Log — Engineering Education Wiki

Append-only history. Entries start with ## [YYYY-MM-DD] <op> | <title>.

[2026-08-04] split | engineering-education-wiki created from _inbox (3 sources)

Spun out when awesome-system-design-resources completed a cluster two learning-roadmaps records had been holding since 2026-06-15 (coding-interview-university, ai-engineering-from-scratch). The coding-interview-university park record had stated the case plainly: no spoke owns CS fundamentals, data structures and algorithms, or CS education, and it had been declined by ../bit-manipulation-wiki (owns bitwise tricks, one sub-section), ../optimization-algorithms-wiki (metaheuristics) and ../research-wiki (knowledge management and formal methods, not study curricula).

The domain is not "lists of links." It is how engineers learn and are assessed, which the three founding sources happen to document from the inside.

[2026-08-05] ingest | Prompt Engineering (NirDiamant)

Routed from the hub (runner-up: ../agentic-tooling-wiki). nirdiamant-prompt-engineering — 22 runnable Jupyter notebooks, one per prompting technique, 7,752★ / 1,006 forks, custom non-commercial licence, T3. The spoke's first source that is not one of the founding three.

Three things it changed. It is a third pedagogical form — neither curated pointers nor build-from-scratch, but a lesson that is itself executable — so executable-curriculum was created alongside meta-curriculum, with the decay contrast (link rot vs dependency drift) as its point. It is a fourth data point on the founding exam/pedagogy correlation that lands on the predicted side but proves little, since a technique domain has no exam by construction — recorded in synthesis as weak support rather than confirmation. And it was the first hard boundary case: prompt engineering is owned by no spoke, so the "route the claim to the claim's owner" rule had no target and the call was made on shape — course-shaped here, reference-shaped stays parked.

Verified against the GitHub API rather than the README: 22 notebooks confirmed, but they sit in one flat directory with no numbering, so the seven-tier sequence exists only in the README prose.

Entities: created nir-diamant (Person; entity-index returned no match ≥ threshold). Sibling repositories and the newsletter are recorded as prose, not noded — evidence is a README mention only. Budget: 1 of 8.

[2026-08-05] ingest | Top 50 LLM Interview Questions (Hoang)

Routed from the hub as a Google Drive PDF over Telegram. llm-interview-questions-hoang — Hao Hoang's self-published 50-question LLM interview-prep doc (May 2025), T3. The PDF is held in raw/llm-interview-questions-hoang.pdf (a Drive link is not durable provenance); origin URL recorded on the page. Created hao-hoang (Person).

First non-repository source and first Article in the spoke, and its first on the AI/LLM- engineering interview (the founding three cover CS fundamentals + system design). It files here on the spoke's own boundary rule: every question is about another spoke's subject (sampling/attention → llm-inference, LoRA/QLoRA/PEFT → machine-learning, GPT-4/Gemini/foundation models → llm-providers, RAG/ CoT → research), but it makes no claim about any of them — it is an assessment artifact, so the form routes here and the substance cross-links.

Synthesis + interview-preparation updated: it is the founding thesis's purest instance — not a curriculum shaped by an exam but the exam itself, printed, with no path around it. Removes the inference step the thesis previously needed. Also characterises what the LLM interview rewards today: recall of mechanism (the flashcard, opposite pole to ai-engineering-from-scratch's build-it). Does not touch the corpus's outcome-data hole.

Entity note: entity-index returned a 0.80 fuzzy MATCH to chao-huang (../research-wiki). Not the same person — Hao Hoang ≠ Chao Huang, a name-similarity false positive. Created a separate hao-hoang node and recorded the distinction on both the node and the index line, per the "read before merging" invariant.

Caveats on the page: self-published/promotional (LinkedIn lead-gen), introductory depth with some invented examples and dating content (GPT-4 vs GPT-3), cites nothing.

[2026-08-05] ingest | developer-roadmap (roadmap.sh)

Routed from the hub (Telegram, via an opensourceprojects.dev write-up; ingested against the repository itself rather than the write-up). developer-roadmap — kamranahmedse, 363.7k★/44.7k forks, T3 (maintainer's own repo, commercial layer beside it). Paged kamran-ahmed.

Note for the record: this is one of the three artifacts the hub named when it spun this spoke out, and it had never been ingested. The spoke built its founding thesis without its own most-cited example.

Gap-relevance is unusually high — it is the exact falsifier the founding open question named, a curated-pointer curriculum with no assessment behind it. It arrived and the exam↔pedagogy correlation broke. The surviving claim is narrower and better: an exam shapes a curriculum's scope (coding-interview-university dropping frontend and SQL), not its pedagogy. And developer-roadmap introduces a third organizing force the first four could not show — job title, one roadmap per role a company hires for, which is neither a syllabus nor a body of knowledge.

