Synthesis — Machine Learning

The evolving thesis. Spun out of the hub _inbox on 2026-07-30 from five sources parked under four different tags across two months, clustered on the human's instruction after the router flagged that 29 spokes existed and none of them owned machine learning.

Current thesis

The hub's ML strays scattered because a model's life is split across spokes, and the split had no centre. Serving mechanics had a home, the provider market had a home, agent frameworks had a home — but the model itself, how it is fitted, adapted and designed into a product, did not. Each stray was correctly declined by the spoke nearest it, and four separate tags accumulated as a result. That is what a missing spoke looks like from the inside.

The corpus that fell out of it covers three layers of one discipline:

• Algorithm — the fitting procedures, in code, from regression to LSTMs mlalgorithms.
• Adaptation — taking someone else's pretrained model and turning it to a task (fine-tuning, PEFT/LoRA), with one explainer for the concepts finetuning-explained-kdnuggets and one toolkit release for the costs unsloth-amd-support.
• System — what gets built around a model in production ml-system-design, as 300+ shipped case studies ml-system-design-case-studies and one headline about the ones that didn't ship ai-projects-fail-infrastructure-people.

A floor under the three layers

The sixth source is not a fourth layer — it is underneath all three. information-theory-inference-learning-algorithms arrives as the corpus's first textbook, first T1, and first source that explains why fitting works rather than how to run it: the loss a network minimizes is a log likelihood, its regularizer is a log prior, and what learning produces is an ensemble of parameter values rather than a point (bayesian-inference). Model complexity is priced by the evidence without anyone adding a penalty term. The same message passing that marginalizes a probabilistic model decodes an error-correcting code (information-theory).

It changes the corpus's shape in three ways.

• The evidence spectrum now has a top. Five sources of vendor posts, catalogs, explainers and one unread headline, and now a university-press textbook with derivations and worked solutions. The other five are about the present; this one is about the mathematics, which is why it can be twenty-three years old and still be the strongest thing here.
• It half-fills the middle gap, from the wrong end. The gap was training a model from data end to end — data, labelling, pretraining, evaluation. MacKay supplies the theory of fitting and model comparison and nothing at all of the practice: no data pipeline, no compute, no cost. The hole is now specifically empirical, not conceptual, and that is a sharper statement of it than before.
• It splits the corpus's centre of gravity. Every earlier source is engineering under time pressure. This one has no cost model and no deployment, and its neural-network chapters stop a decade before the scaling era — Hopfield nets and Boltzmann machines, no ReLU, no attention. The theory is durable, Part V is a historical document, and the spoke should not let the two be read as one.

The founding gap — nothing in the middle

The corpus jumps from how a fitting algorithm works to how to adapt a finished model, with no source on training a model from data end to end: no pretraining, no data collection or labelling, no evaluation methodology. That absence is itself informative about where practitioner attention has gone — adapting a pretrained model is now the common case, and building one from scratch is not — but it is a hole in the corpus before it is a finding about the field. First priority for the next ingest.

One third of it closed, 2026-08-03. timesfm-2-5-forecasting-tutorial supplies the evaluation methodology the gap named: a baseline that isn't zero (seasonal-naive, and MASE which bakes the comparison into the metric), rolling-origin backtesting across folds instead of one holdout, scoring the interval with pinball loss and coverage rather than only the point forecast, and ablating a configuration knob against both accuracy and cost (time-series-forecasting). None of that is specific to the model under test, which is why it counts as method rather than result. Pretraining and the data/labelling half of the gap are still unsourced.

A second third, 2026-08-04. transformers-tutorials supplies the training loop — not described, executable: fine-tuning on real datasets across several dozen models, written three ways (native PyTorch, PyTorch Lightning, the Trainer API) with preprocessing guidance and a stated basis for choosing among them. Until now fine-tuning was a page of concepts from an explainer with no measurements, and unsloth was a tool that hides the loop; this is the loop itself.

So the founding gap has closed from both ends and the remaining hole is specific: data collection and labelling, plus pretraining proper. Worth naming what the pattern of closure shows. The evaluation methodology arrived inside a tutorial for one model, and the training loop arrived inside a tutorial collection for many — both times as method carried incidentally by teaching material, never as a source about the pipeline itself. The middle layer appears to have no literature of its own in this corpus's reach; it is documented only where someone was demonstrating something else.

That also caps what either source settles. Neither reports a number. Knowing how to run a training loop and knowing whether the result is any good remain separate acquisitions, and the corpus keeps paying for that lesson one source at a time.

The last third closes 2026-08-08, and the reading above was wrong. data-cascades (Sambasivan et al., CHI 2021) and pervasive-label-errors (Northcutt et al., NeurIPS 2021) supply data collection and labelling — both peer-reviewed, both about the data layer itself rather than about a model that needed data. So the founding gap is closed except for pretraining proper.

The correction matters more than the closure. This page claimed 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." False. It has a literature; it sits in HCI and in the datasets-and-benchmarks track, studies practitioners and labels rather than architectures, and no amount of searching the ML side would have returned it. The absence was in where this wiki looked.

What the two say, briefly. Data Cascades interviewed 53 practitioners in high-stakes domains and found 92% had hit a compounding data failure, 45.3% more than one in one project — failures seeded upstream and surfacing late. Pervasive Label Errors measured the labels: at least 3.3% wrong on average across ten standard test sets, at least 6% of ImageNet validation, and correcting them can flip a ranking (ResNet-18 over ResNet-50, VGG-11 over VGG-19). Both are consolidated at training-data-quality.

The second result reaches back into this page's own method. The evidence spectrum below sorts sources by what they can establish and treats a neutral benchmark as the strong rung. A neutral benchmark on a 6%-mislabeled test set is weaker than that ordering implies — the ruler has a measured error rate. Growth edge 2 (no independent benchmark) is not retired but its bar moves: what would close it is a benchmark that states its label quality, and almost none do.

The evidence spectrum, stated up front

Three of the six carry a structural bias worth naming, in different directions. The case-study catalog is survivorship — every entry is a company describing a system that worked, so the corpus cannot say how often those same designs failed. The AMD post is first-party benchmarking — the vendor chose the hardware, the baseline and the configuration. The TimesFM tutorial adds a third and sharper form: synthetic data, where the author chose the problem. A vendor benchmark at least runs on a real workload; a forecast evaluated on a series built from a weekly cycle plus noise measures whether the code executes.

The pretrain-once move arrives in a second modality

timesfm is a 200M-parameter foundation model for numeric sequences, used zero-shot, in a field whose classical answer was to fit a fresh ARIMA or ETS per series. That is the language-model pattern — pretrain once, apply everywhere, adapt rarely — landing on a task that already had decades of per-series statistics to displace.

It sharpens this corpus's "don't fine-tune first" rule (fine-tuning) into a stronger claim: not adapt later, but don't adapt. And it makes forecasting the clearest place in the hub to watch whether the pattern generalizes past text, because the comparison is unusually clean — the classical baselines are cheap, well understood, and decades old.

Nothing here settled it for a month. The single source measured the model on data it invented, against one naive baseline, with no ARIMA, no exponential smoothing and no gradient-boosted tabular comparison (time-series-forecasting). The corpus held a method for answering the question and no answer.

Half of it is settled now, 2026-08-31, and the half that closed did not go the way this section expected. gift-eval (Salesforce AI Research, T1) runs 17 baselines over 23 real datasets and 97 dataset/horizon/frequency configurations. Against the classical per-series fits the zero-shot story holds: the foundation models beat Auto_ARIMA (CRPS 0.811 normalised to Seasonal Naive), and they beat it comfortably. Against a supervised transformer trained per dataset it fails — PatchTST takes the best score on every aggregate metric (CRPS 0.563, average rank 5.72), with Moirai-Large second at 0.593.

So "don't adapt" is a claim about what you can skip, not about what wins. The thing zero-shot displaces is the cheap per-series fit, not the expensive per-dataset one. And the split is legible rather than noisy: pretraining wins daily-to-yearly frequencies and univariate series, and loses secondly/minutely, multivariate inputs, and the two domains with the most entropy and least trend. Broad slow structure transfers; noise and cross-series structure do not.

Two caveats belong with the numbers. The foundation models were measured under acknowledged partial leakage — TimesFM, Chronos and Moirai all have pretraining overlap with the evaluation sets, which flatters them, and they still lose. And Moirai is Salesforce's own model, benchmarked by a team that includes its author; the result cuts against the direction that bias would run, which is the reason to trust it here.

The other half is untouched, and now in a more interesting way. GIFT-Eval has five statistical models, eight deep models and four foundation-model families, and no tree ensemble at all — no XGBoost, no LightGBM. The method tabular-deep-learning-benchmark says is still state of the art at ~10K samples, and the one a practitioner tries first, is absent from the field's leading general benchmark. Open question 7's boosted-tree half is not open because nobody got to it; it is open because the benchmark that should have run it did not.

