Books, papers and documentation

Anastasios N. Angelopoulos, Stephen Bates · arXiv:2107.07511 · 2021

Tutorial on split conformal prediction: calibration set, nonconformity scores, finite-sample marginal coverage; Day 26 prediction intervals.

Checked 2026-10-11: Title, authors, submission and latest-version dates via arXiv API record for 2107.07511.

Jon Vaver, Jim Koehler · Google Research technical report · 2011

Checked 2026-10-12: Google Research publication page via search results: title, authors, 2011. Content: non-overlapping geographic regions are randomly assigned to control or treatment and each receives its condition through geo-targeted ads; a later paper notes geo regression is less suitable when few geographic units are available. Abstract-level reading only.

SAS Institute

Checked 2026-10-11: Landing page title and description only: FVA analysis is presented as a way to find waste (steps that do not improve accuracy) in the forecasting process, with guidance on data gathering, analysis and reporting. Full text not read; vendor white paper.

M Open Forecasting Center (organizers of the M5 competition)

Primary description of the M5 accuracy metric and its dollar-sales weights.

Checked 2026-10-12: Downloaded the 16-page PDF from the Mcompetitions/M5-methods repository and read the Evaluation and Weighting sections: 42,840 series, 28-day horizon, ranking by WRMSSE = sum of weight x RMSSE; weights computed from the cumulative actual dollar sales (units times price) over the last 28 observations of the training sample; 12 aggregation levels, all levels equally weighted. Publication date not printed in the sections read.

Vincent Dumoulin, Francesco Visin · 2016

CNN shape arithmetic (day 36).

Checked 2026-10-11: arXiv 1603.07285 confirms title and authors (preprint). Abstract: relationships between input shape, kernel shape, zero padding, strides and output shape for convolution, pooling and transposed convolution.

Priya Goyal, Piotr Dollar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, Kaiming He · Technical report (arXiv) · 2017

Coupling of batch size and learning rate; warmup (days 33, 34).

Checked 2026-10-11: arXiv 1706.02677 confirms title and authors; comment "Tech report". Abstract describes the linear scaling rule for learning rate vs minibatch size and a warmup scheme.

Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, et al. · arXiv:2510.15821 (technical report) · 2025

Chronos-2 adds zero-shot multivariate and covariate-informed forecasting; vendor benchmark claims (fev-bench, GIFT-Eval) are self-reported; Day 27.

Checked 2026-10-11: Title, leading authors, date via arXiv API record 2510.15821; linked as 'Chronos-2 Report' from the official amazon-science/chronos-forecasting README.

Stanford CS231n course staff

Checked 2026-10-11: Course notes page: centred-difference numerical gradient; relative error with rough thresholds (above 1e-2 likely wrong, below 1e-7 good, up to about 1e-4 tolerable with kinks such as ReLU); use double precision; turn off dropout/augmentation during checks; check data loss and regularisation separately; loss at chance (for 10-class softmax about 2.302); overfit a tiny subset of about 20 examples; update-to-weight ratio around 1e-3; random search with learning rate and regularisation on a log scale.

Chenghao Liu, Taha Aksu, Juncheng Liu, Xu Liu, Hanshu Yan, Quang Pham, Silvio Savarese, Doyen Sahoo, Caiming Xiong, Junnan Li · arXiv:2511.11698 · 2025

Successor in the Moirai family (preprint); Day 27 optional reading.

Checked 2026-10-11: Title, authors and version dates via arXiv API record 2511.11698. No peer-reviewed venue confirmed.

Azul Garza, Cristian Challu, Max Mergenthaler-Canseco · arXiv:2310.03589 · 2023

Commercial API time-series foundation model from Nixtla; preprint only, closed weights; Day 27 build-vs-buy discussion.

Checked 2026-10-11: Title, authors and version dates via arXiv API record 2310.03589. Not peer reviewed (preprint).

Ayush Jain, Rajat Sen (Google Research) · Google Research Blog · 2026

Vendor announcement of TimesFM-3 (native multivariate + covariates); benchmark rankings are self-reported, not peer reviewed; Day 27.

Checked 2026-10-11: Title, authors, date August 31, 2026 on the Google Research blog page; page states models up to TimesFM-2.5 (Sept 2025) were univariate-only.

