Log — Machine Learning Wiki

Append-only history. Entries start with ## [YYYY-MM-DD] <op> | <title> (split · ingest · query · lint).

[2026-07-30] split | machine-learning-wiki created from the hub _inbox (5 sources)

Created on the human's instruction ("cluster them") after the router, parking ml-system-design-case-studies, reported that ML strays had accumulated under four separate _inbox tags and that none of the hub's 29 spokes owned machine learning itself.

Migrated from _inbox:

Why these five cohere. Every one of the five park records says the same thing in its own words — no spoke owns this — and each was declined by a different neighbour for a consistent reason: ../llm-inference-wiki is serve-time not train-time (both fine-tuning records), ../optimization-algorithms-wiki owns metaheuristics not gradient fitting (MLAlgorithms), ../platform-ops-wiki operates systems rather than designing them (the case studies). Four tags were four names for one absence. The five are three layers of one discipline: the fitting algorithm, the adaptation of a pretrained model, and the system built around it.

Two records deliberately left parked. ai-test-automation-paradox (ai-productionization, the other member of that tag) is AI-generated brittle tests — software-testing methodology, not shipping an ML system; folding it in would have widened the founding domain to "AI in engineering generally." ai-engineering-from-scratch (learning-roadmaps) stays with its genre cluster on the substance-vs-genre argument the mlalgorithms park record already made: a curriculum is a meta-document, MLAlgorithms is the material. Adjacency notes updated on both.

Founding pages: 5 sources + 5 concepts (machine-learning, classical-ml-algorithms, fine-tuning, ml-system-design, demo-to-production-gap) + 1 tool (unsloth) + spine.

Entities. Created amd (Corporation) and kdnuggets (NewsMediaOrganization); linked the-new-stack cross-wiki from ../ai-governance-wiki rather than duplicating it (entity-index returned a 1.00 match). Authors — Kanwal Mehreen, Meredith Shubel, and the two GitHub maintainers — were kept as prose mentions rather than nodes: single-artifact contributors, below the relevance gate. Budget: 2 of 8 used.

Recorded as founding weaknesses, not discovered later. ai-projects-fail-infrastructure-people was re-fetched during the split and its body is still unreadable — WebFetch twice (including after a redirect to the canonical URL), the Firecrawl fallback, and a direct HTML fetch and parse all returned navigation and signup forms. It is ingested as an honest T4 stub carrying its headline and nothing more, per the hub's never-drop-a-source rule. The corpus also has no middle layer (no end-to-end training source) and no independent benchmark — synthesis makes both the founding gaps rather than hiding them.

avoid-ai-writing run over the founding prose. Site rebuilt + verified at the clean rung, as the hub requires for a split.

[2026-08-02] ingest | Information Theory, Inference, and Learning Algorithms (MacKay, CUP 2003)

Routed here by the hub from Telegram — a book.pdf link with a Facebook click tracker on it, canonically inference.org.uk/itprnn/book.pdf. Downloaded (11.1 MB, 640 pages) and held at raw/mackay-itila.pdf; text extracted with pypdf for reading, since WebFetch can't take a PDF that size. Runner-up spoke: none — see the note below on what nothing owns.

Tier T1, the spoke's first. A Cambridge University Press textbook with derivations, rated exercises and worked solutions, by a researcher whose own LDPC results the book cites. Six sources now, and this is the only one that isn't a vendor post, a catalog, an explainer or an unread headline.

Pages. Source summary information-theory-inference-learning-algorithms (Book), plus three concepts the corpus had no page for: bayesian-inference, information-theory, monte-carlo-methods. Updated machine-learning (a fourth thing under the three layers) and classical-ml-algorithms (its "no conceptual account is paged yet" line is now half-answered — the probabilistic branch only; still nothing on trees, boosting or SVMs, which the book simply doesn't cover).

Entities. Created david-mackay (Person), cambridge-university-press (Organization) and university-of-cambridge (CollegeOrUniversity); entity-index returned no real match for any of them (the 0.75 hit on david-sacks is the fuzzy matcher catching a first name, not a person). Budget: 3 of 8. Shannon and Gallager are subjects of the book, not agents around it, so they stay prose mentions per ENTITIES.md.

Two things recorded rather than smoothed over. Part V is a 1990s Bayesian-neural-network account — single-neuron capacity, Hopfield nets, Boltzmann machines — written a decade before the scaling era, and the source page says so instead of letting "textbook" imply "current." And the book opens the corpus's first live seam: it says learning produces an ensemble of parameter values, while every other source here ships a point estimate and never mentions the difference. Filed in synthesis as open question 6, not as a contradiction — the two sides don't cite each other.

Noted for the router, not acted on: Parts I, II, III and VI — data compression, noisy-channel coding, LDPC and fountain codes — are roughly half the book and no spoke in the hub owns information or coding theory. information-theory holds only what bears on learning, and says in a scope note that a real coding-theory body wants its own spoke rather than growth here.

avoid-ai-writing run over the new prose.

[2026-08-03] ingest | End-to-end forecasting with TimesFM 2.5 (MarkTechPost)

Routed from the hub (Telegram). New: timesfm-2-5-forecasting-tutorial (T3 source summary), timesfm (SoftwareApplication/model), time-series-forecasting (DefinedTerm/task), marktechpost (publisher). Updated: fine-tuning (the zero-shot limit case), demo-to-production-gap (a concrete sample of the contract details that decide it), synthesis (founding gap one-third closed + evidence-table row + new section + open question 7), index. Substance: google/timesfm-2.5-200m-pytorch, 200M-param transformer, zero-shot, ten quantiles per step (index 0 = MEAN, 1..9 = q10..q90, 5 = median), max_context to 1024 (multiple of 32), max_horizon 256 direct / recursive beyond, covariates via XReg in two fusion orders. The tutorial: synthetic retail panel (1,200 days x 6 stores, train 1,144 / holdout 56); MAE, RMSE, MAPE, sMAPE, MASE + pinball loss + 80% PI coverage against seasonal-naive and last-value; rolling-origin backtest (6 folds x 56 days, 36 store-fold combos); context ablation {64...1024} where accuracy plateaus and cost doesn't (512 ~2x faster than 1024 at per_core_batch_size 16); anomaly detection by grading observations against the 80% interval (OK/WARNING/CRITICAL, ~20% flag rate on injected series); direct-vs-recursive long horizon; throughput sweep; NaN/short-context/ determinism robustness; production checklist. Why it matters here: the spoke's founding gap explicitly named "no evaluation methodology" and this closes that third of it. The apparatus (MASE against a seasonal-naive baseline, rolling-origin folds, pinball loss + coverage, ablation against cost) is model-independent, which is why it counts as method. Second finding: timesfm is the pretrain-once pattern arriving in a non-text modality, and the limit case of the corpus's "don't fine-tune first" rule — not adapt later, but don't adapt, in a field whose classical answer was a fresh ARIMA/ETS per series. T3 with the weakness recorded hard: all data is synthetic and the author generated it. Beating seasonal-naive on a series built from a weekly cycle plus noise shows the code runs, not that the model forecasts. Added as a third bias form in the evidence table (survivorship / first-party benchmark / synthetic problem). No ARIMA, ETS or GBM baseline; no independent replication. Entities: marktechpost paged as publisher; Google reused cross-wiki (google, ../llm-providers-wiki) rather than duplicated, per ENTITIES. backtesting cross-linked to ../quant-trading-wiki as the trading-side sibling. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-08-04] ingest | Transformers-Tutorials (NielsRogge)