Open question 1 closed in the negative; a successor opened ("what decides pedagogy, if not the exam?"). The T1 floor is unmoved: this is the fifth T3 and the fifth source measuring intent to learn rather than learning. Third self-publisher-with-a-commercial-layer in a row, which is now a property of the corpus rather than a coincidence and is recorded on kamran-ahmed.

[2026-08-06] ingest | The Self-Directed Disposition: What Computing Students Say (research pass, T1 — the spoke's first)

Sabin, Kumar, MacKellar, McCauley, VanDeGrift & Matsumoto, ITiCSE 2025 (ACM, doi:10.1145/3724363.3729114). Hunted via the hub Research Pass against ## Most wanted #4 — "first T1 (auto-edge: zero-T1 spoke)... needs peer-reviewed CS-education research." This is it. ACM DL hard-blocked WebFetch (403); read via the firecrawl fallback plus the NSF public-access PDF.

Created self-directed-disposition-study (ScholarlyArticle, url-only, T1), mihaela-sabin (Person, lead author) and iticse (Event/BusinessEvent, the venue). Entity budget 2 of 8; five co-authors kept as plain mentions on the source page.

Why it matters beyond the tier. It is the first source here not published by someone with a product adjacent — the pattern kamran-ahmed made visible and that edge #1 had named as the reason the floor was not drifting up on its own. And it examines the disposition every artifact in this corpus assumes and none of them studies.

Two findings folded into a new synthesis thread. Students overwhelmingly equate self-direction with finding external material (148 of 380 coded behaviours, roughly 4× the next category, and more than planning + self-assessment + review combined) — which is precisely the work the meta-curriculum form does for them, opening a new open question the corpus cannot answer either way. And the largest self-reported obstacle to being self-directed is assignment structure (12 of 27), which puts a mechanism under a suspicion the spoke had no way to voice: a curriculum here is an assignment structure with the grade removed, and the most structured form (executable-curriculum) is the one the evidence would predict least self-direction from.

Limitations recorded rather than smoothed: qualitative self-report; responses not disaggregated by course, exercise type or any student characteristic (the authors' own stated transferability limit); no record of the assignment each reflection targeted; US-centric; four of six authors taught 12 of the 18 sections sampled. It measures what students believe self-direction is — the stated research question — not whether those behaviours produce learning.

Growth edge #1 retired as CLOSED; replaced, ranked lower, by the institutional-curriculum half it also asked for (ACM/IEEE Computing Curricula, a university syllabus), which is still absent.

[2026-08-08] ingest | CS2023 — the ACM/IEEE-CS/AAAI curriculum guidelines

T1, via the research pass (hub ## Most wanted edge #5, the institutional-curriculum successor to this spoke's first-T1 edge). cs2023 new.

The corpus's first designed curriculum. Everything else here accreted: a repository, a star count, a maintainer's taste. CS2023 is 17 knowledge areas argued to by a joint task force and endorsed by ACM (18 Jan 2024), IEEE-CS (22 Jan 2024) and AAAI (22 Feb 2024), revising CS2013 — Computational Science dropped, AI/security/PDC/ethics expanded, mathematics enlarged.

What it buys the wiki is a control. Reading CS2023's areas against developer-roadmap and coding-interview-university makes the repositories' popularity selection measurable rather than asserted: Society, Ethics and the Profession and Human-Computer Interaction are first-class areas in the guidelines and near-absent in the repositories, which sort toward interview-preparation. Written into synthesis as its own section.

It does not touch the spoke's real complaint. CS2023 reports no learning outcomes either, so the corpus goes from four sources with no outcome data to five.

Limitation recorded on the page, not hidden: the 400-page final report returned compressed binary through every fetch route, so the CS Core / KA Core tiering and the hour counts are not held — the part of a curriculum that actually binds a department. Successor edge written for exactly that.

Floor: T1 1 → 2. No entity discovery — ACM, IEEE-CS and AAAI are bodies the source names but the spoke has no organization nodes and one ingest is not the moment to open three.

