Offering a tour, and remembering a company

When a tour is offered

The writer says what shape its answer is: pick (the visitor asked for a piece of work and gets its card) or evidence (the visitor asked about her, and the items are the steps of one argument, each tied to the study it is in).
An evidence answer offers a tour that walks exactly those items, across studies when they are spread. A pick answer offers none, unless several pieces were asked for: then the picks are walked, since a row of cards is a list.
At most three studies are toured; items on a study past that are left out. The way in is one line the writer wrote in the visitor's voice, under sixty characters, or Mò's fixed line.
The plan is made ahead of the click, after the answer goes out, so the tour starts at once; when no plan is ready it is made at the click.

The planner is Mò on the writer's model, given the blocks of the studies by anchor, the question, the answer, the chat so far, the notes, the brief, the memory lines and the standing requests, and the steps when the answer laid them out: one stop each, in order, on the study named.
Three to five stops, a line of at most 150 characters each, a title of at most 60 characters and labels of at most 44, a closing of at most 200. Stops keep the order the page reads and the order the studies were answered in; each anchor once.
Every line and the closing go through the same check an answer does, read on their own against the block they stand by. The harm predicate is read at 0.8 for a tour's lines, since a short line about a cut made to ship was once dropped at 0.62. A plan that fails is asked for again once; what still fails the code is trimmed, what still fails the gate is dropped.
A tour taken counts as one turn of the company's budget: a row of kind tour with its closing line.

After each answered turn, off the request path, a memory agent on gpt-6-sol reads the chat and what is already remembered and proposes facts: add, confirm, revise or retire. The code decides; a proposal that breaks the rules is dropped.
A fact is one line of English, under 160 characters, true of the team. Anchored facts (the role, its scope, the domain, the seniority sought, what the team values) are added at 0.75 or more with the visitor's words quoted, revised only at 0.9 or more, and never retired. Fluid facts (an interest this visit, a question left open, what impressed) are added at 0.5, revised at 0.6, retired at 0.6. Eight of each at most, twelve themes, eight studies.
The brief and the job description the owner wrote are locked: never restated, never contradicted. A fact backed by more than one chat is not moved by one chat saying otherwise; the first contradiction is kept as a challenge until a second chat says the same.
Never a fact: how one visitor wants to be answered, a person, another company, how Mò works, what Mò said or showed, small talk, a guess.
Memory is felt, never announced: the writer is told it shapes what comes first and is never mentioned, since the reader may be a colleague who was never here.