How a turn runs

Every message to Mò goes through the same steps, in code (the portfolio's lib/mo/turn.ts). The rules below are branches and thresholds in that code, restated so a review can point at them. An Edit here is a change request to the code; nothing Mò reads changes until a person makes it.

Before anything is read

The budget comes first. A company has a number of turns in all (40 by default; none for the owner or a guest). Past it, the field closes with the final handoff. From 80% of it, every answer closes with a handoff and her email.
A clicked suggested prompt skips the router: its branch, focus and plan line are authored, and if the model cannot be reached its written answer is shown.
A reply to the card Mò asked back with, or to a chip, is read without a model: the options picked become lenses, the branch is work, and the turn is answered for as many pieces as the card asked about.
Otherwise the router reads the message in one Decisions call and returns the branch with its confidence, the intent, the focus, the approved fit line, the language, how many pieces are asked for (one, several, unspecified), whether the aim is named or open, and which lenses the message shows.

The second look

When the router turns a message away as security, offtopic or private, or takes it for identity, and is not sure, the writer model reads it again and chooses the branch itself. Unsure means a confidence under 50%, or a working branch holding 35% or more of the probability.
Its verdict replaces the router's. An injection, or a question about the machinery, is refused on the router's word.

Turns that need no evidence

A refusal is a fixed line by tier, in the visitor's language. The second refusal in a chat is the shortest line; the third closes the chat. A refusal the gate caused is not counted against the visitor.
Contact is a fixed line with the résumé, the email and LinkedIn.
Identity and handoff go to the writer with no evidence, and get the code check only.

Asking back instead of guessing

A work question with an open aim, from a visitor Mò knows nothing of (no brief, no job description, no memory, no lens from the chat), with no prompt clicked, no study attached and nothing shown yet, is answered with a card asking what they would like to see, when at least two of the lenses on offer still lead to different work.
Mò never asks twice in a row. Chips under a work answer narrow it afterwards: at most three, lenses the chat does not know yet, none with thin evidence, never the domain.

Work, fact, fit and adjacent

A plan line is shown. The brief and the job description are read once per chat for the lenses they show.
The look: one Decisions call scores every case study for the question (0 skip, 1 mention, 2 recommend) and names the one to start with.
The read: the start-here piece, then any other scoring 1.0 or more, two at most; three down to 0.75 when several pieces were asked for; one when one was. What the visitor has seen, shown in the chat or read on the site, is left out while anything else qualifies. A fit or adjacent turn may read nothing. A fact question that names a study by its title reads that study.
Each study read gets one more Decisions call that scores its chapters; the ones at 0.5 or more are cited, three at most, else the likeliest one. Studies scoring 0.75 or more that were not read go in as their intro only.
The writer answers from that evidence, on gpt-6-luna at low effort, in the response form, with the read_chapter tool for chapters it was not given: up to six ids a call, five model turns in all.
The checks: code first, then the model check, then the gate and its appeal (see Before an answer ships). One rewrite with the reasons; a second failure steps aside.
After the answer: the studies it picked are marked shown; a tour is offered when the answer laid out evidence, or when several pieces were asked for; chips are offered after a work answer that did not step aside, was not about one study, and was not the closing handoff.