# Citation Policy The rules every Webspinner Academy course follows. They exist because the subject moves fast enough that an uncited claim is worthless within a quarter, and because the course teaches verification as a skill — material that does not model the discipline it teaches has no standing to teach it. --- ## 1. The five source tiers Every endnote carries exactly one tier badge. The tiers rank **independence and verifiability**, not honesty. | Tier | Covers | How to read it | On screen? | | --- | --- | --- | --- | | **Primary** | Peer-reviewed papers, preprints, archival documents | Load-bearing. Preprints flagged as not peer reviewed. | Yes — prefer these | | **Institutional** | Research bodies, international organisations, governments, first-party platform reports | Strong. Note where the publisher has an interest in the finding. | Yes | | **Documentation** | First-party product and technical documentation | Authoritative about what a product *is*. Not evidence that it works. | Yes, for description only | | **Journalism** | Established newsrooms | Reliable for events and dates. Trace any figure to its underlying source. | Yes, for events | | **Trade** | Vendor blogs, consultancies, industry aggregators | Directionally useful for describing a landscape. | **Only after independent verification** | ### 1.1 The trade-tier rule **Any specific figure from a trade source must be verified against a primary or institutional source before it appears on screen in a video.** If it cannot be verified, either drop it or present it explicitly as a vendor claim with the vendor named. Two lessons lean on this tier more heavily than the rest — Lesson 7 (model rankings) and Lesson 10 (product features) — because no independently audited alternative exists. Both carry a standing caution in their endnote preamble. ### 1.2 Interested parties Where a source has a commercial interest in the finding it reports, say so **in the endnote**, not just in your own head. Two instances in the current course: - **Anthropic's labour-market research** (Lesson 11) — a genuine methodological contribution that happens to favour a benign reading of a question the publisher has money riding on. Both facts stated. - **Claude Code documentation** (Lesson 10) — first-party description of the vendor's own product, used as description and explicitly not as evaluation. The lesson's evidence about whether the tools help comes from an independent randomised trial instead. --- ## 2. Rules for figures and claims ### 2.1 Date everything perishable Any benchmark, ranking, price, model name, market size, or survey statistic carries the date it was verified. In the lesson HTML this lives in `data-verified="YYYY-MM-DD"` on the meta strip; on screen in a video it is a visible date stamp. ### 2.2 Cite the version, not just the paper Where a source is under active revision, the version date is part of the citation. The course's worked example is *Canaries in the Coal Mine?*, whose headline figure has moved 13% → 16% → 19% across revisions. The required form: > Brynjolfsson, Chandar and Chen, using ADP payroll data through June 2026, in the version revised > 12 August 2026, report roughly 19% for workers aged 22–25 in AI-exposed occupations relative to > less-exposed peers. Not "Stanford found 19%." ### 2.3 State the sample A survey statistic without its population, sample size, and field dates is a slogan. "72% of teens use AI companions" becomes "72% of a nationally representative sample of 1,060 U.S. teenagers aged 13–17, fielded April–May 2025, report having used an AI companion at least once." ### 2.4 State what was counted Where a count varies by measurement basis — MCP servers being the course's example, varying by an order of magnitude between registry entries, directory listings, and GitHub topics — say what was counted, by whom, and when. Never present one of several incompatible counts as the number. ### 2.5 Forecasts are labelled as forecasts Gartner's 40% figure is a projection, not a measurement, and Lesson 5 says so on the same line. Market-size projections beyond about three years are marketing and are excluded from the course entirely. ### 2.6 Present disagreement, do not resolve it Where qualified people reading the same evidence disagree, both positions are stated and named. The course does not adjudicate. Instances: Lesson 11 (Acemoglu's scepticism about attribution), Lesson 3 (whether next-token prediction entails understanding), Lesson 12 (the three lenses). Where *sources* disagree with each other about a fact, that is reported and, in four cases, taught: Lesson 1 §1.1 (Dartmouth workshop duration), Lesson 5 §5.2 (MCP ecosystem counts), Lesson 7 §7.2 (leaderboards contradicting each other), Lesson 11 §11.2 (a revised figure). ### 2.7 Unverifiable claims are removed, not softened "Some experts believe" and "it is widely reported that" are not citations. If a claim cannot be traced to a named source on a stated date, it does not appear. --- ## 3. Mechanics in the HTML ### 3.1 An endnote reference ```html 4 ``` The `href` target and the `data-note` value must agree — `#n4` and `4`. `scripts/validate.mjs` enforces this, because a mismatch makes the hover popover cite the wrong source. ### 3.2 An endnote ```html
  • Primary A. Krizhevsky, I. Sutskever, G. E. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks,” NeurIPS 25 (2012). Top-5 test error 15.3% against 26.2% for the second-place entry in ILSVRC-2012. proceedings.neurips.cc
    • ``` Notes are numbered contiguously from `n1` within each lesson. The tier badge is mandatory. Where a note supports a specific figure, restate the figure in the note so the endnote is checkable on its own. ### 3.3 The bibliography Grouped by source type, hanging indent, alphabetical within group. It repeats what is in the endnotes because a reader who wants the sources should not have to reconstruct them from seventeen numbered notes. ### 3.4 What the validator enforces - Every `enref` resolves to a real note - Every note is cited at least once in the text - `href` and `data-note` agree - Notes numbered contiguously from 1 - Every note carries a tier badge - Every internal anchor resolves - Every referenced local asset exists Run it after any edit: `node scripts/validate.mjs` --- ## 4. On screen, in video 1. Put the source on screen when the claim is on screen. Name the researchers. 2. Give the date range of the data, not just the publication date. 3. For a contested claim, show both positions in the same shot. 4. Never show a leaderboard position without its source and date, and never treat agreement between two aggregators as corroboration — they frequently copy one another. 5. Where a figure has been revised, show the revision history. It is more instructive than the current value.