[2026-08-09] ingest | Two stale hub parks routed here — a reference and a clone

Both sources were parked in the hub _inbox before this spoke existed (2026-07-26 and 2026-06-28; spun out 2026-08-04) and nobody went back for them. Found by a scan of the holding pen at the curator's request. Both T3.

prompt-engineering-guide — DAIR.AI, 77.4k★/8.5k forks, MIT, 1,589 commits, 13 languages, "3 million learners" as of January 2024. The registry note for this spoke had already named this park as its live boundary case, and the boundary rule decides it: a curriculum teaching a technique domain files here. New entity dair-ai.

adrianhajdin-uber-clone — JavaScript Mastery, 1.9k★/692 forks, no license stated. A ride-hailing app rebuilt across 15+ features with five vendor services. New entity javascript-mastery, new concept clone-project.

What changed in the thesis. Two things, and the first is the bigger one.

1. A controlled comparison. The prompt guide and nirdiamant-prompt-engineering share a subject, an absent exam and a genre, and made opposite pedagogical choices. Five sources of consistent-but-uncontrolled evidence now have one pair that holds the variables still, and it settles the negative half of the standing claim: the assessment regime does not pick the pedagogy. What does is still open.
2. "Build it" split in two. ai-engineering-from-scratch builds from scratch; the clone builds from services and implements nothing. The spoke's curation-versus-construction vocabulary could not separate them, so clone-project now names the second and the rot ordering in synthesis has been extended — a clone decays faster than anything else here, on five vendors' release cycles.

Recorded weaknesses. No license on the clone repo, which unsettles its own portfolio premise. Its 1.9k stars count the answer key rather than the YouTube course that does the teaching. The guide's "3 million learners" is a marketing line unchanged since January 2024, and its per-model pages (GPT-4, Phi-2, Mixtral, Flan, OLMo) are a 2023–24 shelf. Still no outcome data anywhere in the corpus.

[2026-08-10] ingest | Kernel-Exploit-Dojo — the exam with no curriculum in between

Routed from the hub (Telegram). github.com/mito753/Kernel-Exploit-Dojo — 403★/62 forks, 696 commits, 100+ Linux kernel exploitation CTF challenges from 2020–2026, exploit code in C, no license stated. T3: a maintainer-owned repository describing its own material, no external review.

New pages: kernel-exploit-dojo (source), problem-archive (form), capture-the-flag (assessment regime), mito753 (thin Person).

Why it earned three concept pages. The spoke's claim is that where an assessment regime exists it shapes scope. Every earlier case went through an author who wrote a plan. This one doesn't: the scope is the exam — past competition problems kept with the organizers' files, a working exploit and a writeup. The claim survives its most extreme test, and what it produces there is a fifth curriculum form, neither authored nor a set of pointers.

capture-the-flag is the corpus's second assessment regime, and the contrast is what makes it worth a page: public scoring, per-problem results, and problems used once then released. Those properties are why a problem-archive can exist in security and cannot exist around the hiring interview — the form is rare here because the dominant exam keeps its problems, not because nobody wants one.

The find. Per-challenge difficulty grades and solve counts — the first quantitative signal in this corpus that is not a star count. It describes the item (how many teams solved it under time pressure), not the learner, and it is inherited from the competitions rather than measured here. The outcome hole is still open; this moves the corpus off zero on the axis beside it.

Recorded weaknesses. No license, on a repository that redistributes other people's challenge material. Solve counts unsourced within the repo. Coverage follows what got competed on — the popularity selection cs2023 exposed, arriving by a different door. And a new decay mode: not link rot but relevance, as drilled bug classes get patched out of the kernel.

Scope note. Catalogued as a teaching artifact — structure, pedagogy, evidence. The repository's own disclaimer (CTF and local lab only) is recorded on the page; no technique detail or operational guidance is held here.

[2026-08-10] ingest | CS2023 — the tiering and the hours (research pass)

Arrived through the hub Research Pass, against ## Most wanted #5 — the curriculum's prescriptive half. Edge closed.

cs2023 refreshed in place. The page had recorded the CS Core / KA Core tiering and the per-area hour counts as a live gap on the finding that the 400-page report "could not be extracted." That was wrong about the file: the report is published as fourteen section PDFs, and §3.1 Body of Knowledge and §1.3 Introduction to the Knowledge Model both extract cleanly with pypdf. Same failure mode the hub already has a standing rule about — a fetch failure on a PDF says nothing about the document.

Now held: the three-tier scheme in the report's own words (CS Core binding on every curriculum, KA Core recommended depth, Non-core the renamed electives), the full 17-area table of knowledge units and hours, and the four skill levels attached per topic.

270 is the number that binds — total CS Core hours, roughly a year of contact time, and the whole of what CS2023 actually requires. The report refuses to total the KA Core column because those hours are not additive.