Routed here by the hub (runner-up: none — no other spoke covers model training). T2, the same grade and shape as mlalgorithms: a personal MIT teaching repository, not a vendor artifact. The affiliation is recorded rather than penalized — Rogge contributed many of these model implementations to the Transformers library, which makes him expert on the code and an interested party on the choice of library. Nothing here compares Transformers to an alternative.

New pages (2): transformers-tutorials (source), niels-rogge (Person). Updated: fine-tuning context via synthesis, index, growth edge 1.

Closes the second third of the founding gap. The corpus jumped from how a fitting algorithm works to how to adapt a finished model, with nothing on training end to end. timesfm-2-5-forecasting-tutorial supplied the evaluation methodology on 2026-08-03; this supplies the training loop — executable rather than described, fine-tuning on real datasets across several dozen models, written three ways (native PyTorch, Lightning, Trainer) with a stated basis for choosing. fine-tuning had been concepts from an explainer with no measurements; unsloth hides the loop. This shows it.

Remaining hole is now specific: data collection and labelling, plus pretraining proper.

The pattern of closure is itself the finding. Both thirds arrived as method carried incidentally by teaching material — one inside a tutorial for a single model, one inside a tutorial collection for many — never as a source about the training pipeline as a subject. The middle layer appears to have no literature of its own within this corpus's reach; it gets documented only where someone was demonstrating something else. Folded into "The founding gap" rather than opened as a dated section.

No numbers, and none claimed — a teaching resource, so the spoke's record-the-setup rule resolves trivially. Growth edges 2 (independent benchmark) and 3 (evaluation as a discipline) untouched.

[2026-08-05] ingest | Supervision (Roboflow)

Routed from the hub. supervision — Roboflow's model-agnostic Python computer-vision utility library, MIT, 48,725★, T2. Created roboflow (Corporation; entity-index returned no match ≥ threshold). Budget: 1 of 8.

Module structure read from the tree, not the README — 78 modules under supervision/: detection (the sv.Detections structure, NMS, zones), annotators/draw, tracker, dataset (YOLO/COCO/ Pascal VOC conversion), metrics (mAP), plus classification, keypoint, geometry, validators.

This ingest stretches the spoke's scope boundary and the page says so. CLAUDE.md declares train-time and design-time; nothing in this library runs during backprop. It was routed here because the alternatives fit worse — ../llm-inference-wiki is serve-time mechanism for LLMs (logits, sampling, KV cache), and ../dev-tooling-wiki declines application-level runtime libraries under the Ky precedent. Logged as a stretch, not a precedent: if more serve-time vision tooling arrives, either widen the boundary on purpose or admit that computer vision as a modality is unowned hub-wide.

Gap-relevance is real rather than nominal. Growth edge 1's live half is data curation and evaluation, and this source ships both as installable modules — so the edge note now distinguishes holding them as tooling from holding them as a result, which the corpus still does not.

New synthesis section on demo-to-production-gap: the corpus held that gap as a claim with no mechanism (a success catalog with no denominator, an unreadable failure headline). This itemizes it — format conversion, identity across frames, drawing a box, counting line crossings, computing mAP. Weak evidence in one direction only (a library's existence proves demand, not that projects fail without it), and stated that way on the page.

Vendor-adjacency recorded: Roboflow sells a CV platform and the library feeds its inference product. No performance or accuracy claims anywhere in the repo, so nothing to verify.

[2026-08-05] ingest | SkillSmith

Routed from the hub (runner-up: ../agentic-tooling-wiki, declined on the homonym — see below). skillsmith — arXiv preprint "Learning to Compose Parametric Skills and Textual Knowledge" (Dery, Tjandra, Samiei, Kuncoro, Yahav, Shen, Szlam), T2.

The spoke's first primary research on adaptation (prior adaptation sources were an explainer and a toolkit). Method: prefix-tuning stores a skill as weights; an LLM reads existing prefix weights + text and emits new prefix weights for a target skill; reported to beat text-only and weight-space-only baselines. Prefix-tuning is PEFT, so it sits in the fine-tuning family — updated that page with the "adapter as an object you compose, not just train and load" reframing.

Synthesis: added as a third stance on the founding weights-as-distribution vs point-estimate seam — it treats trained weights as manipulable objects, making parameters the unit of analysis. Held loosely: preprint, authors' own evaluation, benchmark caveat in full force.

Homonym recorded: a parametric skill (prefix weights) here vs an agent skill (markdown prompt) in ../agentic-tooling-wiki — both senses arrived in the hub the same day (education-agent-skills there). That homonym is exactly why agentic-tooling was declined: "skills" in the title is weight-space, not prompt-space, so this is model-adaptation research and files here by subject.

Entities: the 7 authors are deferred — a preprint with no affiliations recoverable from the abstract; thin evidence, and paging seven one-off Person nodes would breach the evidence-only rule. Page the lead author on recurrence or if a query needs it.

[2026-08-07] ingest | Phase Transitions in Attention: A Bayesian Theory of Copy Head Emergence

Routed from the hub (Telegram). arXiv 2606.12058v1, Lavie & Fischer et al., 10 June 2026 — T1, and a preprint, so no peer review. The spoke's first primary research on training itself rather than on a tool, a tutorial or a finished model.