The vendor answers the benchmark, and declines to publish the score (2026-09-01)

timesfm-3 landed the day after this spoke read gift-eval, and it is aimed at the exact corner the benchmark said pretraining did not buy. GIFT-Eval's split was that foundation models win broad, slow, single-stream structure and lose noise and cross-series structure; multivariate configurations went to PatchTST and iTransformer, and Moirai was the only foundation model that would accept multivariate input at all. TimesFM-3 is 330M parameters with variate attention pretrained in, covariates native, and the whole horizon decoded in one pass. The gap named in a benchmark in October 2024 has a purpose-built model against it by August 2026, which says the feedback loop between evaluation and release in this field is short.

What did not arrive is a number. The announcement claims the top rank on GIFT-Eval, FEV-Bench and a TIME-leaderboard, in charts with no scores, and only among pre-trained foundation models. That comparison class excludes the model that won the benchmark. So the sentence is true-shaped and unfalsifiable here, and open question 7 does not move: the per-dataset supervised comparison is untouched, and the boosted-tree half is as absent from this release as it was from the benchmark.

The corpus now holds the pattern twice in one subject. The 2.5 tutorial measured the model on data it invented; the 3.0 announcement measures it against a field it chose. Both are T3, both are usable for what the artifact is, neither is usable for how well it works — which is the general rule this wiki keeps rediscovering rather than a fact about forecasting.

One thing did change concretely, and not in the direction the release notes emphasise: the licence. The 2.5 checkpoint is Apache-2.0, the 3.0 checkpoint is timesfm-non-commercial-license-v1.0 (both cards checked 2026-09-01). Open weights, commercial use withdrawn, no mention in the post. The open-weight story in this corpus is usually told about text models in ../llm-providers-wiki; the same narrowing is now visible on a numeric-sequence model, and it happened at the version where the model got good enough to sell against.

A fourth layer arrives, and it is after the model

supervision does not fit the three layers, and that is the interesting thing about it. It is not an algorithm, not adaptation, and not system design — it is the code that runs once a vision model has produced its output: a common detection structure, annotators, multi-object tracking, dataset format conversion, and detection metrics. Nothing in it touches backprop.

It stretches this spoke's declared boundary (train-time and design-time) and was still routed here, because the alternatives fit worse: ../llm-inference-wiki is serve-time mechanism for LLMs specifically, and ../dev-tooling-wiki declines application-level runtime libraries. Recorded as a stretch rather than a precedent — if serve-time vision tooling keeps arriving, the honest response is to widen the boundary deliberately or admit that computer vision as a modality is unowned across the hub, which it currently is.

What it contributes is evidence about demo-to-production-gap, which until now had been a claim with no mechanism: a success catalog with no denominator ml-system-design-case-studies and an unreadable headline about failures ai-projects-fail-infrastructure-people. Here is the gap itemized, in the form of what a company had to write and 48.7k stars' worth of people needed. Format conversion between YOLO, COCO and Pascal VOC. Identity across frames. Drawing a box. Counting what crossed a line. Computing mAP.

That is a modest, one-sided piece of evidence — a library's existence proves demand, not that projects fail without it. But it is the first source in the corpus that says anything concrete about what the distance consists of, as opposed to how wide it is.

The model's caller changes (2026-09-14)

ml-models-as-agent-tools is a T3 explainer with no numbers, kept for one structural claim the design corner did not have: the code around a trained model in production is now, in some systems, an LLM agent — the classifier returns a probability and everything else (retrieval, thresholding, routing, the written justification) is the agent's. The ml-system-design-case-studies are systems built around a model; this is the model as one tool inside a loop it does not control. Its own example undercuts its pitch: the decision is still "score above threshold", a rule, with agent steps wrapped around it. Whether the wrapping helps or adds a new failure mode (a misread score, a stale feature) is exactly the demo-to-production-gap measurement this spoke keeps lacking. The agent machinery is ../agentic-tooling-wiki's; the model's changed position is recorded here.

Weights as a modality

skillsmith is the corpus's first primary research on adaptation (every prior adaptation source was an explainer or a toolkit) and it arrives with an unusual stance. It treats model weights as a thing a model reads and writes: prefix-tuning stores a skill as parameters, and an LLM composes new prefix weights from old ones plus a text description, reportedly beating both text-only and weight-space-only baselines.

This is a third position on the seam this synthesis opened between the textbook and the practice. information-theory-inference-learning-algorithms says learning produces a distribution over parameters; the fine-tuning/unsloth toolchain ships the single point estimate and calls it training. SkillSmith does neither — it takes trained weights as objects to manipulate, making the parameters themselves the unit of analysis rather than the loss or the data. It does not settle the disagreement, and as a preprint on the authors' own evaluation it is held loosely (T2, the domain's benchmark caveat in full force). But it widens the corpus's account of what "adapting a model" can mean: not only moving weights toward a task, but composing weights as reusable parts.

Note the standing homonym it sharpens: a parametric skill (prefix weights) is not an agent skill (a markdown prompt an agent loads) — the same week both senses arrived in the hub, in different spokes.

Training becomes a subject, and open question 6 gets a partial answer

Every source before this one treats training as something you run: a loop, a tool, a cost in VRAM. phase-transitions-in-attention treats it as something to explain — why a capability appears at the data size it appears at — and it is the corpus's first T1 primary research rather than a textbook, tutorial or vendor post.

What it does. Marginalise the weights of a single-head attention layer trained on a copy task, get a closed-form posterior over attention patterns, and reduce it by the task's permutation symmetry to two order parameters: attend uniformly, or attend one position back. The negative log-posterior is then an effective landscape whose minima say which pattern the data prefers, "separately from whether an optimizer reaches it."

Open question 6 — "does the Bayesian account survive at scale?" — moves, without closing. bayesian-inference had a textbook saying learning produces an ensemble and a toolchain shipping point estimates, neither citing the other. This paper puts them in the same equation in one tractable setting: the posterior's MAP is where Adam lands, because minimising the action at large χ corresponds to gradient flow with weight decay, and SGLD sampling agrees with Adam closely. The point estimate is the mode, not a different activity. What buys that result is a symmetry, not a scalable method — nobody integrated over billions of parameters — so the question survives in a sharper form: the correspondence holds where the posterior collapses to two numbers, and no one has shown what happens where it doesn't.

The finding that travels furthest is about monitoring, not training. The order of the transition is set by the attention activation: softmax gives a first-order jump with no precursor, linear attention gives a second-order transition preceded by diverging fluctuations and then a gradual crossover. If that generalises, capability onset in the architecture everyone actually ships is "intrinsically unpredictable from observable quantities," and there are "fundamental limits to what can be monitored" (capability-emergence). That is a claim about AI risk practice derived from statistical mechanics, and this spoke should hold it as a conjecture with a derivation behind it: one layer, one head, i.i.d. uniform tokens, a supervised copy task standing in for next-token training, and asymptotic approximations. Copying is the minimal circuit under in-context learning, not in-context learning.

It also sits oddly against the corpus's evidence habits. Everything else here is unreplicated first-party numbers; this is a derivation whose numerics match its own theory. Both are single-source, and neither kind of weakness cures the other.

Open questions

1. How often do AI projects actually fail, and by what definition? A success catalog with no denominator on one side, an unevidenced failure claim on the other. Nothing measures the rate.
2. When is fine-tuning the right tool? The corpus states the rule — try prompting and retrieval first — and gives no threshold, task taxonomy or cost comparison for applying it.
3. Do the Unsloth speedups survive independent measurement? The identical loss curve is the load-bearing claim, and it is self-reported.
4. What does the middle layer look like? Pretraining, data, evaluation — see the founding gap. Evaluation now has a method (time-series-forecasting); pretraining and the data/labelling half remain unsourced. Answered as a syllabus 2026-08-18 by lindholm-machine-learning-first-course, which is the middle layer taught end to end: k-NN and trees before any parametric model, then regression, regularisation, bias–variance, parametric learning and SGD, neural networks with backpropagation, bagging/boosting, kernels and SVMs, Gaussian processes, mixtures and clustering. Its Chapter 11 also supplies a diagnostic the corpus lacked — carve a training-validation set out of training data and the hold-out error splits into training error, generalisation gap and train-val mismatch, so a practitioner can tell a model problem from a distribution problem. What the book does not touch is what sits above it: adapting pretrained models, and the data-labelling practice of edge 1a. So the question narrows from "what is the middle layer" to "what is between this book and a fine-tuning run". The labelling half answered 2026-08-25 by safetutors and annotation-requirements-autonomous-driving — one annotation operation published in full, one reported by its practitioners. See edge 1a. Pretraining is still the unsourced part.
5. Does the ranking/matching/forecasting concentration reflect the field, or only what companies blog about? The catalog cannot distinguish these.
6. Does the Bayesian account survive at scale? bayesian-inference says the product of learning is an ensemble; nobody samples a posterior over billions of parameters. Whether the point estimate everyone actually ships is an approximation with known cost or a different activity wearing the same vocabulary, no source here answers. Partly answered 2026-08-07 in one tractable case (phase-transitions-in-attention): the MAP of an analytically-derived posterior is where Adam lands, so there the point estimate is the mode. The reduction that made it computable was a permutation symmetry, so the open half is now specific — what holds when the posterior does not collapse to two order parameters.
7. Is capability onset detectable before it happens? (added 2026-08-07) One derivation says the answer depends on the activation function, and that softmax — the one in production — gives no precursor (capability-emergence). Nothing here tests that on a real training run, and the claim would matter a great deal if it held.
8. Do time-series foundation models beat per-series classical models? (added 2026-08-03) Half answered 2026-08-31 by gift-eval: yes against Auto_ARIMA and the naive baselines on 23 real datasets, and no against PatchTST, a supervised transformer trained per dataset, which wins the benchmark outright. The remaining half is the gradient-boosted comparison, which the leading benchmark in the field does not run. The question that replaces it is narrower and better: is per-dataset supervised training worth its cost, given that it wins? — the zero-shot argument was never only about accuracy. Bears directly on question 2. Not moved by timesfm-3 (2026-09-01), despite it being the obvious candidate: Google ranks it only against other foundation models, and publishes charts rather than scores. A release aimed at the benchmark's multivariate finding still leaves the comparison the question asks about unrun. A third register 2026-09-17: younes-introduction-to-machine-learning (JHU, 674 pp, T1) is the graduate mathematical treatment — the optimization theory Lindholm assumes, ~140 pages of graphical models, and a closing chapter of generalization bounds that says why a learner generalizes. It has no exercises and gives neural nets 18 pages. Pretraining, data and evaluation as practice are still unsourced by any of the three books.

