Six layers, two modes, and a record that knows exactly what produced it.
This page is for the technical reader: the evaluator, the computational member of the team, the research office that asks how a session can be trusted. It is the architecture behind what you get. Nothing here changes what a researcher sees; it explains why it holds.
Two are fixed and versioned. Four are composed for your lab, every session.
Most AI features in lab software answer every lab the same way. Methods composes how it listens, what it compares against and what it offers from layers that belong to your lab. Layers 1 and 2 never change between labs; that is what keeps the session honest. Layers 3 to 6 are yours, and they are what your agent's memories write to.
Shared, versioned and tested. One question at a time. Reflect back every few exchanges, in your words. Chase the tacit. Never fill a gap: what you did not state is written as Not stated, and every line of the record traces to something you said.
What to listen for in a bench-to-figure account: the export and first touch, the spreadsheet steps, replicate rules, normalization choices, thresholds set by eye, software settings, what gets re-run; and at the bench, materials, conditions, order, judgment points, controls, failures.
The lens itself. Your field and systems, the instruments and the files that come off them, the software and versions in use, who on the team runs which step, the funder and institutional mandates you work under, the documentation standard you already have or want, what must never be repeated back, and how much guidance you want.
What "complete" means for your method family: open reporting guidelines reused with attribution, proprietary ones cited and paraphrased, and a cross-guideline schema written and expert-reviewed. The interviewer compares what you said against it and asks about what is missing; it never asserts what you should be doing.
Operational skills, each a rule that listens for a pattern in what you say and records an opportunity: a standardized template, a lab-wide documentation standard, automated intake for every technician's results, a script that regenerates the figure, an existing tool that already covers a by-eye step, a fee-free vendor comparison. By default the offers appear at sign-off, tied to the step that prompted them; if you asked for suggestions in the chat, they arrive as you go, one at a time and labeled.
Elicitation by default. Brainstorming only when you open it, and the agent says up front that it is not the scientific expert. At the start of every session you choose whether suggestions appear in the chat as you go or only in the documentation at the end.
Elicitation, by default
The session. It captures, it inventories the parts of your workflow as you describe them, and it produces the record you sign off. Suggestions wait for sign-off unless you asked for them in the chat.
Brainstorming, when you open it
Thinking out loud about hypotheses and next studies. The agent states first that it is not a scientific expert and that you are the authority. It produces notes, never a protocol.
Guidance level, set by you
Detailed: it explains in a clause why it is asking, and the record carries a "why this matters" line per variance point. Minimal: the record only.
Every record knows what produced it
Each session stores which version of each layer was in use. If a line in a protocol is ever questioned, the exact prompts that ran can be read, not reconstructed.
The lens shapes listening, never asserting
A lab profile changes what the agent asks about and what it offers. It never lets the agent fill a gap with what a lab like yours "usually" does. Not stated stays Not stated.
Sign-off is the gate
Nothing exports, and no offer becomes work, until you approve it. See what stays yours for ownership, retention and what happens when something is wrong.