answer, and it is the sharpest question the spoke has.

The biggest reported obstacle to self-direction is assignment structure. Twelve of the 27 not-self-directed responses blamed the shape of the task — "there is really no room to add new tools and techniques that were not discussed" — ahead of unsuccessful effort (7), self-sufficiency (4) and motivation (3). A curriculum in this corpus is an assignment structure with the grade taken off. The finding does not transfer automatically, and the study cannot test it: its setting is a classroom with a deadline. But it puts a mechanism under a suspicion the spoke had no way to voice, and it means the executable-curriculum end of the corpus — the most structured form here — is the form the evidence would predict least self-direction from, not most.

The authors' own resolution is a tension rather than a recommendation: loosening structure to invite autonomy will increase the second-largest impediment, unsuccessful effort. Their answer is scaffolding that is deliberately faded, which no curriculum in this corpus does — a roadmap is the same density at week 1 and week 40.

What the corpus cannot say

Whether any of this works. Four sources, ~433k combined stars, and zero evidence of learning outcomes — no completion rates, no pass rates, no follow-up on anyone who finished. The star counts measure intent to learn, which is not the same as learning and is probably not even correlated with finishing. Nothing here should be read as a claim that these curricula are effective, only that they are what a very large number of people reach for.

Three of the four maintainers sell something adjacent — ashishps1 runs the AlgoMaster platform, nir-diamant has a book and a newsletter above 50,000 subscribers — which now looks like the normal condition of the genre rather than a fact about particular authors. nirdiamant-prompt-engineering also carries a custom non-commercial licence, the most restrictive terms in the corpus and a reminder that "open educational repository" covers a wide range of actual permissions.

The designed curriculum finally arrives

Every source here until now grew by accretion: a repository, a star count, a maintainer's judgment about what belongs. cs2023 is the other kind of object — the ACM/IEEE-CS/AAAI undergraduate guidelines, argued to by a joint task force and endorsed by three societies in January and February 2024. Having one of each lets the corpus say something it could not say with repositories alone.

The repository corpus is popularity-selected, and the selection is visible. CS2023 makes Society, Ethics and the Profession and Human-Computer Interaction first-class knowledge areas. Neither is meaningfully present in developer-roadmap, coding-interview-university or awesome-system-design-resources, which sort toward what gets someone hired. That is not a failing of the repositories; it is what optimizing for interview-preparation does to scope. The guidelines are the control that makes the effect measurable rather than asserted.

And the traffic runs both ways. CS2023 has no counterpart to executable-curriculum — material you run rather than read — and says nothing about the self-direction self-directed-disposition-study measures and finds unevenly distributed. A curriculum specifies content; the repositories specify a practice. The corpus has been treating them as competing answers to one question and they are answers to two.

What this does not settle is the spoke's standing complaint above. CS2023 is a specification, not evidence: it reports no outcomes either. Adding it moves the corpus from four popularity-selected sources with no outcome data to five sources with no outcome data, one of which was designed rather than accreted. The whether-any-of-this-works question is untouched.

One limit worth keeping visible: the 400-page final report would not extract here, so what this wiki holds is the knowledge-area structure and the endorsements, not the CS Core / KA Core tiering or the hour counts. Those are the part that actually constrains a department, and they remain unread.

The exam with no curriculum in between

kernel-exploit-dojo is the standing claim — where an assessment regime exists, it shapes scope — at its limit. Every earlier case reached the conclusion through an author: coding-interview-university drops frontend and SQL because the interview does, but jwasham still wrote a study plan. This archive skips that step. Its scope is not shaped by the exam, it is the exam: 100+ past CTF problems kept with the files the organizers distributed, a working exploit and a writeup, indexed by bug class, primitive, technique, difficulty and solve count.

So the claim survives its most extreme test, and what it produces there is a problem-archive — a fifth form, and the first that is neither authored nor merely a set of pointers. The maintainer's contribution is the classification and the grading.

A second assessment regime, and it is a better instrument. Everything the spoke has concluded about exams came from one kind of exam, whose properties went unexamined because there was nothing to contrast them with. A CTF is public where an interview is private, scored per problem where an interview is pass/fail, and uses each problem once before releasing it. Those last two are why an archive of this shape can exist at all in security and cannot exist for the hiring interview that shapes most of this corpus. The form is not rare here because it is unpopular; it is rare because the dominant exam keeps its problems.