It also moves an existing claim. This page argued the repository corpus under-weights SEP and HCI. True, but the hours show the larger omission is mathematics — MSF at 55 CS Core hours, the biggest single requirement in the report, and the one developer-roadmap and coding-interview-university route around most completely.

2026-08-11 — ingest: ABET computing criteria (quality cycle)

abet-computing-criteria (T1, ABET Computing Accreditation Commission, 2026–2027 criteria). Coverage edge 7 closed. The spoke's second T1 and its first binding document — cs2023 is advice three societies endorsed, this is the rule a US department loses its accreditation for breaking.

Numbers held: 30 semester credit hours of computing under the general criteria; for a computer science program, 40 SCH of CS and 15 SCH of mathematics and statistics with rigor "at least equivalent to introductory calculus"; five general student outcomes plus a sixth for CS; PhD-holding full-time faculty. The 2026–2027 text went through a 180-day public review opened 24 October 2025, comments to 15 June 2026.

The comparison that does not work, recorded before someone tries it. CS2023 counts topic hours (270 in the CS Core); ABET counts semester credit hours (40 in CS). The units do not convert and neither document maps onto the other. Written into the page, the synthesis and the successor edge, so a later cycle does not manufacture a ratio out of them.

What it does for the thesis. The standing claim is that an assessment regime shapes scope. Every earlier case was the hiring interview acting on an author's judgment. ABET acts on the program, with a sanction, and produces exactly the scope the repositories drop: ethics and professional responsibility, teamwork, communication — mandatory attainments, absent from developer-roadmap, coding-interview-university and awesome-system-design-resources.

What it does not do. Edge 1's successor asked whether any department conforms to the CS Core. ABET publishes criteria and an accredited-program list, not per-program findings, and its criteria nowhere require CS2023 coverage — so that question stays open and is now split in two on the page. No outcome data here either.

[2026-08-15] ingest | The MOOC Pivot — the spoke's first outcome data

Quality cycle. This spoke scored depth 1/7 with zero T2 sources on 2026-08-13 (T1 3 / T2 0 / T3 9), and its synthesis.md has carried the same central hole since founding: "Does anyone finish? No completion or outcome data in any founding source." Every page here measured reach — stars, forks — and none measured learning.

Closed it from the side with mooc-pivot (Reich & Ruipérez-Valiente, Science 363(6423), 2019, T1). WebFetch got HTTP 405 on the MIT DSpace post-print; downloaded and extracted locally with pypdf per the standing rule, and the body text gave the figures.

Why it counts. 565 course iterations, 261 courses, 12.67M registrations from 5.63M learners — every MITx and HarvardX MOOC to May 2018. Second-year retention fell 38% → 7% across successive cohorts; only 12% of the peak 1.1M cohort took another course the next year; completion has not improved in six years even among learners who paid, and even among those who told a survey they intended to finish.

Why it does not count as much as it looks. A MOOC is not a meta-curriculum. A repository has no enrolment, no cohort and no completion event, so nothing transfers by arithmetic, and the page says plainly that quoting "7% retention" about a GitHub curriculum is a misuse. What it does is shift the burden: the nearest, better-instrumented population shows collapsing follow-through, so the optimistic reading of a 352k-star repo is now the claim needing evidence.

It also lands on "Who are these for?" — the paper splits registrants by HDI quartile and finds growth concentrated among the already-educated in affluent countries, MOOCs being "a complementary asset for learners within existing systems." That makes "enormous population of self-taught entrants" the weaker of the two readings the spoke had been holding as equals.

Provenance limit recorded on the page: the per-year completion percentages are in Figure 3, an image local extraction cannot read. The flat-over-six-years finding is quoted from the body; the numbers themselves are not reproduced, because this wiki has not seen them.

Entity: justin-reich (Person) — the spoke's first author who measured outcomes rather than shipping a teaching artifact. Ruipérez-Valiente and MIT left as prose (one mention each). Tier floor moves T1 3 → 4.

[2026-08-17] ingest | A Journey in Creating an Operating System Kernel — The 539kernel Book

Arrived via Telegram (Facebook fbclid stripped from the URL before recording), routed by the hub. Downloaded and extracted locally with pypdf — 215 pages, seven chapters, LaTeX, November 2022, free.

539kernel-book, T3 — consistent with every self-published curriculum here: one author, no publisher, no peer review, no outcome data. Two things separate it from the T3 shelf even so. It was read before release by an academic (Dr. Hussain Almohri), the only pre-publication review recorded anywhere in this spoke; and its author has nothing to sell, which takes the publisher-with-a-commercial-layer pattern from 3/3 to 3/5 (mito753 was the first break).