The tabular default finally has sources

The corpus argued about foundation models for a month with no page on the method a practitioner would actually use first. gradient-boosted-trees fixes that, and the two sources behind it say different things worth keeping apart.

xgboost is a story about systems, not statistics. Friedman's algorithm was a decade old; what changed in 2016 was an implementation with sparsity-aware splits, a weighted quantile sketch and out-of-core blocks, hitting scikit-learn's accuracy on Higgs-1M at roughly forty times the speed. The method became the default because it became cheap to try. That is the same mechanism this spoke keeps finding elsewhere — capability follows tooling — and it is worth noting that the dominance evidence (17 of 29 Kaggle winners) is competition results, a venue with its own selection effects.

tabular-deep-learning-benchmark is the durable half. Six years later, with 45 datasets and ~20,000 tuning hours per learner, tree ensembles are still ahead at ~10K samples, and the explanation is structural: uninformative columns, rotation invariance, irregular targets. Those are properties of messy real data, which lines this up with data-cascades — the deciding variable keeps turning out to be the data, and the method that tolerates bad columns keeps winning.

Open question 7 is now precisely stated and still unanswered. The baseline it needs exists here; the comparison against a time-series foundation model does not.

Tensions

Still no head-on contradiction — the sources don't overlap enough to disagree, which remains a weakness in a six-source corpus.

The first real seam has opened, though, and it runs between the textbook and the practice. information-theory-inference-learning-algorithms frames learning as producing a distribution over parameters and treats the single most probable setting as a fallback for when the integral is intractable. Every other source in this corpus ships that fallback and calls it training — fine-tuning and unsloth are entirely point-estimate machinery, and the vocabulary of priors and evidence appears nowhere in them. Neither side argues with the other because neither cites the other. Filed as open question 6 rather than a contradiction: a twenty-three-year-old textbook and a 2026 toolchain disagreeing about what learning produces is a gap in the corpus's coverage before it is a dispute in the field.

The nearest thing to a live tension is between ml-system-design-case-studies and ai-projects-fail-infrastructure-people: one says ML systems ship routinely across dozens of industries, the other that most AI projects fail. Both can be true (different denominators, different eras — the catalog stops at 2023), and neither source cites the other. Recorded as an unresolved question rather than a conflict, because the failure side is a headline and not yet an argument.

A boundary the hub drew and this spoke inherits: the parked ai-test-automation-paradox record argues that AI generating tests at scale mostly scales brittleness — the same demo-to-production shape as this corpus, applied to software testing rather than ML systems. It was deliberately left in _inbox rather than folded in here. If a third source lands on that seam, the boundary should be revisited rather than defended.

Agreement and correctness come apart, and the gap flatters the cheap option (research pass, 2026-08-28)

This wiki adopted inter-annotator agreement as the instrument for label quality when measuring-annotator-agreement arrived, with the caveat that it is well-founded for categorical labels and shaky for everything modern datasets actually use. absa-annotator-comparison-cost adds a sharper caveat, because it measures both ends of the same task: the LLM annotations achieved high agreement and produced models only comparable to student-annotated data, while expert annotations remained the best basis for peak performance.

An agreement score asks whether annotators concur. It cannot ask whether they were right, and an LLM annotating consistently is the case where those two questions diverge most cleanly — a model agrees with itself by construction.

The mechanism is coverage, not mislabelling, which is why agreement misses it. On the harder subtask, student and crowdworker datasets contain systematically fewer instances than expert and LLM ones, concentrated in the rare categories (AMBIENCE, PRICE). The instances are not wrong; they are absent, and they are absent from every annotator's copy, so no pairwise measure computed over what was annotated can see the hole. That is the same shape as pervasive-label-errors one level up: the error is in the test set's construction rather than in any single label.

Filed beside edge 1a's existing flag rather than replacing it: the corpus now holds an instrument, a report that the instrument is rarely published (who-annotates-in-nlp), and a case where the instrument scores the cheap option highest.

Supervision can be manufactured, and the error moves rather than disappears (2026-08-31)

code-as-world-report enters the corpus's most-worked question — where labels come from — with an answer none of the six sources at training-data-quality considers: don't annotate anything.

The setup makes the move legible. The task is reading an object's real speed, size or acceleration off a monocular video, and it is a task nobody can annotate: a crowdworker cannot see metres per second, and an expert would have had to measure the scene while it was filmed. quantiphy gets 569 videos this way, by curation, which is a benchmark and not a training set. So Code-as-World inverts the pipeline — it uses an agentic-discovery-loop to recover an executable-world-representation of a scene, keeps the ones that survive verification, and reads the labels off the simulator, exactly and for free.

Every cost this corpus has spent two months measuring goes to zero. No recruitment, no 140-page guidelines, no adjudication, no £828, no κ — there are no annotators to agree with each other. That is the strongest form of the "data is the bottleneck" thesis being answered, and it is worth naming as a third strategy beside collect-and-label and buy-a-pretrained-model.

But the failure mode does not vanish; it relocates upstream, and out of instrument range. The report's own §4.4 concedes that the loop "may converge to a locally plausible EWR without recovering a mechanistically accurate explanation." pervasive-label-errors found wrong labels on real data. Here the labels are exact by construction and the world they describe may be the wrong one — and no distance function, no agreement measure and no adjudication protocol can detect a label that is perfectly consistent with a mis-specified simulation. The only check on offer is that the world renders like the video, which tests appearance rather than mechanism.

So the two strategies partition by domain rather than compete. Manufacture where a simulator is faithful and a human cannot see the quantity; annotate where the reverse holds. The report's own limits mark the boundary: no camera motion, rotation, deformation, occlusion, fluids or long-horizon multi-object dynamics — that is most of the world, and it is where the annotators stay.

It also gives the RL half of fine-tuning its first reason. GRPO had appeared here only as a timed step in a vendor post. The two-phase recipe behind code-as-world-vl — SFT where cheap derived labels are abundant, RL where labels are few and exact — is the corpus's first stated basis for choosing a post-training stage, from one first-party report and with no ablation between the phases.

And it is the corpus's first source where the agent is a factory rather than a product. The hub's boundary with ../agentic-tooling-wiki holds cleanly and for a reason worth recording: this loop runs offline, produces no user-facing behaviour, and exists entirely to emit training data. Its output is a dataset; the harness that made it is somebody else's subject.

The cross-wiki finding: the loop is scaffolding, not a capability

llms-cant-jump (../research-wiki) argues that abduction — generating a novel explanatory hypothesis for a singular observation — is structurally missing from these models, and names manipulative abduction, hypothesis by embodied simulation, as the human mechanism. Code-as-World builds exactly that, in software: propose a world, run the physics, look at the result, diagnose, revise. The report reaches for the same lineage unprompted, calling world recovery an inverse problem solved by abductive reasoning and citing the heliocentric and Newtonian precedents.

The two do not contradict each other. They agree, and the agreement is the finding — code-as-world-report §5.5 states that the trained model "does not internalize the agentic discovery process itself… hypothesis construction, simulation, diagnosis, and iterative revision remain external to the model." The model learns the outputs of abduction as supervised targets. The jump stays in the harness. That is Zahavy's claim confirmed by a lab trying to route around it, and it is more evidence than either page had alone.

The two label strategies turn out to be consecutive (2026-09-07)

The 2026-08-31 section above set manufacture and annotation against each other and partitioned them by domain: manufacture where a simulator is faithful and a human cannot see the quantity, annotate where the reverse holds. male-fly-connectome shows a mature project running both, in order, and the order is the finding.