The corpus gets its first number that is not marketing. The standing caveat on every source has been that stars measure intent to learn rather than learning. This archive carries per-challenge difficulty grades and solve counts, generated by competitions rather than by the maintainer. That is a signal about the item — how many teams landed the problem under time pressure — not about the learner, and it is inherited rather than measured here, so its comparability across events is unestablished. It still moves the corpus off zero on the axis that has been empty since founding, and it does so from the regime that publishes scoreboards. If outcome evidence ever arrives in this spoke, the shape of this source suggests where to look for it.

What has not changed. No sequence, no progress tracking, no license, and no outcome data. Coverage is decided by what got competed on, which is the popularity selection cs2023 exposed in the repository curricula arriving through a different door. And the archive's own decay mode is new to the corpus: not link rot or dependency drift but relevance, as the bug classes it drills get patched out of the target.

The regime with a sanction behind it

cs2023 gave the corpus a designed curriculum against a shelf of accreted repositories, and the comparison did real work — it exposed the repositories' popularity selection. What it could not do is compel anything. Three societies endorsed it; no department loses anything by ignoring all 270 CS Core hours.

abet-computing-criteria is the version with consequences. It is an assessment regime of a kind this spoke had not held: it examines the program rather than the candidate, and withdrawal of accreditation is a real cost. So the founding claim — where an assessment regime exists, it shapes scope — gets its strongest instance, and the scope it produces is the one the repositories drop. Ethics and professional responsibility, teamwork, and communication are not encouraged there; they are outcomes a program must demonstrate its graduates attain. developer-roadmap, coding-interview-university and awesome-system-design-resources carry none of the three, and that gap is no longer a repository failing to match a recommendation. It is a repository omitting what an accredited degree is required to contain.

The two prescriptive documents will not add up, and that is worth stating plainly. CS2023 measures in topic hours, ABET in semester credit hours. 270 and 40 are not comparable quantities, neither document maps its unit onto the other's, and no arithmetic in this wiki should pretend otherwise. What survives the incommensurability is a shared shape: both fix a floor of required computing and mathematics, and both refuse to name courses. ABET says so outright — "the curriculum requirements specify topics, but do not prescribe specific courses." Two independent bodies, drafting decades apart for different purposes, both stopped at the same boundary between what a curriculum must contain and how a department may arrange it.

And the corpus's hole survives them both. ABET has bound US computing education for decades and reports no outcomes: no evidence that graduates of programs meeting its hours do better than graduates of programs that do not. The does-any-of-this-work question now has two prescriptive documents and one popularity-selected shelf standing over it, and still nothing measuring a result.

The three forms are axes, not boxes (ai-hands-on, 2026-08-22)

The section "Three forms, and each rots differently" above treats meta-curriculum, build-from-scratch and executable-curriculum as a partition — a curriculum picks one, and the choice predicts how it decays. ai-hands-on does not fit, and the way it fails to fit is the useful part: it is all three simultaneously, and the three are separable folder by folder.

Its notebooks run in order (executable). 3.Neural-Network(NN) and 4.Transformer make the reader build neurons, layers and attention by hand (from-scratch). The README carries 19 outbound GEN AI reference links, six books marked "not included" and a framework table (meta-curriculum). One repository, three forms, and no confusion about which is which.

The decay argument is what makes this worth a section rather than a footnote. If each form rots in its own way, a hybrid inherits every failure mode at once — link rot across 19 pointers, toolchain drift under the PyTorch notebooks, and API drift under a 5.RAG module that calls a live model endpoint. The maintenance bill is roughly additive, and it is invisible from the outside: a reader sees one repo and cannot tell it carries three decay clocks. 65 commits is the whole budget being spent against them.

So the taxonomy should be restated. The three forms are axes a curriculum scores on, and the right question for any source here is which axis dominates, not which box it belongs in. The pure cases (539kernel-book, nirdiamant-prompt-engineering) are the ends of the axes, not the normal case — and the corpus does not yet know which is more common, because until now the pages were written as though the choice were forced.

A second thing this source settles about vocabulary. ai-engineering-from-scratch covers a near-identical arc — maths through LLM engineering — in 503 lessons across 20 phases with 17 capstones. ai-hands-on covers it in about ten notebook folders. Both call themselves complete guides. Neither publishes completion data, and the corpus has no basis to rank them. What the pair does establish is that "complete" is doing no work in this genre: it states intent, not coverage, and survives a 50× difference in material without either author noticing a problem. That belongs beside "What the corpus cannot say" — it is not that we lack the evidence to compare, it is that the claim being made is not the kind that could be checked.