Two new concept-adjacent pages. build-from-scratch finally holds the half of the corpus's own from-scratch-vs-from-services split that had been left as a property of one outlier (ai-engineering-from-scratch); the second source lets it be defined against its neighbours, and the sharp new distinction is cumulative vs modular — chapter five will not run without chapter four, which is why the repository ships evolution_by_versions/ so a reader can rejoin at a chapter boundary. A failure mode turned into infrastructure, and no modular curriculum needs one. mohammed-q-hussain is the author node.

The finding is the ending. The spoke's founding claim is that what gets taught is downstream of what gets tested, and every source has fitted it, including the ones that stretched it. This one does not fit at all: chapter 7 walks through five research papers (Endokernel/in-process isolation, nested kernel, multikernel, THINK dynamic reconfiguration, unikernels/HermiTux), names the two exits traditionalist and kernelist, and tells the reader to go implement what the literature has not adopted yet. The terminal state is production of new work, not assessment — a third shaping force beside the exam and the job title, and the first that points away from employment. Recorded in synthesis.md with both available readings, since one source cannot separate "systems programming has no hiring funnel" from "an author with nothing to sell has no job to promise."

Also logged, on the page and in the hub: no spoke in this hub owns operating systems. The book files here as a teaching artifact per the boundary rule; its subject has no home. Two more systems-programming sources make that a spin-out question rather than a filing habit.

[2026-08-18] ingest | CIRR standards — via research pass

Quality-cycle research pass, hunting coverage edge 6 (the CIRR standard plus one audited outcomes report). https://www.cirr.org/our-standards → the ten-page PDF Standards and Implementation Governing Outcomes Reports for 2023-2024 Graduates (rev. 2025-01-06). T1, freshness: stable.

New pages: cirr-standards (source), cirr (Organization). Updated: cs2023 (added the units mismatch against abet-computing-criteria — which also cleared this spoke's one orphan), synthesis (edges 2 and 6), index.

The fetch mattered. WebFetch on the standards web page returned six headings and an explicit "what's missing" list — no denominator rules, no outcome definitions, no audit procedure. All of it is in the linked PDF, which downloaded fine and extracted cleanly with pypdf. Reading the page would have produced a page saying the standard is vague. The standard is the opposite of vague.

What it holds. Intent surveyed by the first day of class and sent to an approved auditor by day 3, with no retroactive administration and no change of intent permitted afterwards — the denominator is fixed and third-party-held before any outcome exists. Ten outcome buckets with documentary requirements (offer or start date, paid, accepted; "in-field" needs a skills statement or a BLS SOC job title; full-time means permanent/at-will/6+ months and 30+ hours; contract-to-hire counts as 1B until a permanent offer lands; "could not contact" needs four attempts across two channels). Salary is base only, the advertised figure must be the CIRR median, and it must carry a clear disclosure of what share of graduates reported salary at all. Audit by an AICPA peer-reviewed firm, board-approved, reconciling the published report against the day-3 roster, filed 1–31 December.

Half close, not a close, and the reason is the finding. cirr.org/schooldata lists three schools (Code Platoon, Codesmith, Hacktiv8), shows "TBD" for every headline metric, and its newest reports cover 2023–2024 — consistent with the governing document itself still being the 2023–2024 version. So the spoke now holds a rigorous method for counting bootcamp outcomes and not one counted outcome. Edge 6 keeps its successor (open one member report at the numbers); edge 2 is re-specified to ask for a mandatory regime's data, since the voluntary one did not survive its own market.

Recorded against this spoke's low tier floor (T1 4 / T2 0 / T3 10 before this ingest): the floor is not a grading habit. Its T3s are curriculum repositories tiered by what they rest on, each with a written justification, and the domain has no independent assessors — which is the same absence cirr-standards just documented from the inside.

[2026-08-21] lint | Tier re-check: the grade stopped carrying the spoke's finding

Acting on the ## Outstanding item raised 2026-08-21 ("Does the tier grade a source, or carry a spoke's finding?"), decided by the curator's delegation the same day.

The problem. The spoke's floor was 0.33, second-weakest in the corpus, on ten T3 curriculum pages. Reading the justifications, the T3s were doing two jobs: grading what each page rests on, and recording the wiki's own finding that these curricula claim results nobody has measured. Only the first belongs in a number that ranks 36 spokes.

The decision: split the jobs, per page, no sweep. Tier by the claim, not the URL (../QUALITY.md, adopted 2026-08-20) already contains both halves — artifact facts at T1, "an effectiveness claim for a curriculum or method" at T3 — so the answer was to apply its own test to each page rather than move eight at once. The result is a split, not the projected sweep.