Google Research and HHMI Janelia released the male Drosophila connectome — 166,000 neurons, 125 million synapses, published in Cell with 111 authors. The segmentation models are trained with synthetic neurons injected into the training data, which the post credits with improving speed and accuracy. Then human experts proofread the model's output for years, and the post names that as the constraint on the whole field: "reducing this need for manual error correction" is what would let bigger brains be mapped "within reasonable budgets and timelines."

So the strategies are not alternatives at this scale. Manufacturing gets you a model; the model gets you a draft; the draft still has to be corrected by people who know what a neuron looks like. And the human step is not the labelling activity this corpus has been measuring — it is label-error correction on a model's output, planned for and budgeted in advance rather than discovered afterwards in a finished test set.

One release history makes the cost legible. The same lab published a fully automated female fly brain in 2019, a human-verified half brain in 2020 (25,000 neurons, 21 million connections), and the complete male CNS in 2026. Automation arrived first and six years went into verification. That is the clearest statement the corpus holds of what the annotation layer actually costs when nobody can skip it.

The publishing pattern repeats, and it is the venue's. Like timesfm-3 a week earlier, this is a Google Research post asserting an improvement — "speed and accuracy" — with no number attached: no error rate, no proofreading hours, no split between automated and corrected output. Two posts, two subjects, same omission. The peer-reviewed paper is where those numbers would live, and this spoke has verified its existence rather than read it.

The dataset ships with a viewer, and the viewer ships a format. neuroglancer arrived hours after the post and is the delivery end of the same pipeline: Apache-2.0, purely client-side, so publishing a connectome means publishing a URL. Its own precomputed format has accumulated the readers and meshing pipelines other labs built around it. Two things are worth keeping. The README states plainly that it is "not an official Google product" — a resource several published connectomes are distributed through, carrying no support commitment. And a dataset anyone can open in a browser is the cheapest route to a second pair of eyes on work that cost years of expert proofreading; whether that produces corrections is unmeasured, and this spoke has no source that asks.

A scope note, recorded rather than settled. Connectomics is neuroscience and no spoke owns it; the pipeline is what routed here. ../psychology-wiki holds behavioural neuroscience but not neuroanatomy. If two more sources of this kind arrive, reopen the question instead of settling it by precedent from this ingest — the same rule that spoke wrote for itself when the polarization model landed there.

Adaptation by vocabulary (2026-09-16)

lfm2-5-230m-chess is a small, legible case of a fine-tuning move this spoke had not paged: instead of teaching a language model to write chess notation, give it 2,106 new tokens, one per legal move, mask sampling to the legal set, and distil Stockfish into a value token followed by a policy token. The base model supplies sequence modelling only; the task lives in the vocabulary and the decoding constraint. Set beside the PEFT pages, it is the opposite bet — change the interface, not the weights' rank — and it comes with a metric that is cheap to check independently (play it). The 2004 Elo with search and 1500 without are the author's numbers; nobody has measured them yet.

Three textbooks, three registers (2026-09-17)

With younes-introduction-to-machine-learning the spoke holds machine learning taught three ways: for engineers with problems to solve (lindholm-machine-learning-first-course), through information theory (information-theory-inference-learning-algorithms), and as graduate mathematics. The page counts are the finding. Younes gives optimization 60 pages and graphical models about 140, and neural networks 18; the words "transformer" and "attention" occur but are not a subject. That is not neglect — the preface says the book is about the mathematics that lets a reader assess and extend the algorithms in the packages — but it means the corpus's most rigorous text is the one furthest from what the rest of this spoke's sources spend their time on. The middle-layer question (open question 4) is answered as theory now as well as syllabus, and still not as practice: none of the three books teaches pretraining, data curation or evaluation.

A category, not a product: three System 1s in one week (2026-09-24)

laya (Convai Innovations) lands in the same architectural slot llm-providers-wiki already tracks: system-one-models-and-jev describes TypeSafe's Jev, a non-autoregressive "System 1" model trained with something it calls RLCD to answer typed questions with calibrated probabilities. Laya is the same shape — bidirectional encoder plus a small decision head, choice/score/noul questions, RLCD training against strictly proper scoring rules — arriving from an unrelated vendor within weeks, and it benchmarks itself directly against Jev by name. It claims to win on accuracy (0.766 vs. 0.727) and calibration (ECE 0.081 vs. 0.246) and to lose on high-cardinality choice questions (Banking77: 0.425 vs. 0.870), all numbers Laya measured itself for both models.

Two things stay separate here. First, nobody replicated anything: these are one vendor's numbers about a competitor, a step up from a vendor's numbers about itself, but still not what growth edge 2 asks for. Second, the difference that doesn't depend on either benchmark table: Jev is closed and metered, Laya is Apache-2.0 and free to self-host, so a reader who wants the architecture rather than the vendor relationship has an option that didn't exist three weeks ago. fine-tuning gets the technical link — Laya's RL stage is a second, independent instance of "GRPO teaches calibration," the same mechanism code-as-world-vl uses in an unrelated domain.

A third, same day: open-jev-9b (Zefan Cai), and it changes the shape of the finding. Where Laya matches Jev's vocabulary and training-method name without comment, Open-Jev's own README says outright what neither Jev's nor Laya's card does: "an independent implementation inspired by TypeSafe's Jev. It does not reproduce proprietary RLCD, private weights or training data." Its architecture is further from both — a LoRA adapter over a 9B generative chat model (Qwen3.5-9B) rather than a bidirectional encoder built for the job, with a decision head initialised from the base model's own yes/no logit rather than trained from scratch. Its card never names RLCD or REINFORCE; the project's documented 27B sibling uses a supervised loss (cross-entropy plus a Brier-score term) instead. So the category this spoke is now tracking is a target shape (typed questions, no generated text, a calibrated probability out) reachable by at least two different training methods — RL against a proper scoring rule, and a supervised loss shaped like one — rather than a single technique with three implementations. Open-Jev-9B's 97.54%/91.97% accuracy figures are not placed beside Laya's or Jev's in this synthesis because they aren't measuring the same thing: its own held-out split, drawn from its own synthetic generator, is not a shared benchmark with either.

Cross-spoke, unintegrated into that spoke: agentic-tooling-wiki's model-routing pages (cache-aware-model-routing, claude-code-model-router) want exactly this primitive — a fast, cheap, calibrated yes/no or choice signal to decide whether a request needs a bigger model. Noted here as a connection; nothing changes in that spoke's own pages.

A fourth artifact, three days later, and it isn't a model (2026-09-27). jev-typesafe-typed-decisions-guide is a how-to for calling Jev in production, from a Hugging Face account whose own profile is a content mill — 19 articles keyed to trending product names, two more about Laya and Jev published within a day of this one. That is worth naming plainly: the category this section tracks has existed for under two weeks and has already attracted SEO content built to rank on its vocabulary, not just three vendors building the thing. Growth edge 2 already holds lindholm-machine-learning-first-course's observation that research doubles as advertising for the lab that publishes it; this is the same incentive one hop further out — a third party with no stake in any of the three models profiting from the category's search traffic. The post's own technical content checks out against jev-ai-sdk-guide on every point tested, so it isn't wrong, only unaccountable to anything but its own claims. Named as typed-decision and folded into ml-system-design and demo-to-production-gap for the parts worth keeping — a decision contract, an explicit "unavailable" result state, and a staged rollout — none of which depend on which vendor's model sits behind the call.

An interpretability result becomes a weight edit (2026-10-04)

llm-abliteration is the spoke's first source where the model-science corner (see induction-head, phase-transitions-in-attention) is used as a tool rather than described. arditi-refusal-direction found that refusal is mediated by a single direction; the toolkit measures that direction by contrast (harmful minus harmless prompts) and orthogonalizes the weights against it, with no training. The same move as isolating a circuit by ablating it, run in the other direction: once you can name the direction, you can delete it.

Two things to keep apart. First, the repo shows the mechanism works well enough to be productised (three switches, a YAML layer plan, six model families listed) but publishes no before/after refusal rate or capability score, so "tested" means "runs". Second, its own caveat narrows the science: removal holds only for refusal types present in the prompt sets, which suggests the single direction is a property of measured refusals, not of everything a model can decline. Implementations also read activations at different residual points, so directions are not interchangeable across tools. The technique and its refinements (norm preservation, biprojection) are tracked in llm-providers-wiki's abliteration; nothing is changed in that spoke here.