And the recurring gap shows up again, in its usual form. ai-hands-on addresses "a beginner or an engineer levelling up" and states no prerequisites, while running from derivatives and vectors to decoder-only attention through two intervening folders. The spoke has seen the opposite discipline — ../mathematics-wiki's grinberg-graph-theory, where a prerequisite list visibly decided what got cut. Unstated entry level is the more common choice in self-directed curricula, and it is worth recording as a property of the genre rather than a flaw of this repo.

One observation about defaults, filed under pedagogy. The 5.RAG module's default provider is reached through a link carrying campaign-tracking parameters that name the repository. Nothing states a commercial arrangement and none is claimed here. What is worth holding is the general point: learners run the default, so whatever a teaching repo defaults to is a pedagogical decision regardless of who benefits from it, and this corpus should read defaults as curriculum content.

The assessors answer, and they are arguing (2026-08-28)

The spoke's oldest structural hole was that every source in it was written by a candidate or for one. hiring-without-whiteboards closes the half of that hole it can: 762 companies, each a line saying how it interviews, published by lauren-tan under MIT with 51.8k stars behind it.

What the employers who bothered to say are doing instead. Counting the entry sentences: 453 mention a take-home exercise, 199 pairing, 48 a code review, 19 system design, 12 live coding. That is work-sample-assessment — assessment by resemblance to the job — and it is now the corpus's third assessment regime, beside the recall interview (interview-preparation) and the scored competition (capture-the-flag). Six in ten of these companies replace the interview problem with homework.

The title names the wrong variable and the README admits it. Whiteboards are "a metaphor, and a symbol for the kinds of CS trivia questions associated with bad interview practices"; the stated test is real-world problem yes, puzzle no, furniture irrelevant. Eight entries describe using a whiteboard inside a process the list endorses. Useful, because it puts the argument on the axis this spoke already runs on — what the exam is about, not how it is administered.

The finding that is actually new: the standard and the practice are in the same repository, and they disagree. RECOMMENDATIONS.md asks companies to pay the candidate, time-box under four hours, and offer a shorter alternative to anyone who cannot spend an evening. Of the 762 listed companies, 27 mention paying and 4 mention a time box. Nothing filters entries against the recommendation, and nothing claims to. The regime's own advocates state the cost, name the mitigation, and then admit companies without it — which is the same shape as cirr-standards: a well-drawn instrument with almost nothing measured through it. Two sectors, two ways of not counting.

Where it fails as evidence. Entries are self-reported, unaudited, opt-in, and corrected only when a candidate opens a "Duds" pull request. And that channel has closed: commits ran 589 in 2017 and 465 in 2019, then 64 in 2024, 8 in 2025, 35 so far in 2026. Two thirds of the file predates 2021, so an entry is a claim about how a company interviewed some years ago, kept alive by a repository that is still pushed and barely edited. It is T3, and the tier is doing real work here rather than marking a genre.

What it therefore does not do: it does not tell us whether work-sample hiring predicts performance, and it does not tell us whether the practice is spreading. So the standing thesis — syllabus follows assessment — gains a rival regime but no evidence that the rival is winning, and the 350k-star curricula still train for the exam this list rejects.

One pattern held, from the other direction. lauren-tan is the third publisher here with nothing to sell (after mito753 and mohammed-q-hussain), against four with a commercial layer. The corpus's read on the free-repository economy survives: when there is no product behind the artifact, what replaces it is advocacy, and advocacy has its own way of not checking itself.

The institution's own offer, and the fourth shaping force (2026-09-09)

harvard-free-courses-catalog is the first source here that is neither a curriculum nor a study of one. It is a supplier's current catalogue: 129 free courses, 128 distinct, every one online, read 2026-09-09.

It breaks the shaping story again, in a new direction. The founding claim was that an assessment regime shapes a curriculum; developer-roadmap replaced the exam with a job title. Harvard's mix is shaped by neither. Counted by the catalogue's own labels: Humanities 32, Data Science 17, Science 15, Health & Medicine 12, Computer Science 12, Social Sciences 11, Business 9, Art & Design 8, Programming 5, Education & Teaching 4, Mathematics 3. No exam and no hiring market produces three mathematics courses against thirty-two in the humanities. A faculty's shape is the fourth organizing force, and it is the only one on this page that predates the internet.

It also shows what an institution can build that an individual cannot. Ten free listings carry the CS50 brand — the intro course forked into Python, Web, SQL, AI, Cybersecurity, Scratch, game development, for Business and for Lawyers. Every curriculum in this corpus is one author's scope decision; this is an intro course extended by audience, which takes staff rather than an opinion. mooc is added as the sixth form to hold that difference, and it is defined by institution and credential rather than by pedagogy, unlike the other five.