New pages: phase-transitions-in-attention (source summary), capability-emergence, induction-head. Updated bayesian-inference with the section that source earns.

Why it matters here. Open question 6 — does the Bayesian account survive at scale — had a textbook on one side and a point-estimate toolchain on the other, neither citing the other. This paper puts them in one equation for a single attention head: the analytically-derived posterior's MAP is where Adam lands (minimising the action at large χ = gradient flow with weight decay), and SGLD agrees. What makes it computable is a permutation symmetry, not a scalable method, so the question is now sharper rather than closed. Recorded as a partial answer, and a new open question 8 for the monitoring claim.

The finding that travels furthest is not about training: transition order depends on the attention activation. Softmax jumps first-order with no precursor signal; linear attention goes second-order then crosses over gradually. The paper's own reading is that capability onset in softmax models is "intrinsically unpredictable from observable quantities," with "fundamental limits to what can be monitored." Held as a conjecture with a derivation behind it — one layer, one head, i.i.d. uniform tokens, supervised copy task, equivalent-kernel and d_model → ∞ approximations, and copying is the minimal circuit under in-context learning rather than in-context learning itself.

The PDF would not extract through WebFetch (raw object stream); pulled the text locally with pypdf.

Entities: 3 created (itay-lavie, zohar-ringel, moritz-helias), 0 linked. Six authors and five affiliations; I paged the corresponding author and the two senior authors whose groups the work joins, and left Fischer, Lekov and Van Maele as plain mentions along with all five institutions. Same call as the hub's 2026-08-06 precedent on thin one-off paper authors — revisit if this line of work recurs, since Fischer is a co-first author and would be the next node.

[2026-08-08] ingest | Data Cascades + Pervasive Label Errors — the founding gap's last third

Two peer-reviewed sources (both T1), arrived via the research pass hunting growth edge 1, the data-collection-and-labelling half of the missing middle layer.

data-cascades — Sambasivan, Kapania, Highfill, Akrong, Paritosh, Aroyo, CHI 2021, google Research. 53 practitioner interviews across India, East and West African countries and the USA, all in high-stakes domains. 92% had hit at least one compounding data failure; 45.3% two or more in one project. Cascades are "pervasive, invisible, delayed, but often avoidable."

pervasive-label-errors — Northcutt, Athalye, Mueller, NeurIPS 2021. Ten standard test sets, confident learning plus crowdsourced confirmation: ≥3.3% of labels wrong on average, ≥6% of ImageNet validation, 51% of algorithmically flagged candidates confirmed. Correcting the labels reverses model rankings — ResNet-18 over ResNet-50, VGG-11 over VGG-19.

New pages: data-cascades, pervasive-label-errors, training-data-quality (the concept node consolidating both), nithya-sambasivan, curtis-northcutt. Entity budget 2 of 8; google linked cross-wiki rather than duplicated.

A correction to synthesis, not just an addition. This wiki had written that the middle layer "appears to have no literature of its own in this corpus's reach" and is documented "only where someone was demonstrating something else." That was wrong. The literature exists in HCI and in the datasets-and-benchmarks track; it studies practitioners and labels instead of architectures, so ML-side searching never returned it. The old text is kept and the correction written under it.

Contradiction recorded, not resolved. ai-projects-fail-infrastructure-people (T4, body never readable) says systems fail on infrastructure and people; data-cascades (T1, 53 interviews) says the cause sits in data decisions taken before any model exists. Both kept on demo-to-production-gap. Also flagged there: 92% is prevalence among practitioners recruited to discuss data practice in high-stakes domains, not a failure rate for AI projects, so the corpus's open question about a denominator stays open.

Edge 2 moves rather than closes. The label-error result means a "neutral third-party benchmark" — the strong rung on this spoke's evidence spectrum — inherits its test set's error rate. The edge now asks for a benchmark that states its label quality.

Successor edge 1a written: nothing here says what competent labelling looks like — no annotation guidelines, no inter-annotator agreement, no sampling design, no cost.

[2026-08-08] ingest | Measuring Annotator Agreement — the instrument for the data layer

Research pass (20-edge run), successor edge 1a, written hours earlier the same day. measuring-annotator-agreement new (T1, Braylan/Alonso/Lease, arXiv 2212.09503, ACM 2022).

Completes a sequence rather than adding to a pile: data-cascades says the practice fails (92% of practitioners), pervasive-label-errors says the output is wrong (≥3.3% of labels), and this says the instrument you would check with is well-founded only for the simplest label shapes. Agreement across bounding boxes, keypoints, sequence tags, ranked lists, free-text translation, numeric vectors and syntax trees needs a distance function chosen first, and the metric inherits that choice.

Cross-corpus rhyme noted on the page: er-evaluation-framework in ../research-wiki found the naive entity-resolution metric biased upward. Different field, same shape — the error hides in the measurement of the measurement.

Edge 1a marked partly closed: the corpus now has a metric and still has no documented labelling practice — no guidelines, no annotator training, no sampling design, no cost.

[2026-08-09] ingest | Gradient-boosted trees — XGBoost and the tabular benchmark (via research pass)

Coverage edge 4 closed with exactly the pair it named. xgboost — Chen & Guestrin, KDD 2016, T1, extracted from the PDF locally. tabular-deep-learning-benchmark — Grinsztajn, Oyallon & Varoquaux, NeurIPS Datasets & Benchmarks 2022, T1. Concept page: gradient-boosted-trees.

What was missing. classical-ml-algorithms carried the entire non-deep story, and the method practitioners reach for first on tabular problems had no page — while open question 7 was explicitly asking for it as the baseline a time-series foundation model should be measured against.

Two distinct claims, recorded separately. XGBoost's contribution is a system: the algorithm was Friedman's, and the paper's own numbers are engineering ones — 0.68 s per tree against scikit-learn's 28.51 s at the same AUC on Higgs-1M, 1.7 billion Criteo examples processed on a single machine. The 17-of-29-Kaggle-winners figure is adoption evidence from a venue with its own selection effects, and is labelled as such.

Grinsztajn et al. is the claim that ages well: 45 datasets, ~20,000 compute hours of tuning per learner, tree ensembles still state-of-the-art at ~10K samples, with three named inductive biases behind it — robustness to uninformative features, orientation preservation against rotation-invariant networks, and irregular target functions. That is a structural explanation, so it predicts rather than merely reports.

