Sepia Files

A patch release with one behavioural fix: the Gemini prose-layer table now covers only the Gemini releases the vendor names.

What you will notice

• A review whose author or executor is Gemini 3.5 or Gemini 3.8 Flash reports that role's prose layer as prior instead of operative. Gemini 3 and 3.1 are unchanged.

Model fingerprints

• The Gemini table was tagged "Gemini 3 series" and said the vendor's "less verbose" statement ran from Gemini 3 Flash through Gemini 3.8 Flash. The current pages, re-read 2026-09-23, do not support that upper bound: the developer guide says "Gemini 3 is less verbose" and lists models up to 3.1, the page for Gemini 3.5 writes "Gemini 3 and 3.1 is less verbose", and the prompting strategies page names neither 3.5 nor 3.8 in its text. The table is now tagged Gemini 3 and 3.1, and later releases in the same numbering read it as a prior (#277).
• The model-identity rule in SKILL.md used Gemini 3 as its example of a series-wide tag. It now says a table tagged with named releases matches those releases, and a later release that shares the numbering but is not named reads the table as a prior (#277).
• The research ledger and the fingerprint coverage row in all three READMEs match (#277).

Not in this release

• A prose-layer row for Gemini 3.5 or 3.8 Flash. Google has published no statement about their default writing style.
• Any change to the measured Gemini 3.8 Flash column in languages/zh.md, which is a corpus measurement, not vendor guidance.

How this was found

An independent search for new vendor prompting guides, run through Grok, found no new guide but surfaced the Gemini 3.5 page's wording. Each of the three pages was then read directly before the change.