The free/paid line sits after the learning, not around it. "Audit for Free / Add a Verified Certificate for $219" (checked on four courses; one is $299). Every self-published curriculum here is free with a book or a course sold beside it; the MOOC sells the evidence you did it and gives the teaching away. Same word, opposite structure.

And the outcome hole gets sharper rather than smaller. mooc-pivot is this spoke's only outcome evidence, built from every HarvardX and MITx course to May 2018 — 12.67M registrations, second-year retention 38% falling to 7%, completion flat across six years of effort. Eight years later the same institution publishes 129 courses on the same platform with no enrolment, completion or retention figure anywhere in the catalogue. The standing caveat here has been that a star count measures intent rather than learning. This is worse than a star count: the operator has the numbers, published them once for researchers, and shows a prospective learner none of them. Growth edge 2 does not move.

Open questions

Growth edges

Ranked; each names the kind of source that would close it (see ../QUALITY.md → Growth edges).

  1. First T1, and the spoke has no tiered floor at all CLOSED 2026-08-06, and its institutional-curriculum successor CLOSED 2026-08-08 by cs2023 (ACM/IEEE-CS/AAAI, T1). Floor now T1 2 / T3 6. Its prescriptive-half successor CLOSED 2026-08-10 (research pass): the report extracts fine as fourteen section PDFs, and cs2023 now holds the CS Core / KA Core / Non-core scheme, the full 17-area hour table, and the four per-topic skill levels. 270 CS Core hours is the binding total. What replaces it: nobody here has checked whether the guidelines are followed. CLOSED 2026-09-03 (research pass) by cs2023-accreditation (T1), and the answer is that the question has no subject. ABET's 2024 computer-science criteria are, in Oudshoorn's words, "a subset of those identified in CS2013" — a subset of the guidelines CS2023 replaced. Artificial intelligence (12 CS-Core hours), graphics (4), HCI (8) and software engineering (6) are in CS2023's core and absent from the accreditation criteria entirely. So nobody is assessed on conformance to the CS Core and nobody publishes it (corrected 2026-10-05: one department has; see 1a), and a department can hold current ABET accreditation for a computer science degree while teaching no AI. The lag is procedural and therefore permanent: a criteria change needs two commission votes and a comment period, at least two years, so every guideline revision meets an accreditor still enforcing the previous one. 1a. What the 40 credit hours actually contain. CLOSED 2026-10-05 (research pass) by curriculum-alignment-uaeu (T1, arXiv 2606.19469): the ABET-accredited BSc in Computer Science at united-arab-emirates-university, 41 courses and 728 syllabus and outcome items traced against CS2013 and CS2023 by two raters (κ 0.64 / 0.69). It covers 48.4% of CS2023's knowledge units and 59.4% of its recommended hours, and 69% of covered units rest on one course. Parallel and distributed computing is 0 of 5 units, and the paper marks it unmet against ABET too. Depth reached falls from 95% of delivered units under CS2013 to 76% under CS2023: a program last accredited in 2023 teaches to the guideline before the current one. That is the lag cs2023-accreditation described, now measured in one department. The authors measure their own program, and the figures are the lenient end of a range about 7 points wide. Original text: New 2026-09-03, successor to the above. Both documents are now held (cs2023, abet-computing-criteria) along with the gap between them, and all three describe requirements. Not one source in this spoke describes what a single named department teaches. — needs: one accredited program's published degree map or syllabus set, traced against either list. T1/T2, and it is the cheapest edge on this page — hundreds of departments publish theirs.
  2. Any outcome data — the corpus holds popularity and nothing else. — needs: a study of self-directed learning completion or hiring outcomes, T1/T2. Sharpened 2026-08-18 by cirr-standards, and the news is about the supply of evidence rather than about outcomes. The one body that audits outcomes to a published methodology in this domain lists three member schools, shows "TBD" for every headline metric, and has not issued a cycle since its 2023–2024 reports. So the absence this edge names is not the corpus failing to look. Two populations, two results: a university consortium instrumented itself and published a collapse (mooc-pivot); the bootcamp sector wrote a better instrument and stopped feeding it. — needs now: a mandatory reporting regime's data (state regulator filings, gainful-employment-style disclosures) rather than another voluntary standard.
  3. What decides pedagogy, if not the exam? New 2026-08-05. developer-roadmap broke the founding correlation — curated pointers with no exam behind it — leaving two builders and three curators with nothing sorting them. — needs: a curriculum whose author states the choice and the reason for it, or a comparison across curricula that treats pedagogy as the variable. A third force named 2026-09-09, not the answer this edge asks for. harvard-free-courses-catalog is organised by a faculty's shape — 32 humanities courses against 3 in mathematics — which is neither an exam nor a job title, and it comes with staff enough to fork one intro course ten ways. That is a scope determinant again, though, so the edge's actual question stands: nobody here has stated why they teach the way they teach.
  4. The assessment side, from the assessors — every source here is written by candidates or for them; nothing represents the people setting the exam. — needs: T1/T2 on technical hiring practice or interview validity research. HALF CLOSED 2026-08-28 by hiring-without-whiteboards (T3, hub-routed from Telegram): 762 companies stating their own interview process, which is the practice half — take-home 453, pairing 199, code review 48 — and the first employer-authored source in the spoke. Not struck, because the edge asked for two things and this is neither at the tier asked for nor the validity half at all: self-reported, unaudited, opt-in, and thinning (8 commits in 2025). Successor, unmet: does work-sample assessment predict job performance? work-sample-assessment is now paged as a regime with no evidence under it, and the intuition behind it is strong enough to pass for a finding. — needs: selection-validity research from industrial/organizational psychology, or an employer's own hiring-outcome data, T1/T2. Note the shape this repeats: the spoke can now describe two assessment regimes and score neither, exactly as edge 2 leaves it able to describe how bootcamp outcomes should be counted and unable to state one.