Growth edges

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

1. The missing middle layer — partly closed 2026-08-04 by transformers-tutorials, which supplies the training loop as working code across three frameworks and several dozen models. The method half is answered. Data curation and evaluation are not — the repo teaches how to run the pipeline and reports nothing about whether the result is good. — needs: T1/T2 on data curation and on evaluation, which have separated out as edge 3 and are the live half of this one. Touched, not closed, 2026-08-05: supervision ships a dataset module (YOLO/COCO/Pascal VOC conversion) and a metrics module (mAP), so the corpus now holds curation and evaluation as installable tooling. It still holds neither as a result — nobody has reported what good looks like on a task. A library is where the answer would be computed, not the answer. CLOSED 2026-08-08 (research pass) by data-cascades and pervasive-label-errors, both T1, both about the data layer as a subject: a 92% cascade prevalence across 53 practitioners, and at least 3.3% wrong labels across ten standard test sets. Successor written as edge 1a. 1a. How the data work is actually done well. New 2026-08-08. Both closing sources are diagnoses: one measures how often the practice fails, the other how wrong the labels are, and neither says what competent labelling looks like. Partly closed the same day by measuring-annotator-agreement (Braylan, Alonso & Lease, T1), which supplies the instrument — and reports that inter-annotator agreement is well-founded for categorical labels and shaky for the bounding boxes, sequence tags, ranked lists and free text that modern datasets actually use. Still missing the practice: annotation guidelines, annotator training, sampling design, cost. Hunted again 2026-08-10 (research pass) — and the hunt returned the reason, not the artifact. who-annotates-in-nlp (Kunilovskaya et al., T1) audits 2,667 annotation tasks from 1,603 ACL-venue papers and finds the practice mostly unpublished: annotator training reported for 18.7% of tasks, released guidelines 34.1%, an agreement value 35% — against recruitment 90.4% and total items 86.0%. Papers say who was hired and how much was labelled, and omit what would let a reader judge whether the labels mean anything. So this edge was never a search failure; the base rate is low. Flagged, because it undercuts the instrument above. measuring-annotator-agreement analyses a statistic that a third of tasks report a value for, under conditions — who annotated, trained how, told what — that are documented less often still. A kappa without its guidelines describes an unspecified procedure. Re-specified rather than closed. — needs: a single labelling operation documented end to end (guidelines, training, sampling, cost) — a dataset paper's full annotation appendix, an industrial data-operations write-up, or a datasheet that actually fills the fields. A second meta-study of reporting would not close it. CLOSED 2026-08-25 (research pass), by both named artifacts at once. safetutors (Hazra et al., TU Eindhoven / IIT Kharagpur / Cisco Research / NYU, arXiv 2603.17373, T1) publishes the appendix: three annotation stages with three workforces and three stated pay rates (6 undergraduates $5/hr, 3 doctoral students $15/hr, 24 Prolific workers $10/hr), a four-step calibration ending at an ≥80% agreement-with-gold gate, a 10-item qualification test whose pass rates are printed (24 of 33 crowd recruits, 72.7%), verbatim guidelines per stage, the tools named, an adjudication protocol with rates (4.2% of crowd instances escalated to a senior reviewer; 3.8% cross-stage conflicts, 21 dropped and 36 flagged), Fleiss' κ split by stage, subject and turn count, and 15,300 labels totalled. annotation-requirements-autonomous-driving (Saeeda, Mohamad, Knauss, Horkoff & Nouri — chalmers-university-of-technology, RISE, Gothenburg, Volvo Cars, arXiv 2511.15859, T1) supplies the industrial side from 19 interviews at six automotive companies: guidelines past 140 pages, manual review cut 10% → 5% under budget pressure, rework at 3× the cost of labelling properly, only ~20% of collected data ever annotated, 15% of one 3D point-cloud set destroyed by version conflicts. Successor (edge 1b), new 2026-08-25: neither source states what the work cost. SafeTutors gives hourly rates and per-item times but no total; the interview study gives a ratio (3× for rework) with no base. — needs: a total cost of ownership for one labelling operation — spend, hours, or cost per usable label, with the scope it covers. T1/T2. CLOSED 2026-08-28 (research pass) by absa-annotator-comparison-cost (Donhauser et al., Regensburg, arXiv 2605.03624, T1, read from the PDF): £828 total, £0.41 per three-way annotated example including platform fees, thirty Prolific crowdworkers at £9/hour, two days per study — a total, a per-usable-label figure, and its scope. What it adds beyond the number, and this is the part worth keeping. The paper prices four annotator types for the same task, and two of them have no price: students were paid in course credit and took weeks, and expert refinement took "several hours" only because it revised existing labels. Meanwhile "LLM annotation" spans $0.0025 self-hosted to $0.36 on a frontier API — a 144× spread that straddles the human price, so the cheaper-than-humans claim is unfalsifiable until the tier is named. Successor (edge 1c): every figure here is for 1,000 sentences of restaurant reviews with a short guideline. The industrial case in annotation-requirements-autonomous-driving runs 140-page guidelines over 3D point clouds, and nothing lets £0.41 travel to it. — needs: a cost per usable label from an operation with a guideline of that size, or a study of how per-item cost scales with task specification. T1/T2. Not eligible before 2026-09-11. Hunted 2026-10-03 — nothing ≥ bar. The closest find is an industrial case study from Hyundai (Kim, Ju, Feldt & Yoo, ESEC/FSE 2020, arXiv 2006.00894, T1). It says outright that the company's internal labelling cost "will remain confidential" and prices labels from Google Cloud's public list instead, which this edge excludes. That is the edge's own mechanism stated by an industrial author: the operation that could answer will not publish the number. Not eligible before 2026-10-17. One thing the pair also explains: the academic operation could publish because it was small, bounded and finished; the industrial one has unstable guidelines and, by its own practitioners' account, missing traceability, so there is no settled artifact to attach to a paper. That is a mechanism for the 18.7%/34.1% reporting rates above, not merely a restatement of them.
2. No independent benchmark, and the bar just moved. Every speed/VRAM figure is a vendor first-party snapshot — but pervasive-label-errors shows a neutral benchmark inherits its test set's label error rate, at least 6% on ImageNet validation, enough to reverse model rankings. — needs: a T2 neutral benchmark that states its label quality, not merely a third-party one. Explained, not closed, 2026-08-18. lindholm-machine-learning-first-course §12.2 teaches the incentive as a fact about the field: much ML research is done by large companies whose PR departments publicise it, so "research is (in part) a form of advertising," and "the scope of the results can potentially be exaggerated for commercial gain." A T1 textbook saying plainly what this spoke had been inferring one vendor post at a time. It is a statement about publication incentives, not a benchmark, so the edge keeps its ask — but it stops the corpus treating the pattern as its own observation. A second data point, 2026-09-24: laya's benchmark table against Jev (system-one-models-and-jev) is vendor-vs-competitor rather than vendor-vs-itself, a step up in rigor, but it is still Convai measuring both models and reporting the result on its own card — not the neutral T2 comparison the edge asks for. A third, same day: open-jev-9b doesn't even attempt a head-to-head — its 97.54%/91.97% figures are scored on its own held-out synthetic split, which the edge's neutral-benchmark ask rules out by construction. Three sources into this pattern and the corpus still has zero examples of a benchmark none of the parties involved built.
3. Evaluation as a discipline — the demo→production gap page implies it and nothing covers it. — needs: T1/T2 on model evaluation and monitoring.

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

4. Gradient-boosted trees. CLOSED 2026-08-09 (research pass) — xgboost (Chen & Guestrin, KDD 2016) and tabular-deep-learning-benchmark (Grinsztajn et al., NeurIPS D&B 2022), both T1, with gradient-boosted-trees as the concept page. Successor: run the comparison question 7 actually needs. The baseline now exists on paper and nobody has pointed it at a time-series foundation model. — needs: an M-competition-style evaluation, or any published result putting a foundation model against boosted trees and ARIMA/ETS on real series. HALF CLOSED 2026-08-31 (research pass) by gift-eval (T1): 23 real datasets, 97 configurations, and the ARIMA/ETS half is answered — foundation models beat Auto_ARIMA, and a per-dataset supervised transformer beats them in turn. The boosted-tree half is untouched, and the reason is now the finding: GIFT-Eval's 17 baselines include no tree ensemble of any kind. Residue as edge 4a below. 4a. The boosted-tree comparison, which nobody runs. (new 2026-08-31, residue of edge 4) gift-eval is the field's general forecasting benchmark and its 17 baselines contain no XGBoost, no LightGBM, no tree ensemble at all — while tabular-deep-learning-benchmark says trees still lead on messy medium-sized data. Somebody has to put the two on the same series. — needs: an M-competition result, a Kaggle-style forecasting write-up with both, or any paper running boosted trees beside a time-series foundation model on real data.

5. Model selection and validation. Cross-validation, leakage and the split protocol are what pervasive-label-errors and every benchmark claim here silently depend on. — needs: a textbook chapter or a paper on leakage in published work.

6. The machinery under demo-to-production-gap. Feature stores, drift monitoring and retraining triggers are named by implication and never described. — needs: a system report from a team running one, or a tool's architecture docs.

7. Interpretability. SHAP, LIME and partial-dependence plots are how deployed models get explained to the people affected by them, and no page covers any of it. — needs: the SHAP paper (T1) plus one critique of its limits.

