What Mò reads, and how it picks

The corpus

Each case study is its registry details (role, duration, collaborators, headline, reported outcome, thesis) and its chunks: the intro, then one chunk per section of its page, each with a one-sentence brief. The ids are meteor#intro, meteor#3.
The owner's notes in What Mò also knows go with their study when the heading is a study's slug, and to every turn otherwise (site, stories, background, now, process, research, technical). Mò may state them as fact.
A study can carry a caution, a line from the owner that the writer and the check both read. Engram's figures are under one: point at the case study instead of quoting them.
The corpus is generated at build time; a stale copy stops the dev server. Mò never reads a file at runtime.

The look across every study

The look is one Decisions call given the question, the intent, the focus, the owner's notes for the company, what has been shown, the curation notes, the brief and the memory lines, the lenses the chat knows with where the site shows each, and what the visitor has read. It returns the one piece to start with, with its confidence, and a relevance score per study.

What is read in full

With a study attached to the chat, that one.
Else the start-here piece, then the next best at 1.0 or more, two at most. Asked for one piece, one. Asked for several, three, down to 0.75, so a tour has more than one study to walk.
Pieces the visitor has seen are avoided while anything else qualifies. The look is told so, and has still ranked a read study first; code sets it aside.
A work or fact question always reads its best piece, even under the bar, so the writer has something to point at. A fit or adjacent question reads one only when the look named it.

Chapters

One Decisions call per study read scores its chapters. At 0.5 or more they are cited, three at most, else the likeliest one alone. The writer sees the other chapters by id and title, and may open them.

The read_chapter tool

The writer's one tool: open chapters it was not given, by id, up to six in one call. It is told to use it only when what it has does not answer the question.
Five model turns in all: the model, a call, the model, one more call, the model. An evidence answer that opened two studies' chapters one by one once failed the whole turn, so the tool takes several ids at once.
What the writer opened is read by the check too, so a fact from an opened chapter is grounded. The turn records which chapters were opened: a turn that had to open chapters is a turn the look did not serve well.

Testimonials and the approved line

On a fit question, or a work question focused on working with engineers, the testimonials go in as evidence.
On a fit question the approved line the router named goes to the writer, and the check reads it as grounded.