Open question 7 is unchanged in status and better specified: the baseline now exists in the corpus, the comparison still does not. Edge 4's successor asks for it directly.

Entities: 0 created — no new author or org pages; one paper each is not a recurrence.

[2026-08-10] ingest | Awesome Agentic Time Series — 243 papers, and the shape of them

Routed here by the hub (runner-ups: ../agentic-tooling-wiki for the agent-architecture vocabulary, ../quant-trading-wiki for the finance slice). Source: https://github.com/TROUBADOUR000/Awesome-Agentic-Time-Series, T3, MIT, 188★/10 forks, created 2026-06-01, read 2026-08-10. Companion to an unpublished survey, The Landscape of Agentic Time Series Systems.

Pages: awesome-agentic-time-series. time-series-forecasting gained a section on the field around it.

Why this spoke. The subject is time-series analysis, which this wiki already owns (time-series-forecasting, timesfm). The agents are the method. That follows the hub's standing precedent — a tool that merely uses LLM agents files by subject, with the mechanism cross-linked — and the list's own domains settle it: weather, IIoT, motion and multimodal classification outnumber finance, so ../quant-trading-wiki would have held a tenth of it.

The finding is the distribution, not any entry. I counted the sections rather than reading the papers: 133 papers on building agentic time-series systems, 16 on evaluating forecasts or agentic decisions, 7 on reliability. Under a tenth of the field is about whether any of it works. That is demo-to-production-gap arriving from the literature side, and it lands directly on time-series-forecasting's standing complaint that nothing here compares a pretrained forecaster against classical baselines — the comparison is the field's thinnest corner, not this wiki's blind spot.

Also recorded: 31 foundation-model entries (Falcon-X, Toto 2.0, Timer-S1 at billion scale, TS-Memory) mean the pretrain-once move in this modality is crowded, and timesfm is no longer the whole story here.

T3 and honestly so. A three-month-old maintainer-curated list with no published inclusion criteria, a "200+" badge that actually counts to 243, and a BibTeX entry with no venue or year. Nothing in it was read; every claim about a paper is that paper's title. The counts are what this page stands behind.

[2026-08-10] ingest | Who Annotates in NLP? (research pass)

Arrived through the hub Research Pass, against ## Most wanted #3 — a documented labelling operation, the edge left half open on 2026-08-08 when measuring-annotator-agreement supplied an instrument and no practice.

who-annotates-in-nlp written (T1, arXiv 2606.02255; preprint, flagged as such). It audits 2,667 annotation tasks from 1,603 ACL-venue papers and validates its own LLM extraction against a 72-task adjudicated gold standard at Krippendorff's α 0.606, above the 0.585 human–human figure on the same data.

The edge does not close, and the reason is the finding. The practice this wiki kept failing to locate is largely unpublished: annotator training reported for 18.7% of tasks, released guidelines 34.1%, an agreement value 35%, against recruitment at 90.4%. The search was not at fault; the base rate is. Re-specified on the hub shortlist rather than struck.

Placed against data-cascades, pervasive-label-errors and measuring-annotator-agreement — the four now say, in order: the damage is real, the labels are wrong at measurable rates, the instrument is ill-suited to most label shapes, and the conditions needed to interpret the instrument are usually missing.

[2026-08-14] ingest | Simple and Effective Paraphrastic Similarity from Parallel Translations

Arrived via Telegram as a bare arXiv link. Routed here over ../research-wiki (RAG/LLM-products) and ../llm-inference-wiki (serve-time mechanism): this is train-time representation learning, and no spoke owns NLP as its own subject, so it lands on the closest fit — model training methodology.

paraphrastic-similarity-parallel-translations written (T1, ACL 2019). Trains sentence embeddings directly on bitext with a margin loss and hard-negative mining, skipping paraphrase-corpus construction; beats deep cross-lingual baselines on STS while encoding ~330× faster. New concept page paraphrastic-sentence-embeddings and author page john-wieting (first/corresponding author, also the source of the mega-batching procedure and the 10,000-GPU-hour paraphrase corpus this paper supersedes).

Worth keeping past the summary: Section 3.3's random-encoder result — an untrained averaged trigram embedding is competitive with or beats every literature baseline monolingually, because STS without a cross-lingual gap rewards lexical overlap, not semantics. Filed as another instance of the same caution tabular-deep-learning-benchmark and pervasive-label-errors raise from opposite angles: an unchecked baseline can win for the wrong reason.

[2026-08-15] ingest | Information Theory, Inference, and Learning Algorithms — re-seen, refreshed in place

The same inference.org.uk/itprnn/book.pdf link arrived again via Telegram (again with a Facebook click tracker attached). It is already held at raw/mackay-itila.pdf and paged as information-theory-inference-learning-algorithms since 2026-08-02, so this is a refresh, not a second ingest — no new page, no duplicate.

What changed: the Cross-spoke context section said no spoke in the hub owned information or coding theory. That stopped being true within hours of the original ingest. ../mathematics-wiki was spun out the same day and now names information & coding theory as a growth edge, and its scope note says explicitly that the MacKay routing stands and the next such source lands there. ../cryptography-wiki (spun out 2026-08-15) owns the other slice — codes as construction material inside proof systems, not the coding theorems. Both are now named on the page.

Nothing in the book's substance changed and no claim was revised, so synthesis.md and index.md are untouched. updated: bumped to 2026-08-15.

[2026-08-18] ingest | Lindholm et al., Machine Learning: A First Course for Engineers and Scientists

Routed by the hub from Telegram. https://smlbook.org/book/sml-book-draft-latest.pdf — 348pp draft of 4 March 2026, cambridge-university-press, T1, freshness: volatile. The submitted link carried a Facebook fbclid tracking parameter; stripped before recording, per the hub convention.

New pages: lindholm-machine-learning-first-course (source), andreas-lindholm, niklas-wahlstrom, fredrik-lindsten, thomas-schon, david-sumpter (Person). Updated: synthesis (open question 4 and growth edge 2), index. cambridge-university-press already existed from the MacKay ingest and was linked, not recreated — entity-index got that one right, and got the two author queries wrong, returning thomas-judson and david-stirzaker as MATCH for "Thomas Schon" and "David Sumpter". Same first-name fuzzy failure as the Charles University case today.