8. New 2026-08-14 — representation learning, opened rather than closed. paraphrastic-similarity-parallel-translations is the corpus's first source where training a model is the subject rather than adapting one (fine-tuning) or fitting tabular data (gradient-boosted-trees): a contrastive objective over sentence pairs, with hard-negative mining as the technique that makes the loss actually bite. It also feeds edge 2 from an angle nothing else here has: its own baseline check found an untrained random encoder beating published models on monolingual STS by exploiting lexical overlap the metric doesn't control for — the same "an unchecked baseline can win for the wrong reason" caution tabular-deep-learning-benchmark and pervasive-label-errors raise, but caught by the paper's own authors rather than by an outside audit. — needs: a second representation-learning source (vision or audio embeddings would widen the pattern past text) to say whether the random-baseline trap and the pooling-beats-depth result generalize past sentences.

9. New 2026-08-31 — independent reproduction of manufactured supervision. The manufacture-don't-annotate strategy above rests on one first-party report with no peer review, no venue, no named authors and no released discovery agent — the loop that produces the supervision is described, not shipped (code-as-world releases inference and evaluation only). The one number this spoke trusts is the internal 31.2 → 50.6 delta against the model's own base. — needs: a T1/T2 result from someone other than mirros on simulator-derived training data, or any published attempt to reproduce the QuantiPhy gain. Failing that, a source on synthetic-data generation as a training practice in general, which would let the corpus say whether the right-render/wrong-mechanism failure is specific to physics or generic to the method.

Index — Machine Learning Wiki

Catalog of every page, grouped by schema.org @type. Spine: synthesis (thesis), log.md (history), this file (catalog). Spun out of the hub _inbox on 2026-07-30 (5 sources, four tags). Most speed and memory numbers in this domain are first-party and unreplicated — record the setup beside the number, or say the source gave none.

DefinedTerm (concepts)

• training-data-quality — the founding gap's last third: how training and test data get collected, labelled and checked. Two T1 halves — a 92% cascade prevalence across practitioners, and ≥3.3% wrong labels across the standard test sets · practice-area

• machine-learning — the root page: fitting a model to data, treated as an engineering discipline; the three layers this corpus covers and the one it skips · root

• classical-ml-algorithms — the pre-deep-learning core (regression, SVM, trees, boosting, k-means, PCA); currently a pointer to an implementation rather than a conceptual account · algorithm-family

• fine-tuning — adapting a pretrained model: full vs PEFT (LoRA/QLoRA), catastrophic forgetting, the "don't fine-tune first" rule, and the BF16 merge/de-merge drift gotcha · technique

• ml-system-design — designing the system around the model (target, features, serving, retraining); what actually ships is ranking, matching and forecasting over behavioural data · practice

• demo-to-production-gap — the distance between a demo that works once and a system that runs; the corpus's biggest hole — a success catalog with no denominator · problem

• bayesian-inference — learning read as inference: the loss is a log likelihood, the regularizer a log prior, and the product an ensemble rather than a point; Occam's razor comes free with the evidence · framework

• information-theory — entropy, compression and channel capacity, and the claim that they are the same subject as learning; coding theory proper is unowned across the hub · foundation

• time-series-forecasting — predicting a sequence from its own past, and the corpus's first full evaluation design: seasonal-naive + MASE, rolling-origin backtesting, pinball loss and interval coverage, context ablation; plus forecasting reused as anomaly detection · task

• monte-carlo-methods — sampling instead of integrating, because the posterior is intractable; the ML-side half of the standing homonym with metaheuristic search · technique

• capability-emergence — whether an ability arrives abruptly or gradually, and whether the abruptness is real or a metric artifact; the transition's order (first vs second) decides whether anything is visible before it, which makes it a monitoring question · phenomenon

• induction-head — the … A, B … A → B circuit behind in-context token statistics, and its first-layer copy sub-circuit, the part that is hard to learn; held here as how it forms in training, not as a serving concern · mechanism

• connectomics — cellular-scale wiring diagrams of nervous systems, held as an ML task: EM sections segmented by flood-filling networks, then proofread by hand. 166k neurons for a fly against 86 billion for a human, and one lab's 2019 → 2020 → 2026 releases as a record of what the verification step costs · task

• executable-world-representation — a physical scene written as code a simulator can run: composition / dynamic evolution / visual appearance, kept separate so the mechanism survives a re-render. Executable is testable and editable; it is not the same as true · representation

• agentic-discovery-loop — propose → instantiate → execute → render → verify, run by an LLM against a physics engine until a world explains the evidence or the budget rejects it; beats Best-of-5 sampling at a matched budget. An offline data factory, not a product · method

• gradient-boosted-trees — sequential ensembles of shallow trees, each fitted to the last one's residual; the default for tabular data, and the missing baseline in open question 7. Wins on inductive bias (ignores uninformative columns, orientation-preserving, handles irregular targets), not on maturity · method

• paraphrastic-sentence-embeddings — sentence encoders trained so same-meaning sentences land close in vector space, from paraphrase pairs or, more cheaply, directly from bitext; the corpus's first source on representation learning as its own training-time subject · technique

• typed-decision — a model call that returns a bounded judgment (declared answer space, matching response shape, a defined role in control flow) rather than free text; the vocabulary — choice/score/noul — that Jev, laya and open-jev-9b converged on independently, plus the decision contract for calling one safely · pattern

SoftwareApplication

• lfm2-5-230m-chess — maxime-labonne / liquid-ai, via Telegram: LFM2.5-230M-Base turned into a chess engine by adding 2,106 move tokens, an 80-token board prompt, a value-then-policy double pass with legal-move masking, and Stockfish distillation; ~1500 Elo one-pass, ~2004 Elo with depth-3 search (author's figures, no opponent pool stated) · source · T3 · huggingface.co
• code-as-world-vl — mirros's vision-language models for reading physical quantities off video; 4B/9B released Apache-2.0 on Qwen3.5 bases (27B trained, not released). SFT on pixel measurements then GRPO on simulator-derived labels; 31.2 → 50.6 on quantiphy against its own base · model
• timesfm — Google Research's foundation model family for time-series forecasting; zero-shot, quantiles per step, covariates via XReg at 2.5 and pretrained variate attention at 3.0 — the limit case of "don't fine-tune first," in a field that used to fit a model per series · model
• neuroglancer — github.com/google/neuroglancer: "WebGL-based viewer for volumetric data" — Apache-2.0, TypeScript, 1,364★/381 forks, created 2016, and "not an official Google product" by its own README. Purely client-side (a dataset is published as a URL), UI thread plus a WebWorker backend because browsers have no shared memory; WebGL 2.0 + EXT_color_buffer_float. Reads N5, Zarr v2/v3, OME-Zarr, DVID, NIfTI and its own precomputed format, which grew the ecosystem around it (CloudVolume, BigArrays.jl, Igneous, TensorStore). The delivery end of connectomics · source · T1 · github.com/google
• unsloth — fine-tuning framework (Triton kernels, ~80% VRAM reduction claimed); AMD/ROCm support since 2026-07; refuses LoRA merge/de-merge on BF16 drift grounds · tool
• laya — convai-innovations's non-autoregressive "System 1" decision model: ModernBERT-large/mmBERT-base with a decision head, typed choice/score/noul questions answered with calibrated probabilities in one forward pass, trained with RLCD against proper scoring rules. Benchmarks itself directly against Jev — same acronym, same question vocabulary, unrelated vendor — claiming higher accuracy and 3× better calibration, all self-measured; open Apache-2.0 weights are the one undisputed advantage · source · T3 · huggingface.co
• open-jev-9b — zefan-cai's LoRA adapter (rank 8/alpha 16) plus a Yes-minus-No-initialised scalar head on Qwen3.5-9B: a third choice/score/noul implementation, and the only one of the three whose README states outright that it's "an independent implementation inspired by TypeSafe's Jev" rather than staying silent on the relationship. 97.54%/91.97% test/OOD accuracy on its own held-out synthetic split — not comparable to laya's or Jev's numbers · source · T3 · huggingface.co

SoftwareSourceCode (sources)

• code-as-world — MirroS-Lab, Apache-2.0, 198★, Python (created 2026-08-27): inference + quantiphy evaluation for the two code-as-world-vl checkpoints, a vLLM recipe and one MuJoCo example. The discovery agent itself is not in the repository — what ships is the model trained on its output · source · T1 (artifact facts only; numbers are T3 in the report) · github.com

• llm-abliteration — NousResearch fork of jim-plus/llm-abliteration, 150★: measure a refusal direction from harmful vs harmless prompts, then orthogonalize weights against it; optional --projected and --normpreserve (norm-preserving biprojected variant). Interpretability used as a weight-editing tool; lists six tested model families but publishes no before/after scores · source · T1 (artifact facts only) · github.com

• mlalgorithms — rushter, MIT, ~11.1k★: from-scratch NumPy implementations from linear regression to LSTMs and deep Q-learning, written for legibility over speed · source · T2 · github.com

• transformers-tutorials — NielsRogge, MIT, 11.7k★: one folder per model, notebooks for inference/fine-tuning/evaluation across DETR, ViT, SAM, DINOv2, LayoutLM v1–v3, Donut, TrOCR, BLIP-2, LLaVA, TAPAS and more; fine-tunes three ways (native PyTorch / Lightning / Trainer). The spoke's first source on the middle layer — the training loop as executable code, and no numbers anywhere · source · T2 · github.com