Five moved to T1 — the wiki reads them for inventory, sequence and scope, which is the artifact describing itself: 539kernel-book (215 pages, seven chapters, prerequisites, version-per-chapter, read from the PDF), ai-engineering-from-scratch (503 lessons / 20 phases / 17 capstones + the stated build-it → use-it → ship-it rule), nirdiamant-prompt-engineering (22 notebooks, and the finding that the seven-tier sequence exists only in the README), kernel-exploit-dojo (per-challenge template, technique and difficulty taxonomies; the exploits and writeups are in the repo, so it is not a link catalogue), adrianhajdin-uber-clone (feature list, five-vendor stack, absent licence — the integration-not-implementation reading comes off the dependency list).

Five held at T3, each with the reason written on the page: awesome-system-design-resources and developer-roadmap (curated-pointer catalogues, which the rule carves out explicitly; the first also declares a commercial funnel), coding-interview-university (same carve-out — and the flagged boundary case, since what it excludes is checkable artifact fact), llm-interview-questions-hoang (its load-bearing claim is that it represents the exam, which is the author's and unverified), prompt-engineering-guide (already claim-keyed, and correctly: it summarizes other people's papers and self-reports its reach).

Floor 0.33 → 0.67, not 0.87. Flipping coding-interview-university alone would make it 0.73.

Where the finding went. meta-curriculum now carries it in prose, with the note that the argument did not change: a curriculum whose contents are first-party fact is still a curriculum nobody has shown teaches anyone anything, and the missing outcome data is still this spoke's largest hole.

[2026-08-21] lint | The boundary case resolved: coding-interview-university → T1

Follow-up to the tier re-check earlier today. The curator took all five amendments proposed for ../QUALITY.md, one of which re-keyed the catalogue clause to the claim rather than the artifact: a curated list is primary for the inclusion and exclusion decisions it states, and never for the material it links to.

That resolves the page this spoke flagged rather than decided. coding-interview-university states its own exclusions — frontend, full stack, SQL — and that stated omission is the reason the source is in the wiki at all. T3 → T1. Spoke floor 0.67 → 0.73.

The rest of the pass stands. Four pages remain at T3: awesome-system-design-resources and developer-roadmap curate without publishing criteria, llm-interview-questions-hoang rests on its own unverified claim to represent the exam, and prompt-engineering-guide summarizes other people's papers while self-reporting its reach.

[2026-08-22] ingest | ai-hands-on — AI Engineering: Hands-on (hub-routed, Telegram, URL-only, T1)

ramakm/ai-hands-on, MIT, 1.4k★ / 291 forks / 65 commits. Notebook curriculum for AI engineering: 1.Math → 2.PyTorch → 3.Neural-Network(NN) → 4.Transformer, with 5.RAG and 6.OCR run separately and a Basic ML Model Implementation folder beside the deep-learning line. Created ai-hands-on (T1, README read in full) and the Person node ramakm.

Not a duplicate of ai-engineering-from-scratch — different author (rohitg00), different scale (503 lessons / 20 phases / ~33k★ against ~10 folders / 1.4k★). Checked before creating.

The finding: the three curriculum forms are axes, not boxes. The spoke's taxonomy treated meta-curriculum, build-from-scratch and executable-curriculum as a partition. This repo is all three at once and cleanly separable by folder — notebooks that run, two folders that make the reader build neurons and attention by hand, and a README carrying 19 outbound reference links, 6 books marked "not included" and a framework table. Restated the taxonomy on executable-curriculum and build-from-scratch and in synthesis: ask which axis dominates, not which box applies.

Why that matters rather than being a filing detail. executable-curriculum already argued each form decays differently. A hybrid therefore carries all three decay clocks at once — link rot across 19 pointers, toolchain drift under the notebooks, API drift under a 5.RAG module calling a live endpoint — an additive maintenance bill invisible from outside the repo, against 65 commits of budget.

A vocabulary point worth keeping. Two sources now cover a near-identical arc, one in 503 lessons and one in ten folders, and both call themselves complete guides. Neither publishes completion data. So "complete" states intent, not coverage in this genre, and survives a 50× difference in material unremarked. Filed beside "What the corpus cannot say" — the issue is not missing evidence but an uncheckable claim.

Recorded, not asserted. The 5.RAG default provider (Atlas Cloud, deepseek-ai/DeepSeek-V3-0324) is reached through a link carrying utm_source=github&utm_medium=link&utm_campaign=ai-hands-on. No commercial arrangement is stated in the repo and none is claimed here; what is observable is a campaign-tracked default in a teaching module. The general point kept: learners run defaults, so a teaching repo's defaults are curriculum content whoever benefits.