The founding gap, taught. This spoke has argued since 2026-07-30 that it holds the top of the stack and the bottom and nothing in between. Chapters 2–10 are the in-between: k-NN and decision trees before any parametric model, linear and logistic regression, regularisation and GLMs, bias–variance, loss functions and likelihood, SGD and hyperparameter search, neural networks with backpropagation derived in an appendix, bagging through gradient boosting, kernels to the representer theorem, Bayesian linear regression and Gaussian processes, mixtures, clustering and dimensionality reduction. Second textbook here, after MacKay, and the first ordinary one.

Chapter 11 is the part that is rarer than the algorithms. "User Aspects of Machine Learning" — defining the problem, improving a model, what to do when more data is unavailable, practical data issues, and "Can I Trust my Machine Learning Model?". Its improvement procedure carves a training-validation set out of training data so hold-out error splits into training error, generalisation gap and train-val mismatch, and the third term tells you whether you have a model problem or a distribution problem. The corpus had the vocabulary for this and no procedure.

§12.2 speaks to this spoke's oldest complaint. Chapter 12 was contributed in full by david-sumpter, and its second section argues that because much ML research is done by large companies with PR departments, "research is (in part) a form of advertising," with AlphaGo's company-financed film and Musk's autonomy claims as the worked examples. Growth edge 2 has been recording, source by source, that every number here is a vendor's own. This does not close it — that edge wants a neutral benchmark stating its own label quality — but the corpus can now cite the mechanism instead of inferring it. Recorded as explained, not closed, per the half-close rule.

What it does not do: nothing on fine-tuning or PEFT (open question 2), nothing on time-series foundation models (question 7), and §11.4 is a first course's page on data issues rather than the documented labelling operation edge 1a asks for.

Licence, recorded because the framing invites a mistake: free to read is not open. "Personal use only. Not for re-distribution, re-sale or use in derivative works."

[2026-08-25] ingest | SafeTutors + RE for AI in Practice — growth edge 1a closed (via research pass)

Two sources, hunted together against the edge that had failed three times: one labelling operation documented end to end. Both ingested; the edge is closed and a narrower successor opened.

safetutors (Hazra et al., TU Eindhoven / IIT Kharagpur / Cisco Research / NYU, arXiv 2603.17373, T1) is a benchmark paper about whether LLM tutors quietly damage learning. Its subject belongs with ../engineering-education-wiki and ../ai-governance-wiki, and the page says so. What this spoke took it for is Appendix D, which prints the whole annotation operation: recruitment criteria and pay per stage ($5/hr undergraduates, $15/hr doctoral students, $10/hr Prolific workers), a four-step calibration ending at an ≥80% agreement-with-gold gate, a 10-item qualification test with its pass rates (24 of 33 crowd recruits, 72.7%), verbatim guidelines per stage, the annotation tools by name, an adjudication protocol with rates (4.2% of crowd instances escalated, 3.8% cross-stage conflicts of which 21 dropped and 36 flagged), Fleiss' κ split three ways, and 15,300 labels totalled.

annotation-requirements-autonomous-driving (Saeeda et al., Chalmers / RISE / Gothenburg / Volvo Cars, arXiv 2511.15859, T1) covers the industrial half by interview — 20 practitioners at six automotive companies. Guidelines past 140 pages; manual review cut from 10% to 5% under budget pressure; rework at about 3× the cost of labelling properly; ~20% of collected data ever annotated; 15% of a 3D point-cloud set destroyed by version conflicts.

What made the pair worth two pages rather than one. The edge named two acceptable artifacts — a full annotation appendix and an industrial data-ops write-up — and each source supplies exactly one. Together they also explain who-annotates-in-nlp's reporting rates: the academic operation could publish because it was small, bounded and finished, while the industrial one has unstable guidelines and, on its own practitioners' account, no traceability, so there is no settled artifact to publish.

Recorded as weaker evidence than it looks. Every industrial number above is one informant's recollection quoted with an interview ID, not an export from a system of record. The page says so in its own words rather than in a footnote.

Flagged on the SafeTutors page: the paper reads its falling κ (0.81 → 0.73 → 0.68) as an expertise effect, but stage, task and workforce all change together, so expertise is confounded with task subjectivity. Its cleaner patterns — single-turn beats multi-turn by Δκ ≈ 0.05, maths beats physics — stand.

Successor edge 1b: neither source states total cost. Rates without a total on one side, a ratio without a base on the other. — needs: total cost of ownership for one labelling operation.

New entity pages: rima-hazra, hina-saeeda, chalmers-university-of-technology. Updated: training-data-quality, who-annotates-in-nlp, measuring-annotator-agreement, synthesis.md (edge 1a, open question 4), index.md.

[2026-08-28] ingest | Annotation quality in ABSA: experts, students, crowdworkers and an LLM — research pass, T1

Donhauser, Fehle, Hellwig, Weinberger, Kruschwitz & Wolff (Universität Regensburg), arXiv:2605.03624, 5 May 2026, 16pp — read from the PDF (the abstract page carries none of the numbers; extracted locally with pypdf per the standing rule). Ingested with claude-opus-5 at high effort, via the research pass hunting growth edge 1b.

Created absa-annotator-comparison-cost (T1). Updated synthesis (edge 1b closed, edge 1c written, plus a new section on agreement-vs-correctness) and index.md. Not a duplicate: safetutors and annotation-requirements-autonomous-driving describe operations; neither prices one.

Edge 1b is CLOSED. "Crowdworker studies were completed within two days per study at a total cost of roughly £828 (≙ £0.41 per three way annotated example), including platform fees." Thirty Prolific annotators at £9/hour, 200 sentences each, over 1,000 GERestaurant sentences in two ABSA subtasks. A total, a per-usable-label figure, and the scope — exactly what the edge specified.

The comparison is worth more than the number. Four annotator types, same task:

• crowdworkers £0.41/example, two days, price = whole cost
• self-hosted LLM ≈ $0.0025/example (~10k tokens at $0.05–0.25/MTok)
• frontier API ≈ $0.36/example (GPT-5.2 Pro, $21/$168 per 1M in/out)
• students: course credit, several weeks
• experts: several hours, and only because they revised pre-existing labels — the authors say plainly that "without pre-existing labels, large-scale expert annotation would be costly and difficult to scale"

Two observations recorded from that: "LLMs are cheaper" spans 144× and straddles the human price, so the claim is empty without naming the tier; and the two cheapest options carry no money at all, paying in elapsed time and in work somebody else already did.