• supervision — Roboflow, MIT, 48.7k★: model-agnostic Python utilities for the layer after a vision model — sv.Detections, annotators, multi-object tracking, YOLO/COCO/VOC dataset conversion, mAP metrics. Stretches the spoke's train-time boundary; the corpus's most concrete evidence on demo-to-production-gap · source · T2 · github.com

Collection (sources)

• awesome-agentic-time-series — TROUBADOUR000, MIT, 188★: 243 papers on LLMs and agents for time series, companion to an unpublished survey. Taxonomy of agent roles (perception / reasoning / planning-action / memory / world-model) and 31 foundation models beside timesfm — but 16 papers on evaluation against 133 on building, which is demo-to-production-gap seen from the literature side · source · T3 · github.com
• ml-system-design-case-studies — Engineer1999, ~10.9k★: 300+ production ML case studies from 80+ companies (2017–2023), indexed by industry and use case; a bibliography, and a survivorship-biased one · source · T3 · github.com

Book (sources)

• younes-introduction-to-machine-learning — laurent-younes (johns-hopkins-university AMS), arXiv 2409.02668 v2 (May 2025, 674 pp, CC BY 4.0), held in full, via Telegram: the graduate mathematical register — 60 pp of optimization (SGD, Adam, proximal, ADMM), ~140 pp of graphical models, 18 pp of neural nets, a closing chapter of generalization bounds; no exercises. Third textbook register beside Lindholm (engineers) and MacKay (information theory) · source · T1 · arxiv.org
• lindholm-machine-learning-first-course — Lindholm, Wahlström, Lindsten & Schön, Machine Learning: A First Course for Engineers and Scientists (cambridge-university-press; online draft of 4 March 2026, 348pp, personal use only): the spoke's missing middle layer as a syllabus — k-NN and trees first, then regression and regularisation, bias–variance, parametric learning and SGD, neural nets with backprop, bagging/boosting, kernels and SVMs, Gaussian processes, mixtures and clustering. Chapter 11 adds the practice (the train-val mismatch decomposition, "Can I Trust my Machine Learning Model?") and Chapter 12, contributed entirely by david-sumpter, argues in §12.2 that company ML research "is (in part) a form of advertising" — the spoke's own evidence complaint, stated by a textbook · source · T1 · smlbook.org
• information-theory-inference-learning-algorithms — David MacKay, Cambridge University Press 2003 (held copy: v7.2, 4th printing, 2005): 50 chapters unifying Shannon's information theory with Bayesian inference and neural networks; the spoke's only textbook and only T1 · source · T1 · inference.org.uk

Article (sources)

• ml-models-as-agent-tools — MachineLearningMastery (Vinod Chugani, 2026-09-10): a trained classifier becomes a tool an LLM agent calls — the agent plans, fetches, thresholds and routes; the model only predicts. One insurance-fraud example, LangChain named, nothing measured; the pattern's shape and its hidden threshold rule · source · T3 · machinelearningmastery.com
• finetuning-explained-kdnuggets — KDnuggets (Kanwal Mehreen, 2026-07-10): pretraining vs fine-tuning, full vs PEFT, and the caution against reaching for fine-tuning first; vocabulary without measurements · source · T3 · kdnuggets.com
• timesfm-2-5-forecasting-tutorial — MarkTechPost (2026-08-01): end-to-end timesfm walkthrough — zero-shot forecast, rolling-origin backtest, XReg covariates, interval-based anomaly detection, context ablation, throughput and robustness, plus a production checklist. All of it on synthetic data the author generated, so the method transfers and the accuracy numbers don't · source · T3 · marktechpost.com
• timesfm-3 — Google Research blog (Jain & Sen, 2026-08-31): TimesFM 3.0 — 330M parameters, natively multivariate (causal temporal + full variate attention), covariates and multiple targets pretrained in, whole horizon in one forward pass, >1T training time points. Claims the top rank on gift-eval, FEV-Bench and TIME-leaderboard among foundation models only, in charts with no numbers; and the licence goes Apache-2.0 → non-commercial, unmentioned · source · T3 · research.google
• male-fly-connectome — Google Research blog (Januszewski & Jain, 2026-09-03): the complete male Drosophila CNS connectome — 166,000 neurons, 125 million synapses, led by HHMI Janelia, in Cell (189(18), 111 authors, verified via Crossref, not read). Flood-filling networks segment; synthetic neurons in the training data improved the PATHFINDER system; then years of expert human proofreading, which the post names as the field's constraint. Manufacture and annotation running consecutively, and no number published for either · source · T3 · research.google
• unsloth-amd-support — AMD developer blog (2026-07-20): Unsloth on Instinct/Radeon/Ryzen AI via ROCm; 1.39× faster and 1.33× leaner than TRL+FA2 on Llama-3.1-8B LoRA SFT, all first-party; plus the BF16 adapter-drift argument · source · T3 · amd.com
• ai-projects-fail-infrastructure-people — The New Stack (Meredith Shubel, 2026-07-06): failure is infrastructure and people, not model quality. Body never readable — four extraction routes failed; the wiki holds the headline only · source · T4 · thenewstack.io
• laya-doom-medium — Christian Graham on Medium (2026-09-18): a weekend spent teaching laya to play Doom — one focused noul question to fix "never shoots," hand-written guard rules on top, exploration still unsolved. One practitioner's account, no measurement, corroborates only that the model runs locally and fast enough for a real-time loop · source · T4 · medium.com
• jev-typesafe-typed-decisions-guide — Hugging Face community post, byline "bna" (2026-09-25): a how-to for calling Jev safely — a decision contract, hand-rolled runtime validation instead of a type cast, a WorkflowResult type with decision/review/unavailable states, a staged rollout. Technically accurate against jev-ai-sdk-guide, but from an account whose own profile is an SEO content mill · source · T4 · huggingface.co

ScholarlyArticle (sources)

