Reference
How the assistant works
Enough of the shape of it to calibrate your trust — what it is genuinely good at, and the specific ways it fails.
It is not a search engine, and not an oracle
The useful mental model is a capable colleague who has read an enormous amount, has your sequence files open, and can run the analysis tools — but who is overconfident, and who will produce a fluent answer whether or not they know.
That last part is the whole reason the product is shaped the way it is. It does not ask you to trust prose. It reads the actual bases, runs deterministic analyses against them, and shows you the result — the translation, the cut positions, the diff, the verification report. The prose is a summary of those; the artefacts are the evidence.
What it is good at
- Reading a construct you did not build
- Working out what is actually on a plasmid from its sequence rather than its labels, and telling you when the two disagree.
- The bookkeeping of cloning
- Frames, junctions, cut sites, orientation, spacing. Exactly the arithmetic that is easy to get wrong at 6pm.
- Knowing what a design needs
- That a chromoprotein needs a cofactor its host may not make; that a disulfide-bonded protein will not fold in a reducing cytoplasm; that a strong promoter may be the wrong choice.
- Finding the parts
- Searching the repositories and bringing back records with accessions, rather than recalling a sequence from memory.
- Writing up
- Turning a conversation into a protocol or a notebook entry.
Where it goes wrong
These are not hypothetical. They are the failure modes worth knowing on sight.
| Failure | What it looks like | How to catch it |
|---|---|---|
| Confident fabrication | A plausible accession, a citation, a variant name — that does not exist. | Open every identifier. If it does not resolve, it is not real. |
| The nearly-right part | The right gene from the wrong organism, or the wild type where you asked for a variant. | Check the record it fetched, not the name it used. |
| Miscounting bases | Claims about positions or codons made from reading a sequence rather than computing it. | Ask for the analysis rather than the assertion — a translation, a site search, a diff. |
| Answering the question you asked | You asked for expression, you got expression; nobody mentioned that the protein needs a cofactor to do anything. | State the functional goal, not just the cloning goal. |
| Agreeing with you | Push back on a correct answer and it may fold. | Ask it to justify the original answer instead of asserting the opposite. |
Why it insists on sources
You will notice it fetching parts from repositories rather than writing out a sequence it “knows”, and attaching accessions to what it brings back. That is deliberate: a recalled sequence is indistinguishable from an invented one, and a fetched record is something you can open and check.
When it cannot find a source for something, the right answer is to say so — not to fill the gap. If you catch it asserting a sequence with no stated origin, treat that as the finding. Where the data comes from.
The model behind it
What actually changes for you
Requests that need more reasoning take longer and cost more of your usage allowance than simple ones. A long design conversation with many steps consumes considerably more than a single question. Usage limits.
A conversation also has a finite working memory. A very long thread that has wandered across several constructs will start to lose the earlier detail — which is the practical reason to start a new thread when the topic changes rather than continuing forever in one.
What happens to your sequences
Your sequences are not used to train models. They are sent to the model provider only as needed to answer your request, under commercial terms that prohibit training on them, and your project data stays yours.
If your work is under an agreement with specific data-handling requirements, check them against our privacy policy before uploading, and get in touch if you need something in writing.
Calibrating how much to check
Scale the checking to what a mistake would cost.
| What you are doing | How much to verify |
|---|---|
| Getting oriented on an unfamiliar plasmid | Read it, spot-check anything surprising. |
| Choosing between cloning strategies | Confirm the cut sites yourself; the strategy follows from them. |
| Building a construct you will transform | Read the diff, translate the coding sequence, run verification. |
| Ordering synthetic DNA | All of the above, plus personally confirming every item verification could not establish. |
The last row is the one that matters. Everything before synthesis is recoverable in an afternoon.