The quality finding lands on this wiki's own instrument. "High agreement does not necessarily translate into superior model performance, as shown by LLM annotations, which achieved high IAA but only comparable performance to student-annotated data." measuring-annotator-agreement gave the corpus agreement as the label-quality measure; here it scores the cheap option highest. And the mechanism is coverage. On TASD, student and crowdworker datasets hold systematically fewer instances than expert and LLM ones, concentrated in rare categories (AMBIENCE, PRICE). The labels are not wrong, they are absent — from every annotator's copy — so no pairwise agreement computed over what was annotated can detect it. Same shape as pervasive-label-errors, one level up.

Limits recorded on the page: one language, one domain, 1,000 sentences, one crowd platform, and the frontier model priced from a list rather than run. Successor edge 1c: nothing lets £0.41 travel to the 140-page-guideline industrial case.

Entity note: no new nodes. Six authors at one university, none of whom carries a claim here beyond authorship; Universität Regensburg left as prose per ../ENTITIES.md's relevance gate.

[2026-08-31] ingest | Code-as-World — supervision you simulate instead of annotate (routed from hub)

Routed from the hub by the router; arrived via Telegram as github.com/MirroS-Lab/Code-as-World. Runner-up spoke was ../llm-providers-wiki (two open-weight checkpoints on HuggingFace) and the dominant substance is train-time — a method for manufacturing training supervision and the models fitted to it — so it lands here.

Two source pages for one release, deliberately. The repository code-as-world is T1 and carries only the artifact's own facts (Apache-2.0, 198★, Python, what runs). The technical report code-as-world-report is T3 and carries everything measured, because the benchmark is mirros's own, on models MirroS trained, against baselines MirroS configured — ../QUALITY.md's three-layer rule, same call already made for unsloth-amd-support. Report read from the PDF (30pp), not from a summary.

New pages: code-as-world · code-as-world-report · executable-world-representation · agentic-discovery-loop · code-as-world-vl · quantiphy · mirros (entity). Refreshed: training-data-quality · fine-tuning · machine-learning · synthesis.

What it does

Represents a physical scene as code a simulator can run, then recovers that code from a video or a text prompt with a five-stage loop — propose → instantiate → execute → render → verify — run by an LLM against MuJoCo until the replay matches the evidence or the budget rejects the hypothesis. The verified worlds expose exact geometry and state trajectories, so quantitative question–answer pairs can be generated with labels read off the simulator. code-as-world-vl is a family of VLMs trained on those pairs: SFT on pixel-space measurement, then GRPO on world-space calibration.

The number worth keeping

Qwen3.5-4B scores 31.2 on quantiphy; Code-as-World-VL-4B scores 50.6 — same size, same protocol, and the one controlled comparison in the release. The headline (9B at 55.4 beats Gemini-3.1 Flash at 54.8) is the weakest claim in the table, not the strongest: a 0.6-point margin in a table where ChatGPT-5 Pro lands at 19.5, below its own sibling and below a 0.26B SmolVLM, and MiniCPM-V 4.5 scores exactly 0.0 on both 3D subsets — which reads as a harness failure, and is not discussed. Recorded on the pages as claim, not measurement.

One control the report does run properly: the discovery loop is compared against Best-of-5 independent sampling at a matched five-evaluation budget, with the scoring metrics held out of the verifier's signal. Iterative refinement wins. That is the right experiment and it is rarely offered.

Why it changes the spoke's most-worked question

Six sources at training-data-quality treat labelling as work done to existing data — recruit, train, guide, adjudicate, and £828 for 2,000 examples. This one refuses the premise for a task nobody can annotate: no crowdworker can see metres per second. Every one of those costs goes to zero, and a new one appears — §4.4 concedes the loop "may converge to a locally plausible EWR without recovering a mechanistically accurate explanation." The labels become exact and the world becomes the thing that can be wrong, which no agreement measure, distance function or adjudication protocol can detect. Same failure as pervasive-label-errors, moved one level upstream and out of instrument range. Folded into synthesis as its own section; growth edge 9 opened for an independent reproduction.

Cross-wiki finding

llms-cant-jump (../research-wiki, Zahavy/DeepMind) argues abduction — and specifically manipulative abduction, hypothesis by embodied simulation — is structurally missing from these models. This report builds that loop in software, reaches for the same abductive lineage unprompted, and then says in §5.5 that the model "does not internalize the agentic discovery process itself." The two agree, and the agreement is the finding: the jump stayed in the harness. Linked from both synthesis and agentic-discovery-loop, not duplicated.

Entity note

One new node: mirros (Organization), same-as its GitHub and HuggingFace orgs, both checked. No Person pages — the report names no authors beyond "MirroS Team". stanford-university and qwen already existed and are linked cross-wiki. quantiphy's authors (incl. Fei-Fei Li) are the subject's agents, not this source's, so no pages per ../ENTITIES.md. Budget: 1 of 8.

Recorded as unverifiable

The BibTeX cites arXiv:2608.xxxxx — a placeholder. No venue, no peer review, no named authors, no stated location or funding for the lab, and the discovery agent is not in the released code. The 27B result changes scale and response protocol together and its checkpoint is not published. None of this blocks the ingest; all of it is on the pages.

[2026-08-31] ingest | GIFT-Eval — open question 7, half answered (via research pass)

Hub research pass, quick-win 12. gift-eval (Aksu, Woo, Liu et al., Salesforce AI Research, arXiv 2410.10393, T1) read from the PDF with pypdf; the abstract carries none of the result tables. 23 real datasets, 144k series, 177M points, 97 configurations, 17 baselines. PatchTST wins overall (CRPS 0.563, rank 5.72) ahead of every time-series foundation model, with Moirai-Large second — so foundation models beat Auto_ARIMA and lose to a supervised transformer trained per dataset. Recorded with both caveats the authors state: acknowledged partial pretraining leakage for TimesFM, Chronos and Moirai, and Moirai being Salesforce's own model. Half-closes coverage edge 4's successor and open question 7; residue written as edge 4a, because the benchmark runs no boosted-tree baseline and that is the half question 7 still needs.