Evan Miller · 2024

Recommends reporting standard errors, clustering them when questions are grouped, comparing two models with paired question-level differences, and using power analysis to choose the number of questions. Supports insights 0555 and 0556.

Checked 2026-10-11: arXiv abs page (title, sole author, abstract; no venue, preprint); recommendations checked in the arXiv HTML full text (Sections 2.1, 2.2, 4.2, 5).

Fanjia Yan, Huanzhi Mao, Charlie Cheng-Jie Ji, Ion Stoica, Joseph E. Gonzalez, Tianjun Zhang, Shishir G. Patil · Gorilla project blog, UC Berkeley · 2024

Supports insight 1513. Methodology of one benchmark, not a universal standard.

Checked 2026-10-12: Blog page opened 2026-10-12. AST evaluation parses each model output into a syntax tree and checks function name, required parameters present, no undocumented parameters, type and value checks (strings compared case-insensitively after normalisation, lists in order); executable evaluation runs the calls and compares outputs (exact match, 20% tolerance for live numeric data, structural checks); AST is used because not all functions are easily executable (for example Java); relevance detection gives functions that do not fit the request and expects no call. The title field here is descriptive (the exact post title was not recorded); the URL, authors and last-updated date were read on the page.

OpenAI (280 listed authors) · 2023

Figure 8 shows calibration of answer probabilities worsening after post-training. Vendor technical report about one model; supports insight 0565.

Checked 2026-10-11: arXiv abs page (v6 2024-03-04; no journal-ref, not peer reviewed). Section 5 (Limitations) and Figure 8 read in the PDF: on an MMLU subset the pre-trained model's answer log-probabilities are well calibrated and post-training (PPO) significantly reduces calibration.

Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever · OpenAI technical report · 2018

Decoder-only generative pretraining followed by supervised fine-tuning; supports day 43 (pretraining and fine-tuning).

Checked 2026-10-11: Opened the OpenAI-hosted PDF (≈0.5 MB); first page shows title and four authors (OpenAI). PDF has no printed date; year 2018 from the GPT-2 report, which cites "the OpenAI GPT model (Radford et al., 2018)". Not peer reviewed. The openai.com blog page returned HTTP 403.

Saurav Kadavath et al. (36 authors) · 2022

Larger models are well calibrated on diverse multiple-choice and true/false questions in the right format; models can be trained to predict whether they know an answer, with weaker calibration on new tasks. Supports insight 0565.

Checked 2026-10-11: arXiv abs page (v4 2022-11-21) and abstract; no journal-ref, no venue confirmed (preprint).

Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever · OpenAI technical report · 2019

Zero-shot task transfer from a larger decoder-only LM trained on WebText; supports days 42-43 (emergence of prompting from LM pretraining).

Checked 2026-10-11: Opened the OpenAI-hosted PDF (≈0.6 MB); first page shows title and six authors. Year 2019 from the citation entry in the official openai/gpt-2 GitHub README. Not peer reviewed.

Isaac Ong et al. · arXiv preprint 2406.18665 · 2024

Learned router choosing between a strong and a weak model per query, evaluated as a cost-quality curve; supports project-044.

Checked 2026-10-11: arXiv export API confirms title, first author Ong, eight authors, first submission 2024-06-26; no journal reference recorded in the arXiv entry.

Karl Cobbe et al. (OpenAI) · arXiv preprint 2110.14168 · 2021

Introduces GSM8K; supports project-044 dataset description.

Checked 2026-10-11: arXiv export API confirms title, first author Cobbe and first submission 2021-10-27; no journal reference recorded.

Google SRE team

Checked 2026-10-12: Opened 2026-10-12. Defines burn rate relative to the SLO; shows that a single long window has a long reset time; recommends multiwindow, multi-burn-rate alerts where a short window (about 1/12 of the long one) must also exceed the threshold; example for a 99.9% SLO: page at 14.4x over 1 hour (5 minute confirmation) and 6x over 6 hours (30 minutes), ticket at 1x over 3 days (6 hours).

Alec Warner, Stepan Davidovic et al., Google

Checked 2026-10-11: Page title and chapter 16 on sre.google; canary = partial, time-limited deployment compared against a control; small set of user-visible, attributable metrics; pause and roll back when the canary looks bad.

Michael Nygard · Cognitect blog · 2011

Origin of the short architecture decision record format; supports the decision-record elements in Day 58 and the ADR template of the S05 scenario.

