Method
Everything here is the method behind every edition. Where a step is done by code, we say so. Where models are involved, we say how many had to agree.
We fetch each tracked document every day: public pages only, logged out, at a polite rate, with an honest user agent. We do not log in and we do not get around bot walls. When a site blocks automated clients we use its sitemap date or Internet Archive copies and mark the source degraded. Every capture is stored as fetched, with its SHA-256 hash, and never edited.
History goes back to January 2024 where the Internet Archive kept copies. As of 19 Oct 2026 we hold 3,520 captures of 59 documents.
We compare consecutive versions of a document line by line. A change is dated to the window between the last capture that still showed the old text and the first that showed the new text, for example “between 13 and 15 Oct”. We never pick a single day inside that window. When the new text states an effective date, notice days run from the window to that date, as a range. If the effective date falls on or before the window, the change took effect without notice we could observe.
For each vendor and plan tier, models from four labs (Anthropic, OpenAI, xAI and Google) read the documents that govern that plan and answer each question independently, with a supporting quote. A value is accepted when at least three agree and an agreeing quote is found word for word in our capture. The quote you see is cut from the capture, not copied from a model. “Not stated” is a value, and it names the documents searched.
Cells that do not reach agreement stay “in review” and keep their last accepted value, or are settled by a documented review call that gives the reason and a quote checked against the capture. A review call stops applying when a document it rests on changes. In the 19 Oct 2026 matrix, 272 of 289 cells were accepted by agreement (94%), 8 were settled by review calls and 9 were in review.
Notice days, retirement lead times, changes per vendor per quarter and days since the last change are computed by code from the captures, and each one carries its derivation. Models write the sentences around the numbers. They never produce the numbers. Rebuilding an edition from the same captures gives the same tables, byte for byte.
Every edition states the documents tracked, which ones missed a capture that week and which sources are degraded. When we get something wrong, the next edition lists it as a correction, apart from vendor changes, and it goes on the corrections page. If a cited fact does not reproduce against its source, we refund the purchase under our refund policy.
Vendor documents belong to the vendors. We quote short excerpts, 30 words or fewer, with attribution and a link to the source. We never republish whole documents.
Materiality rubric
Material: a change a privacy, procurement or vendor-risk team should act on. It touches what the vendor does with customer data, how long it keeps it, who sees it, who owns what, which law applies, how much notice the vendor gives, or when a model is retired. Notable: a real change in wording that does not touch those topics, such as new examples in a usage policy or a new model added to a lifecycle table. Editorial: formatting, navigation, page furniture, dates in headers and footers, clause renumbering, small wording fixes. Recorded, never reported.
The rules are applied by code, in this order, and every change in our data records the rule that graded it. Only changes the rules cannot call go to a vote.
| Rule | Grade | Fires when |
|---|---|---|
| E1 | Editorial | Only line order, line breaks, case, punctuation or typographic quotes changed, or the same words appear in a new order (a menu or list reordered). |
| E3 | Editorial | The only changed lines are date stamps: "last updated", "effective date", copyright lines, "3 days ago". |
| E4 | Editorial | The only changed lines are navigation or link labels: six words or fewer with no sentence punctuation, digits or material terms, or a standalone question that links to another article. |
| E5 | Editorial | The only changed lines are page chrome: feedback prompts, cookie banners, "skip to content", product announcement banners in a page’s opening text. |
| E2 | Editorial | Only small connecting words changed (articles, prepositions, pronouns). |
| E7 | Editorial | Only clause numbers or cross-references changed ("A.1.", "(b)", "Section 4"). |
| M1 | Material | The changed words themselves concern a material topic (list below). |
| M2 | Material | A number or date was replaced or removed inside a passage about a material topic ("30 days" became "60 days" in a retention sentence). |
| M3 | Material | In a model lifecycle document, an existing date was replaced. |
| A1 | Vote | A material topic is in the changed passage, but none of the changed words is a material term. Four models grade it independently. |
| N3 | Notable | In a model lifecycle document, rows with dates were added and nothing existing changed. |
| N2 | Notable | Usage policy wording changed by five or more substantive words. |
| N1 | Notable | Five or more substantive words changed outside material topics. |
| E6 | Editorial | Fewer than five substantive words changed outside material topics. |
Matched at the start of a word, ignoring case. Examples of the terms for each topic:
Each A1 change goes to four models, one per lab, with the old and new text and this rubric. Each answers on its own with a grade and a one-sentence reason. A grade is accepted when a majority of the models that answered agree, with at least three agreeing when all four answer. Otherwise the change is graded Notable and goes to review.
Before a terms or policy revision from the last 90 days is printed as Material, the four models grade the revision as a whole. A majority can confirm or downgrade it; without a majority the rule’s grade stands. Model lifecycle pages are not re-graded; their dates feed the retirement calendar.
Some vendors serve a different edition of a page by visitor region. We track the edition our own capture receives, so a regional switch is never reported as a change. Bot checks and error pages are kept in the raw archive but never become a version.
Coverage
As of 19 Oct 2026. Degraded sources: none. The date is our first capture of each document, counting Internet Archive copies.