[2026-09-01] ingest | TimesFM-3 — the multivariate answer, without a score

Routed from the hub (Telegram). timesfm-3 (Google Research blog, Ayush Jain & Rajat Sen, 2026-08-31, T3). 330M parameters, natively multivariate via alternating causal-temporal and full-variate attention, past and past-future covariates, several targets at once, nine quantiles, non-autoregressive decoding of the whole horizon (Contiguous Patch Masking), 32-step patches,

1 trillion training time points; 20 layers, dim 1280, 16 heads from the model card.

Two findings the post does not lead with. The benchmark claim carries no numbers — charts only — and ranks TimesFM-3 among pre-trained foundation models, a class that excludes PatchTST, which won gift-eval outright. So open question 7 does not move, and the mid-table rank on timesfm stands. The licence changed: the 2.5 card is apache-2.0, the 3.0 card is timesfm-non-commercial-license-v1.0 (both checked 2026-09-01, in the model-card frontmatter).

Refreshed timesfm, time-series-forecasting and gift-eval in place; new synthesis section plus a "not moved" note on open question 7.

Entity note

No new entity pages. The authors (Jain, Sen) are evidenced but the relevance gate wants a checked identity URL and none was verified, so they stay plain mentions per ../ENTITIES.md — a guessed same-as: is worse than none. google already exists in ../llm-providers-wiki and is linked. Budget: 0 of 8.

[2026-09-07] ingest | The complete male fruit fly connectome (Google Research)

Telegram source, routed here by the hub over ../psychology-wiki. male-fly-connectome — T3, the grade this spoke already gives Google Research posts (timesfm-3): first-party about the poster's own work, so the pipeline description is reliable and the quality claims are ungraded homework.

Verified the paper rather than trusting the blog. Crossref returns "Sexual dimorphism in the complete Drosophila male central nervous system connectome," Cell 189(18), September 2026, 111 authors, first Stuart Berg, last Gregory S.X.E. Jefferis. The blog renders the title in a different word order. The paper is where the error rates live; it was verified as a citation and not read, and the page says so.

Why a fly brain is in a machine-learning wiki. The subject is neuroscience and no spoke owns that. What routes here is the pipeline, and it lands on the spoke's most-worked question — where labels come from — with both answers running in one system:

• Segmentation by flood-filling networks, with synthetic neurons injected into the training data to improve the PATHFINDER reconstruction system. That is manufactured supervision, the strategy code-as-world-report introduced on 2026-08-31.
• Then years of expert proofreading at HHMI Janelia, which the post names as the constraint on the entire field.

The synthesis section from 2026-08-31 had the two strategies partitioned by domain. This source shows them consecutive: manufacture trains the model, the model produces a draft, people correct the draft. And the human step is not the labelling activity this corpus has been measuring — it is label-error correction on model output, budgeted in advance instead of discovered afterwards. training-data-quality gains a section for that, and who-annotates-in-nlp's question gains a third answer: expert staff at the institute that owns the science.

The release history is the cost, written down. Fully automated female fly brain 2019 → human-verified half brain 2020 (25,000 neurons, 21 million connections) → complete male CNS 2026 (166,000 neurons, 125 million synapses). Automation came first and six years went into verification.

Same omission as last week. timesfm-3 claimed benchmark wins in charts with no numbers; this post claims improved "speed and accuracy" with no number at all — no error rate, no proofreading hours, no split between automated and corrected output. Two subjects, one venue, one pattern.

New pages: connectomics (DefinedTerm, the task), viren-jain and michal-januszewski (Persons, same-as their research.google profiles, both checked by title), hhmi-janelia (ResearchOrganization, same-as Wikipedia). Viren Jain is explicitly distinguished on his page from Ayush Jain of timesfm-3 — same employer, same spoke, a week apart, different people.

Scope note, not settled. Neuroanatomy has no owner in the hub. Recorded in synthesis: if two more sources of this kind arrive, reopen the question rather than settling it by precedent from this one.

Also announced and recorded without pages: a Columbia-led Nature paper on the elephantnose fish hindbrain, an upcoming Harvard zebrafish collaboration, the ZAPBench and Fish Fire&Wire datasets, and Neuroglancer as the open-source viewer.

[2026-09-07] ingest | Neuroglancer — the viewer the connectome is published through

Second Telegram source in the same cluster, hours after male-fly-connectome and pointing at the tool that post names. neuroglancer — T1, the repository itself.

What it is. Apache-2.0, TypeScript, 1,364★ / 381 forks / 235 open issues, created 2016-05-27. A WebGL viewer for volumetric data that is "purely a client-side program" — no server of its own, so a 166,000-neuron connectome is published by publishing a URL. The frontend/WebWorker split is forced by the platform, and the README says why: browsers have no shared memory and require DOM and WebGL work on the UI thread. Needs WebGL 2.0 plus EXT_color_buffer_float.

The fact worth keeping. "This is not an official Google product" — second line of the README, under the google/ org. Several published connectomes are distributed through it and it carries no support commitment. Recorded flat, without reading it as either reassuring or alarming.

The format outgrew the viewer. It reads N5, Zarr v2/v3, OME-Zarr, BOSS, DVID, Render, NIfTI and Deep Zoom, and its own precomputed format now has an ecosystem: CloudVolume and BigArrays.jl as Python and Julia interfaces, Igneous for meshing and downsampling, neuroglancer-scripts for converting BigBrain, mesh-n-bone from Janelia CellMap, and Google's own TensorStore underneath. That list is the project's own and nothing here measures adoption; noted with that caveat.

The open question it leaves. Proofreading the fly connectome took years of expert work (training-data-quality). A viewer anybody can open in a browser is the cheapest possible route to a second pair of eyes on that work, and no source here measures whether public explorability produces corrections. Added to synthesis alongside the 2026-09-07 section rather than as a new one — same day, same cluster.

Not examined: rendering performance, dataset-size limits, how the precomputed format encodes multi-resolution meshes, and the documentation site (checked to resolve, not read).

[2026-09-14] ingest | How to Combine Traditional ML with Agentic Reasoning (MachineLearningMastery)

Routed in from the hub (Telegram).

Source: ml-models-as-agent-tools, T3 — unmeasured conceptual explainer on a tutorial site, one worked example, no code. published: 2026-09-10. Ingested for the design pattern (trained model as a tool an LLM agent calls) and the gap it leaves (nothing measured, threshold rule still inside).