• xgboost — Chen & Guestrin, KDD 2016 (T1): the system that made boosting the default — 17 of 29 Kaggle 2015 winning solutions, every KDDCup 2015 top-10 team; sparsity-aware splits, weighted quantile sketch, cache-aware out-of-core blocks; Higgs-1M 0.68 s/tree vs scikit-learn's 28.51 s at equal AUC, 1.7 billion Criteo examples on one machine · source · T1 · arxiv.org
• tabular-deep-learning-benchmark — Grinsztajn, Oyallon & Varoquaux, NeurIPS D&B 2022 (T1): 45 tabular datasets, ~20,000 tuning hours per learner — tree ensembles remain state-of-the-art at ~10K samples, and the three reasons are inductive biases (uninformative features, rotation invariance, irregular targets) · source · T1 · arxiv.org
• gift-eval — Aksu, Woo, Liu et al., Salesforce AI Research, arXiv 2410.10393 (T1): 23 real datasets, 144k series, 177M points, 97 configurations, 17 baselines — and the winner overall is PatchTST, a supervised transformer trained per dataset (CRPS 0.563, rank 5.72), ahead of every time-series foundation model. Foundation models beat Auto_ARIMA (0.811) comfortably and win daily-to-yearly and univariate; deep models win noisy, high-frequency and multivariate. No boosted-tree baseline anywhere in it, so open question 7 half-closes · source · T1 · arxiv.org
• phase-transitions-in-attention — Lavie/Fischer et al., arXiv 2606.12058: a closed-form Bayesian posterior over an attention layer, reduced by symmetry to two order parameters, showing the copy sub-circuit appears as a phase transition in training-set size — first-order (no precursor) for softmax, second-order plus crossover for linear attention. Adam lands on the MAP, so the point estimate is the posterior mode here. One layer, one head, synthetic copy task · source · T1 · arxiv.org
• data-cascades — nithya-sambasivan, Kapania, Highfill, Akrong, Paritosh, Aroyo (Google Research), CHI 2021: interviews with 53 AI practitioners in India, East/West Africa and the USA on high-stakes AI (cancer detection, suicide prevention, landslide detection). Names data cascades — compounding downstream failures seeded by upstream data decisions — and measures them: 92% prevalence, 45.3% with two or more in one project; "pervasive, invisible, delayed, but often avoidable." Closes the founding gap's data half, and it is HCI, not ML, which is why searching the ML side never found it · source · T1 · research.google
• measuring-annotator-agreement — Braylan, Alonso & Lease, arXiv 2212.09503 (ACM, 2022): inter-annotator agreement works for categorical and ordinal labels and breaks down on the seven task shapes real datasets use — bounding boxes, keypoints, sequence tags, ranked lists, free-text translation, numeric vectors, syntax trees — because each needs a distance function chosen before agreement means anything. Proposes two more interpretable measures. The instrument for catching the failures the other two sources measure · source · T1 · arxiv.org
• pervasive-label-errors — curtis-northcutt, Athalye, Mueller, NeurIPS 2021 (Datasets & Benchmarks), arXiv 2103.14749: label errors in 10 standard CV/NLP/audio test sets found by confident learning and confirmed by crowdworkers (51% of flagged candidates real). ≥3.3% wrong on average, ≥6% of ImageNet validation. Correcting labels reverses rankings — ResNet-18 beats ResNet-50, VGG-11 beats VGG-19 — so the field's ruler has a measured error rate · source · T1 · arxiv.org
• who-annotates-in-nlp — Kunilovskaya et al., arXiv 2606.02255 (2026): task-level audit of annotation reporting across ACL venues 2018–2025, 2,667 tasks from 1,603 papers. Operational facts get reported (recruitment 90%, expertise 86%, total items 86%); the facts that let a reader judge validity do not — training 19%, guidelines released 34%, IAA value 35%. Explains why the corpus could never find a documented labelling operation: the base rate is low · source · T1 · arxiv.org
• absa-annotator-comparison-cost — Donhauser et al. (Regensburg, arXiv 2605.03624), read from the PDF: the same German ABSA task annotated by experts, students, crowdworkers and an LLM, with the cost of each and the downstream models trained on each. The corpus's first total cost of a labelling operation — £828, £0.41 per three-way annotated example including platform fees, 30 Prolific workers at £9/hour. Two of the four options have no price (course credit; expert refinement of existing labels), and "LLM annotation" spans $0.0025 self-hosted to $0.36 frontier, straddling the human price. The quality half: high agreement, only student-level model performance — and the failure is missing rare-category instances, which no agreement measure can see · source · T1 · arxiv.org
• safetutors — Hazra et al., arXiv 2603.17373 (2026): a pedagogical-safety benchmark for LLM tutors, held here for Appendix D — one labelling operation documented end to end. Three stages, three workforces, three stated pay rates ($5/$15/$10 per hour), an ≥80% agreement-with-gold gate, printed qualification pass rates (24/33 crowd recruits), verbatim guidelines, adjudication rates (4.2% escalated) and 15,300 labels. The artifact edge 1a spent three hunts looking for · source · T1 · arxiv.org
• annotation-requirements-autonomous-driving — Saeeda, Mohamad, Knauss, Horkoff & Nouri, arXiv 2511.15859 (2025): 19 interviews at six automotive companies on how annotation requirements are written and governed. Guidelines past 140 pages, manual review cut 10%→5% for budget, rework at 3× proper labelling, ~20% of collected data ever annotated, 15% of a point-cloud set lost to version conflicts. Practitioner testimony, not logs — recorded as such · source · T1 · arxiv.org
• skillsmith — Dery et al., arXiv preprint: weights as a modality — an LLM composes new prefix-tuning weights from existing prefix weights + text, beating text-only and weight-space-only baselines; the spoke's first primary research on adaptation, and a third stance on the weights-as-distribution seam · source · T2 · arxiv.org
• paraphrastic-similarity-parallel-translations — Wieting, Gimpel, Neubig & Berg-Kirkpatrick, ACL 2019: trains cross-lingual sentence embeddings directly on bitext, skipping paraphrase-corpus construction; beats deep NMT-encoder baselines on 4/6 cross-lingual STS sets while encoding ~330× faster (855,571 vs 2,601 sentences/sec). Also shows random untrained encoders beat literature baselines monolingually via lexical overlap alone — a measured "check your baseline" result · source · T1 · arxiv.org

Report (sources)

• code-as-world-report — mirros, 30pp, 2026-08-27, authored as "MirroS Team", no venue and a placeholder arXiv id: the phenomenon–mechanism argument, the agentic-discovery-loop algorithm, the QuantiPhy table, and two limitations sections the release did not have to write — the loop can converge on a plausible-but-wrong world (§4.4), and the trained model never internalises the loop (§5.5) · source · T3 (self-run benchmark on self-trained models) · mirros.ai

Dataset

• open-jev-dataset — the mostly-synthetic training/eval data behind open-jev-9b and its 2B/27B siblings: 12 configs, 520,199 rows across 60 splits (heavily overlapping — don't sum), CC0-1.0 on generated content only. Its "OOD" split is source-specific variation within the generator's own configs, not held-out real-world domains · training data
• quantiphy — stanford-university + UST, CVPR 2026: 3,300+ instances over 569 videos asking a VLM to estimate size/velocity/acceleration in world units from monocular video given one reference quantity; 2S/2D/3S/3D subsets scored by MRA. The yardstick for code-as-world-vl, and by its own scope a narrow one · benchmark

Person (entities)

• laurent-younes — JHU applied mathematician; author of younes-introduction-to-machine-learning · author

• andreas-lindholm — first-named author of lindholm-machine-learning-first-course · author

• niklas-wahlstrom — co-author of lindholm-machine-learning-first-course · author

• fredrik-lindsten — co-author of lindholm-machine-learning-first-course · author

• thomas-schon — senior author of lindholm-machine-learning-first-course · author

• david-sumpter — wrote Chapter 12 (ethics) of lindholm-machine-learning-first-course; the source of §12.2 on research-as-advertising · contributor

• niels-rogge — HuggingFace Transformers contributor (ViT, DETR, LUKE and others) and author of the tutorials; expert on the code, interested party on the library choice · author

• nithya-sambasivan — lead author of data-cascades, published through Google Research; HCI method — 53 practitioner interviews, no model trained · author

• curtis-northcutt — first author of pervasive-label-errors; confident learning, the method behind the label-error estimates · author

• david-mackay — Cambridge information theorist; author of the 2003 textbook, and a co-rediscoverer of the LDPC codes it teaches · author

• itay-lavie — Hebrew University / Harvard SEAS; co-first and corresponding author of phase-transitions-in-attention · author

• zohar-ringel — Racah Institute, Hebrew University; senior author, statistical-mechanics methods applied to learning · author

• moritz-helias — Jülich (IAS-6) / RWTH Aachen; senior author, field-theoretic and Bayesian feature-learning theory · author

• john-wieting — Carnegie Mellon University; first/corresponding author of paraphrastic-similarity-parallel-translations and the recurring name across the paraphrastic-embeddings line it builds on · author

• rima-hazra — TU Eindhoven; corresponding author of safetutors, the paper whose annotation appendix this spoke keeps · author

• hina-saeeda — chalmers-university-of-technology; first author of annotation-requirements-autonomous-driving, working the requirements-engineering side of labelling · author

• viren-jain — Research Scientist, Google Research; co-author of male-fly-connectome and lead of the connectomics partnership with hhmi-janelia. Not Ayush Jain of timesfm-3 · author

• michal-januszewski — Research Scientist, Google Research; co-author of male-fly-connectome, on the flood-filling-network line of automated reconstruction · author

• nandha-kishor-m — CEO of convai-innovations; former PI, IIT Palakkad Technology IHUB Foundation; GitHub author of laya · founder

• zefan-cai — author of open-jev-9b and open-jev-dataset; no stated affiliation beyond "I help community use large language models" · author

Organization (entities)

• johns-hopkins-university — CollegeOrUniversity: AMS department behind the Younes text · thin node

• hhmi-janelia — ResearchOrganization: led the male fly connectome and did the proofreading. The third answer to who annotates — not crowdworkers, not a vendor, but expert staff at the institute that owns the science, for years · research site

• amd — Corporation: Instinct/Radeon/Ryzen AI silicon and the ROCm stack — the non-CUDA training path · vendor

• roboflow — Corporation: computer-vision platform company; maintainer of supervision, with a commercial inference product the library integrates · vendor

• kdnuggets — NewsMediaOrganization: data-science practitioner outlet; orientation-tier explainers · publisher

• marktechpost — NewsMediaOrganization: AI-news outlet running long runnable Colab walkthroughs; checkable API surface, synthetic datasets, no visible editorial verification · publisher

• cambridge-university-press — the publisher that let a full university textbook stay free on the author's site · publisher

• university-of-cambridge — CollegeOrUniversity: where the textbook was lectured for nine years before it was a book · institution

• chalmers-university-of-technology — CollegeOrUniversity: Gothenburg; academic base of the annotation-requirements study, co-authored with RISE, Gothenburg and Volvo Cars · institution

• mirros — Organization: the lab behind Code-as-World; "an OS where the world model and the agent co-evolve." No team, founders, location or funding published, and the report names no authors · lab

• convai-innovations — Organization: maker of laya; a prior track record of Hindi-language and medical (ECG) models on Hugging Face, not a single-product account · vendor

Linked cross-wiki, not duplicated: The New Stack (the-new-stack, ../ai-governance-wiki) · Google (google, ../llm-providers-wiki) as timesfm's maker · backtesting (backtesting, ../quant-trading-wiki) as the trading-side sibling of rolling-origin evaluation · Qwen (qwen, ../llm-providers-wiki) as code-as-world-vl's base · Stanford University (stanford-university, ../agentic-tooling-wiki) as quantiphy's home · LLMs can't jump (llms-cant-jump, ../research-wiki), whose missing-abduction argument code-as-world-report §5.5 confirms.