Checked 2026-10-11: Page opened: title, author Michael Nygard and date November 15, 2011 on the Cognitect blog. Describes a short record per architecturally significant decision with title, status (for example proposed, accepted, superseded), context, decision and consequences, so later readers can see why a choice was made and whether it still holds.

Google SRE team

Checked 2026-10-12: Opened 2026-10-12. Triage section: the first aim during an outage is to make the system work as well as it can, not to find the root cause; emergency measures include moving traffic away, dropping traffic and disabling subsystems; evidence such as logs should be preserved for later analysis; possible data corruption can justify freezing the system.

Andrew Stribblehill (ed. Kavita Guliani), Google

Checked 2026-10-11: Page title and chapter 14 on sre.google; incident commander, operations and communication roles, a live incident document, explicit handoffs, and declaring incidents early.

Rob Ewaschuk (ed. Betsy Beyer), Google

Checked 2026-10-11: Page title and chapter 6 on sre.google; pages should be urgent, actionable and need human judgment; four golden signals; noisy alerts get ignored and non-urgent issues belong on dashboards.

John Lunney, Sue Lueder (ed. Gary O'Connor), Google

Checked 2026-10-11: Page title and chapter 15 on sre.google; describes blameless postmortems focused on contributing causes, triggers defined in advance (user-visible downtime, data loss, rollback, monitoring failure), and preventive action items.

Martin Zinkevich (Google) · last updated 2025-08-25

Practitioner guidance from Google engineering; supports 'rules before ML' and 'simple first model' (Day 1).

Checked 2026-10-11: Title, author and last-updated date on the Google for Developers page; Rule #1 (launching without ML is fine; heuristics first) and Rule #4 (keep the first model simple, get the infrastructure right).

Mike Ulrich (chapter author); Google · Site Reliability Engineering (O'Reilly Media), free online edition at sre.google

Retry, backoff and deadline discipline reused for LLM and tool calls on day 52.

Checked 2026-10-11: Page loads; chapter 22 by Mike Ulrich. Guidance: randomized exponential backoff, limited retries per request, a per-process retry budget, retries at several layers multiply (three retries at three layers can mean 64 attempts), do not retry permanent errors, set deadlines and propagate the remaining deadline downstream.

Martin Wattenberg, Fernanda Viegas, Ian Johnson · Distill · 2016

Why 2D projections of embeddings are not evidence (day 38).

Checked 2026-10-11: Distill page loads; authors and date Oct 13 2016 shown; citation DOI 10.23915/distill.00002. Sections state cluster sizes in a t-SNE plot mean nothing and distances between clusters might not mean anything.

Kimball Group

Checked 2026-10-12: Opened 2026-10-12. Fully additive facts sum across all dimensions; semi-additive facts (account balances) sum across all dimensions except time; non-additive facts (ratios) cannot be summed, so store the additive components and compute the ratio from the aggregated totals.

Kimball Group

Checked 2026-10-12: Opened 2026-10-12. Header-level facts such as a freight charge should be allocated to the line grain with business-defined rules so they can be rolled up by every dimension; a separate header-level fact table is often unnecessary.

Kimball Group

Checked 2026-10-12: Opened 2026-10-12. Dimensions are conformed when shared attributes have the same names and values, so data from separate fact tables can be combined in drill-across reports; they are defined once and reused.

Kimball Group

Checked 2026-10-12: Opened 2026-10-12. Factless fact tables record events with no numeric measure; to show what did not happen, use a coverage table of all possible events and an activity table of events that occurred, and subtract.

Kimball Group

Checked 2026-10-11: Page title and organisation; content: declaring the grain is the pivotal early step, the grain states what one fact-table row represents, prefer the atomic grain, and different grains must not be mixed in one fact table.

Kimball Group

Checked 2026-10-12: Opened 2026-10-12. Operational transaction header/line (parent/child) schemas; header-level dimension keys and degenerate dimensions should be carried onto the line-level fact table.

Kimball Group

Checked 2026-10-11: Page title 'Late Arriving Dimension'; facts can arrive before their dimension context; post the fact anyway with special dimension rows holding the unresolved natural key and unknown attributes, then overwrite them (type 1) when the context arrives.