Also recorded: no prerequisites stated despite a derivatives-to-attention gradient; no exercises, assessment or capstone — a read-and-run curriculum where the notebooks are the whole contract. 1.Math is the boundary with ../mathematics-wiki: taught as a means to an engineering goal, so it stays here.

Entity note: "Atlas Cloud" left as prose, not a node — vendor mention, tangential per ../ENTITIES.md (entity-index returned a false 0.62 match to alibaba).

[2026-08-28] ingest | hiring-without-whiteboards — companies that don't do whiteboard interviews (hub-routed, Telegram, URL-only, T3)

poteto/hiring-without-whiteboards, MIT, 51.8k★ / 3,926 forks / 2,175 commits / 421 contributors / 1,322 merged PRs, created 2017-03-15, last push 2026-08-24. Read from the raw README (829 lines) plus RECOMMENDATIONS.md and the GitHub API. Ingested with claude-opus-5 at high effort.

Created hiring-without-whiteboards (T3), the concept page work-sample-assessment, and the Person node lauren-tan. Updated interview-preparation, synthesis and index.md.

Growth edge 4 is half closed. The edge asked for the assessment side from the assessors, and every source here was written by a candidate or for one. This is 762 companies each describing its own interview process — employer-authored, and arguing against the exam the spoke's curricula train for. Not struck: the edge wanted T1/T2 and hiring-validity research; this is T3 practice claims and no validity at all. Successor recorded: does work-sample assessment predict job performance?

The counts, from the entry sentences (762 company entries, 757 with a described process): take-home 453, pairing 199, code review 48, paid 27, system design 19, live coding 12, explicit time box 4. Six in ten replace the interview problem with homework.

The finding worth keeping: the standard and the practice sit in the same repository and disagree. RECOMMENDATIONS.md — thirteen lines, the only prescriptive text here — asks companies to pay the candidate, time-box under four hours, and offer a shorter alternative to anyone who cannot spend an evening. 27 entries mention paying, 4 mention a time box. Nothing filters an entry against the recommendation. That is the cirr-standards shape again: a decent instrument, almost nothing counted through it. Filed on work-sample-assessment because it decides who can afford to be assessed this way.

Second finding: the correction channel has closed. Commits by year — 589 (2017), 75, 465, 318, 233, 207, 181, 64 (2024), 8 (2025), 35 (2026 to date). The only mechanism for removing a stale entry is a candidate opening a "Duds" pull request, and that flow has dried up while the repository keeps being pushed. So two thirds of the file describes how companies interviewed before 2021, and an entry is weak evidence about the present. Recorded as the reason for T3 rather than as a genre judgement.

Third: the title names the wrong variable, and the README says so. Whiteboards are "a metaphor…a symbol for the kinds of CS trivia questions"; the test is real-world problem yes, brainteaser no. Eight entries use a whiteboard inside a process the list endorses. Useful for this spoke because it puts the argument on the axis already in play — what the exam is about, not how it is administered.

Publisher pattern, third break. lauren-tan has no course, newsletter or paid tier beside the repository, after mito753 and mohammed-q-hussain. Against four publishers with a commercial layer. What replaces the commercial motive here is advocacy — and advocacy does not audit itself either.

Entity note: one Person node created, well inside the budget. Cursor and the React compiler team are recorded as prose from the GitHub profile, not noded — employer mentions, tangential per ../ENTITIES.md. The blog bio still says Facebook; both self-reported, the disagreement is on the page and dated.

Boundary recorded in synthesis → Cross-spoke adjacency: the hub's parked awesome-job-boards was declined here because a job board is a listing venue. This routed because it is entirely about how candidates are assessed. If a tech-hiring spoke ever spins out, the assessment stays here, the market goes there.

[2026-09-03] ingest | CS2023 Curricula Guidelines and Computer Science Accreditation

Research pass, quality cycle. Growth edge 1's successor asked whether anybody checks that a department covers the CS Core. cs2023-accreditation (Oudshoorn, High Point University, ICACIT 2024, hosted on ACM's own csed.acm.org) answers it structurally: the body that could check does not ask for it.

WebFetch returned PDF binary; extracted locally with pypdf, six pages.

The load-bearing sentence: "The computer science topics identified in the 2024 accreditation criteria are a subset of those identified in CS2013." Not CS2023 — the guidelines CS2023 replaced. Named as present in CS2023's core and absent from the criteria: artificial intelligence (12 CS-Core hours), graphics and interactive techniques (4), human-computer interaction (8), software engineering (6).