Coverage edges (added 2026-08-08, at the curator's request for a wider backlog). These widen what the spoke covers instead of answering an open question above; one ordinary solid source closes any.

  1. Where learners actually are. freeCodeCamp, Coursera and The Odin Project deliver most of this material and none has a page, so the spoke holds curricula (developer-roadmap, cs2023) with no vehicle underneath. — needs: platform documentation plus enrolment or completion figures.
  2. Bootcamps, and the one place outcomes get published. Growth edge 2 wants outcome data; the CIRR reporting standard is where it exists, contested as it is. — needs: the CIRR standard plus one audited outcomes report. HALF CLOSED 2026-08-18 (research pass): cirr-standards is in, T1, read from the ten-page PDF rather than the summary page — day-3 auditor submission of the intent survey, ten documented outcome buckets, BLS SOC codes for "in-field", base-salary-only medians, an AICPA peer-reviewed auditor reconciling against the day-3 roster. Not struck, per the half-close rule now in ../QUALITY.md: no member report was opened, so the corpus holds the methodology and still no audited number. Successor, unmet: one member school's audited CIRR report, read at the numbers — placement by bucket, the share of graduates reporting salary, the non-reporting rate. Three schools publish one; without it this spoke can describe how bootcamp outcomes should be counted and cannot state a single counted outcome.
  3. Accreditation. CLOSED 2026-08-11. abet-computing-criteria (T1, the 2026–2027 Computing Accreditation Commission criteria) is the binding document: 30 semester credit hours of computing generally, 40 SCH of computer science and 15 of mathematics/statistics for a CS program, six mandatory student outcomes including ethics, teamwork and communication. Successor, unmet: the two prescriptive documents do not share units. CS2023 counts topic hours (270 in the CS Core); ABET counts semester credit hours (40). Nothing in either maps onto the other, so the corpus cannot say whether an ABET-accredited program covers the CS Core. — needs: a crosswalk between the two, or a department's degree map traced against both. HALF CLOSED 2026-10-05 by curriculum-alignment-uaeu: one department's courses are now traced against both documents. But the paper reports coverage over all 161 CS2023 units, not the CS Core alone, and counts recommended hours, never ABET's semester credit hours. So one program sits on both lists and the unit question is still open. 7a. The crosswalk itself. New 2026-10-05, residue of 7. — needs: a published mapping from CS2023's CS Core hours to ABET's 40 semester credit hours, or a department's CS Core coverage stated on its own. T1/T2.
  4. Findings, not just venues. iticse is a conference page with no research under it — pair programming, worked examples, retrieval practice all have decades of results. — needs: a SIGCSE or ICER study, T1.

Cross-spoke adjacency

Index — Engineering Education Wiki

Catalog of every page, grouped by schema.org @type. Spine: synthesis (thesis), log.md (history), this file (catalog). Spun out of the hub _inbox on 2026-08-04 (3 sources: a CS-fundamentals study plan, an AI/ML build-it curriculum, a system-design resource collection). Star counts and "N learners" are marketing numbers on a dated snapshot — record the date, and never read popularity as evidence that the curriculum works.

DefinedTerm (concepts)

Collection (sources)

Person

Organization

SoftwareSourceCode (sources)

Book (sources)

Article (sources)

ScholarlyArticle (sources)

Report (sources)

Event

Synthesis

Notes

Related pages

Links to

Linked from