Kimball Group

Checked 2026-10-12: Opened 2026-10-12. A fact arriving after newer dimension rows were loaded must be assigned the surrogate keys of the dimension rows in effect when the event happened, not the latest rows.

Kimball Group

Checked 2026-10-12: Opened 2026-10-12. Joining two fact tables through shared keys in BI SQL makes the result row count unpredictable and gives wrong answers; query each fact table separately (drill across) and merge the result sets on shared attributes.

Kimball Group

Checked 2026-10-12: Opened 2026-10-12. For systems spanning time zones record both a universal reference time and the local time of each event: two foreign keys to a date dimension playing different roles, plus a time-of-day dimension if needed.

Kimball Group

Checked 2026-10-12: Opened 2026-10-12. A fact row can relate to several dimension values (several diagnoses per treatment) through a group key into a bridge table with one row per value; with type 2 dimensions the bridge needs effective and expiration timestamps and must be filtered to a point in time. The page does not discuss weighting factors.

Kimball Group

Checked 2026-10-12: Opened 2026-10-12. Natural keys come from source systems under rules the warehouse team does not control (an employee number can change after a rehire); a durable key is assigned by the warehouse and never changes for an entity.

Kimball Group

Checked 2026-10-12: Opened 2026-10-12. Null measurements are acceptable because aggregate functions handle them; foreign keys must never be null (it breaks referential integrity) and should point to a placeholder dimension row for unknown or not-applicable values.

Kimball Group

Checked 2026-10-12: Opened 2026-10-12. Each row summarizes measurement events over a fixed period (day, week, month); the grain is the period, not the transaction; tables are densely populated and usually add a row even for periods with no activity.

Kimball Group

Checked 2026-10-12: Opened 2026-10-12. One physical dimension (for example the date dimension) is reused several times in a fact table, each foreign key pointing to a separate view with distinct column names for its role.

Kimball Group

Checked 2026-10-12: Opened 2026-10-12. A Type 1 change overwrites the attribute so history is lost; any aggregate fact tables or OLAP cubes depending on the changed attribute must be recomputed.

Kimball Group

Checked 2026-10-11: Page title 'Type 2: Add New Row'; a change adds a new dimension row with a new surrogate key that fact rows use from the change until the next change; recommends effective date, expiration date and current-row indicator columns.

Kimball Group

Checked 2026-10-12: Opened 2026-10-12. Rapidly changing attributes are split into a separate mini-dimension with its own key; the fact table carries the keys of both the base dimension and the mini-dimension.

Python Enhancement Proposals

Checked 2026-10-12: Opened 2026-10-12. fold attribute (0 or 1) distinguishes the two occurrences of a repeated local time when clocks move back; for nonexistent local times in a forward gap fold selects which UTC offset is applied; naive comparison ignores fold.

IANA / tz database maintainers

Checked 2026-10-12: Opened 2026-10-12. Governments control time-zone and DST rules and change them with little notice, so the tz database is updated periodically and end users may receive updates late; region names such as America/New_York are preferred over abbreviations, which are ambiguous and not unique identifiers.

General Data Protection Regulation (text reproduced on gdpr-info.eu) · 2016

Why deletion must reach derived stores such as indexes and caches; check with counsel for the actual obligations.

Checked 2026-10-12: Page opened (unofficial reproduction of Regulation (EU) 2016/679, not EUR-Lex). Article 17(1): the data subject may obtain erasure of personal data and the controller must erase it without undue delay where one of the listed grounds applies, for example data no longer necessary for its purpose or consent withdrawn. Exceptions in 17(3) were not reviewed. Not legal advice.

European Parliament and Council of the European Union (text); gdpr-info.eu (reproduction) · 2016

Checked 2026-10-12: Page opened: 22(1) right not to be subject to a decision based solely on automated processing, including profiling, that has legal or similarly significant effects; 22(2) exceptions (contract necessity, authorising law with safeguards, explicit consent); 22(3) safeguards at least include the right to obtain human intervention, express a point of view and contest the decision; 22(4) special-category data limits. Unofficial reproduction.

European Parliament and Council of the European Union (text); gdpr-info.eu (reproduction) · 2016

Checked 2026-10-12: Page opened: Article 4(5) defines pseudonymisation as processing personal data so that it can no longer be attributed to a specific person without additional information, provided that information is kept separately and protected by technical and organisational measures. The official EUR-Lex text did not render in the fetch tool (JavaScript page); the reproduction is unofficial, so check EUR-Lex for legal use.