Refreshed: ml-system-design — new Who calls the model section. Spine: synthesis section The model's caller changes; index under Article. Author deferred (one source).

[2026-09-16] ingest | LFM2.5-230M-Chess (mlabonne model card)

Source: lfm2-5-230m-chess, T3, via Telegram (msg 2128), routed by the hub (runner-up llm-providers). Entity liquid-ai paged in llm-providers-wiki (thin); maxime-labonne reused cross-wiki. Synthesis section Adaptation by vocabulary.

[2026-09-17] ingest | Younes, Introduction to Machine Learning (arXiv 2409.02668)

Source: younes-introduction-to-machine-learning, T1, published: 2024-09-04, via Telegram (msg 2149), routed by the hub (runner-up mathematics-wiki). PDF v2 held in raw/ and read (TOC, preface, full text searched). New entities laurent-younes, johns-hopkins-university. Open question 4 annotated; synthesis section Three textbooks, three registers.

[2026-09-24] ingest | Laya (Convai Innovations model card)

Source: laya, T3, via Hugging Face model card (curator-sent, "ingest this"), cross-checked against the GitHub repo (github.com/NandhaKishorM/laya) and PyPI (laya) for dates and authorship. published: 2026-09-18 (HF/GitHub/PyPI creation dates agree). Runner-up spoke llm-providers-wiki, recorded as cross-spoke context on the page — dedup found system-one-models-and-jev already held there (TypeSafe's Jev, the model Laya benchmarks itself against by name). Also folded in an already-held, previously unfiled raw source on the same subject: laya-doom-medium (T4, Medium, Christian Graham, published: 2026-09-18).

New entities: convai-innovations (Organization), nandha-kishor-m (Person, CEO, verified via GitHub profile + org membership). Refreshed: fine-tuning — new RL for calibration, a second instance section, linking Laya's RLCD training to code-as-world-vl's GRPO calibration stage. Spine: synthesis section A second System 1, and a second copy of the same benchmark problem; growth edge 2 annotated with Laya as a second no-independent-benchmark data point; index updated under SoftwareApplication, Article, Person, Organization.

[2026-09-24] ingest | Open-Jev-9B (ZefanCai model card)

Source: open-jev-9b, T3, via Hugging Face model card (curator-sent, "ingest this", Haiku-classified to this spoke, runner-up llm-providers-wiki, high confidence). Cross-checked against the GitHub repo (github.com/Zefan-Cai/Open-Jev) and the dataset card for dates, authorship and the project's stated relationship to Jev. published: 2026-09-20 (HF/GitHub creation dates agree). Dedup: rg for "open-jev|zefan" and entity-index -- "Zefan Cai" found nothing existing.

Integrated with the same-day laya ingest as a third instance of the Jev-shaped typed-decision category: unlike Laya, Open-Jev's own README states plainly it is "an independent implementation inspired by TypeSafe's Jev," and its architecture (LoRA + Yes/No-initialised head on a 9B generative Qwen3.5 base) and its card's silence on RL/REINFORCE stand in contrast to both Jev's and Laya's RLCD. New pages: open-jev-dataset (Dataset, its training/eval data) and zefan-cai (Person, thin — no stated affiliation).

Refreshed: laya — two new paragraphs under Training: RLCD naming Open-Jev as the independence-question answer and the architecture/eval contrast. fine-tuning — new paragraph contrasting Open-Jev's apparent supervised loss against Laya's/code-as-world-vl's RL-for- calibration pattern. Spine: synthesis section retitled A category, not a product: three System 1s in one week, new paragraph; growth edge 2 annotated a third time. Index updated under SoftwareApplication, Dataset, Person.

[2026-09-27] ingest | Jev AI TypeSafe AI: A Practical Guide to Typed Decisions in Production (Hugging Face)

Source: jev-typesafe-typed-decisions-guide, T4, via hub router (routed here, runner-up agentic-tooling-wiki). published: 2026-09-25. A Hugging Face community post on calling Jev safely in production. Checked the author's own HF profile (account sora-2, byline "bna"): 19 articles keyed to trending product names, two more on Laya and Jev published within a day of this one — a content-mill account, not a practitioner or lab. No Person/Organization entity page created for it: not domain-relevant, evidence gate fails. Dedup: rg for "jev" and "typed decision" across all spokes found the existing llm-providers-wiki coverage (system-one-models-and-jev, jev-ai-sdk-guide, typesafe-ai) and this spoke's laya / open-jev-9b; this source adds no new fact about any of them and is filed for its productionization pattern instead, which is why it routes here rather than duplicating the market-side coverage.

New page: typed-decision (DefinedTerm) — names the pattern the three System 1 implementations share (declared answer space, matching response shape, a role in control flow) and the choice/score/noul vocabulary table across Jev, Vercel's AI SDK, Laya and Open-Jev-9B.

Refreshed: ml-system-design — new section on the decision contract (minimal state, one ID per decision, catch-all option, policy outside the model) and the WorkflowResult/staged-rollout pattern. demo-to-production-gap — new miniature: shadow mode and assist mode as the two stages a demo has no reason to build. Spine: synthesis new paragraph under A category, not a product naming the SEO-content-farm angle as one hop further out than the vendor-PR incentive already tracked in growth edge 2; index updated under DefinedTerm and Article.

[2026-10-03] lint | quality cycle refresh of unsloth; edge 1c hunted

unsloth re-checked from the GitHub README (site behind a bot check): "70% less VRAM" kept beside the July ~80%; Desktop/Studio/Core forms and Intel/Vulkan support added. Edge 1c hunted via research pass, nothing at T1/T2.

[2026-10-04] ingest | llm-abliteration (NousResearch fork of jim-plus) — refusal-direction weight editing

Routed from the hub, runner-up: none strong (mechanistic interpretability applied as a weight edit; no attack-tool or policy angle). New source page llm-abliteration (SoftwareSourceCode, T1 for artifact facts; no scores published, no published: date stated). Links to the technique page abliteration, arditi-refusal-direction and heretic in llm-providers-wiki and to nous-research in agentic-tooling-wiki; nothing written there. Synthesis: new section "An interpretability result becomes a weight edit". Dedup: no existing page for this repo; the method was already paged cross-spoke.