Rewrite the public site prose in a plain teaching voice

[Cuuper22]

Finishes the documentation rewrite by covering the two files it missed: the public site's own pages. Earlier work rewrote the project's markdown and its JavaScript comments, but the prose a visitor actually reads on index.html and observatory.html was still in the old style.

What changed

Visible text only, across roughly 75 edits. Technical terms now carry a defining clause where they first appear — held-out split, regret, resolver, scope, root input, dependency cone, scenario fixture, HBM, MFU, PUE on the landing page; held-out, collective, preregistered, prior, residual, durable frontier, PDU, WAN on the observatory.

Jargon gave way to plain equivalents where the plain phrasing says the same thing: "epistemic ranking regions" became "ranking under uncertainty", "inadmissible" became "cannot support a claim", "learning-noninferior" became "learned just as well".

The result reads like the rest of the project. For example, the page now explains that "a held-out split is data the model never saw while it was being fitted, so a prediction on that data is a real test," and that "a root input is modeling debt you can see. That beats hidden debt papered over with a fudge factor."

What deliberately did not change

No HTML structure. Not one tag, attribute, class, id, data-* attribute, inline style, or href/src reference.

No numbers. The docs-stats gate parses the <b>NNN</b> cells in the stat grid and compares them against values computed live from the registry, so a single altered digit would both fail CI and publish a false fact.

No string that JavaScript matches on. 28 shared strings on the landing page and 75 on the observatory were identified up front and frozen. That is also why the legend item "Collective" keeps its exact wording — observatory.js matches it — with the term defined in the surrounding prose instead.

Verification

• docs-stats gate prints docs-stats: OK.
• Ordered tag-skeleton comparison: 757 → 757 tags on index.html and 2058 → 2058 on observatory.html, with zero differences. Because the comparison is over the ordered token list, this proves no structural or attribute change anywhere.
• Every numeric literal extracted and compared: 182 → 182 and 576 → 576, all identical.
• Full test suite run locally before merge.

Known follow-up, not addressed here

Seven panels on the landing page have their copy duplicated inside app.js, which overwrites the HTML on load, so those panels render the app.js copy. That prose is already plain and reads well, and three of its sentences are pinned by docs_stats_check.py regexes that require exact wording around the digits, so rewriting it was judged more risk than benefit. Anyone editing app.js later must keep these intact: "The registry currently names N variables and N equations", "N equations are currently covered by unit checks", and "N root inputs are still visible in the current summary".

🤖 Generated with Claude Code