Also recorded: the CS2013→CS2023 hour ledger (279.4 → 270, with Software Engineering losing 17.6 of 23.6 while Software Development Fundamentals holds 43 unchanged), the faculty-survey ratings that displaced calculus as the stated maths requirement (probability 98%, statistics 98%, linear algebra 89%), and the two-year minimum for any criteria change — two commission votes plus a comment period, which makes the lag permanent rather than accidental.

Edge 1 closed; successor 1a opened, and it is the cheapest edge on the page: three sources now describe requirements and none describes what a single named department teaches. Hundreds publish their degree maps. avoid-ai-writing run.

[2026-09-09] ingest | Harvard's free-course catalog (pll.harvard.edu/catalog/FREE)

Arrived via Telegram with the title "Free Courses | Harvard University"; routed here by the hub (runner-up: none — no other spoke owns curricula, and the catalogue's non-computing majority belongs to no spoke at all).

Read by counting, not by summarizing. The listing paginates, so the whole result set (13 pages) was fetched and the 129 course cards parsed for subject label, title, duration and modality. That is where every number on the page comes from; the catalogue itself displays only "129 results". Four individual course pages were then opened to resolve the Free * asterisk the listing never expands.

Source: harvard-free-courses-catalog, T1 — a first-party institutional catalogue, which is the best evidence of what Harvard offers and no evidence about teaching quality. freshness: volatile: course instances expire and prices move, so every figure is a 2026-09-09 reading.

What the count shows. 128 distinct courses, all online. 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. Computing is 34 (37 with maths), the largest bloc after humanities. 97 of 128 state a length: 1–16 weeks, median 8.

Two structures the repositories cannot produce. Ten free listings carry the CS50 brand — the intro course forked into Python, Web, SQL, AI, Cybersecurity, Scratch, 2D game development, for Business and for Lawyers, which is extension by audience and takes staff rather than an opinion. And eight Data Science: X courses are HarvardX's paid professional certificate served free in pieces.

The asterisk, resolved. Course pages say "Audit for Free — Add a Verified Certificate for $219" (CS50, Data Science: R Basics, Pyramids of Giza); Introduction to Data Science with Python is $299. Four of 128 checked, so the page calls $219 the common case rather than the rule. The teaching is free and the evidence you did it is not — the inverse of every self-published curriculum here, where the credential does not exist and the commercial layer sits beside the free thing.

New concept page: mooc, the corpus's sixth curriculum form and the first not published by an individual. The other five are defined by how they teach; this one by who stands behind it and what is sold, which is why it is the only form with a measurable completion event — and why mooc-pivot's numbers do not transfer to a GitHub repository by arithmetic.

Folded into synthesis as a new section. The catalogue is shaped by a faculty, not by an exam (interview-preparation) or a job title (developer-roadmap) — a fourth organizing force, noted on growth edge 3 without closing it, since it determines scope rather than pedagogy.

The sharp finding, recorded as a gap that grew. mooc-pivot exists because Harvard and MIT published their own outcome data: every course to May 2018, 12.67M registrations, second-year retention 38% → 7%, completion flat over six years. Eight years on, the same institution lists 129 courses on the same platform with no enrolment, completion or retention figure anywhere in the catalogue. Growth edge 2 (any outcome data) does not move; that page updated with the supply-side comparison.

Entities: harvard-university reused cross-wiki from ../mathematics-wiki. HarvardX and edX are named in prose and deferred — the catalogue evidences the platform only through outbound course links.

Unchecked, and said so on the page: certificate prices on the other 124 courses, the per-course difficulty grades (the facet exists; the cards were not parsed for it), and whether the 129 figure is stable week to week.

[2026-10-05] ingest | Measuring curriculum alignment: one accredited CS degree against CS2013 and CS2023 (via research pass)

curriculum-alignment-uaeu (T1, arXiv 2606.19469). Hunted for Most wanted #37 / edge 1a. UAEU's ABET-accredited BSc CS traced course by course against both guidelines: 48.4% of CS2023 units, 69% of covered units on one course, parallel computing 0/5, depth 95% → 76%. Edge 1a closed; coverage edge 7's successor half closed, residue 7a (the crosswalk). Entity: united-arab-emirates-university (same-as verified: Wikipedia, Wikidata Q1544739, uaeu.ac.ae). Retraction in place: cs2023-accreditation and the synthesis said nobody publishes conformance to the CS Core. This paper does. Struck and corrected with the date.

[2026-10-05] quality | quality cycle

Scorecard in the hub quality-log.md.