European Parliament and Council of the European Union (text); gdpr-info.eu (reproduction) · 2016

Checked 2026-10-12: Page opened: Article 5(1)(b) purpose limitation: collected for specified, explicit and legitimate purposes and not further processed in a manner incompatible with them; further processing for archiving in the public interest, scientific or historical research or statistical purposes is not considered incompatible, subject to Article 89(1) safeguards. Unofficial reproduction.

European Parliament and Council of the European Union (text); gdpr-info.eu (reproduction) · 2016

Checked 2026-10-12: Page opened: Article 7(3) the data subject can withdraw consent at any time; withdrawal does not affect lawfulness of processing before withdrawal; the data subject is informed beforehand; withdrawing must be as easy as giving consent. Unofficial reproduction.

OWASP Gen AI Security Project · 2025

Risk page behind the docs-owasp-llm-top10 list entry; supports day 53.

Checked 2026-10-11: Page title 'LLM01:2025 Prompt Injection - OWASP Gen AI Security Project'. Defines direct and indirect injection; states that RAG and fine-tuning do not fully mitigate it; mitigations include deterministic validation of output formats, least privilege with app-held API tokens, human approval for high-risk actions, segregating and labelling external content, and adversarial testing.

OWASP Gen AI Security Project · 2025

Supports day 53 (validate model output before SQL, HTML, shell).

Checked 2026-10-11: Page title 'LLM05:2025 Improper Output Handling - OWASP Gen AI Security Project'. Model output passed unvalidated to downstream components can enable XSS, SSRF, privilege escalation or remote code execution; mitigations: treat output as untrusted (zero trust), context-aware output encoding, parameterised queries, content security policy, logging.

OWASP Gen AI Security Project · 2025

Supports days 49 and 53 (tool allow-lists, narrow tools, approvals, least privilege).

Checked 2026-10-11: Root causes: excessive functionality, excessive permissions, excessive autonomy. Mitigations: minimise extensions and their functionality, avoid open-ended extensions such as shell runners, least privilege, execute extensions in the user's context with limited scopes, human approval for high-impact actions, complete mediation (authorisation enforced downstream, not by the LLM), plus logging and rate limiting to limit damage.

OWASP Gen AI Security Project · 2025

Risk page for vector stores in RAG; supports day 53.

Checked 2026-10-12: Page title 'LLM08:2025 Vector and Embedding Weaknesses - OWASP Gen AI Security Project'. Names weak access control, cross-context leaks in multi-tenant vector stores, embedding inversion and data poisoning as risks; mitigations include permission-aware vector stores with logical and access-level partitioning, source validation, data classification and detailed retrieval logging.

OWASP Gen AI Security Project · 2025

Supports days 53 and 54 (quotas and budgets as security controls).

Checked 2026-10-11: Page title 'LLM10:2025 Unbounded Consumption'. Excessive uncontrolled inference can cause denial of service, financial loss (denial of wallet), model extraction; mitigations include input size limits, rate limiting and per-user quotas, resource allocation management, timeouts and throttling, monitoring.

MITRE

ATT&CK-style matrix of adversarial techniques and case studies against AI systems; supports day 53 threat modelling.

Checked 2026-10-11: atlas.mitre.org loads with title "MITRE ATLAS" (content is rendered client-side). The official mitre-atlas/atlas-data GitHub README describes ATLAS as a public knowledge base of adversary tactics, techniques and procedures targeting AI systems.

European Parliament and Council of the European Union (text); gdpr-info.eu (reproduction) · 2016

Checked 2026-10-12: Page opened: pseudonymised data that could be attributed to a person with additional information is to be considered information on an identifiable person; identifiability is judged by all means reasonably likely to be used (cost, time, available technology); anonymous information is outside the principles. Unofficial reproduction of the official recital.

OWASP Foundation

Applies to any URL-fetching tool an agent can call.

Checked 2026-10-12: Opened 2026-10-12 (via a page summary). Allow-lists preferred over bypass-prone deny-lists; cloud metadata address 169.254.169.254 named as a typical SSRF target for credentials; disable redirect following; validate resolved IP addresses against permitted networks (DNS rebinding) and connect only to validated addresses; network-layer segmentation so the application can reach only its legitimate flows.