Second Nature, by 29Court
You be the boss. We train the staff.
29Court is the consigliere for your AI staff. It takes care of things, keeps the riff-raff out, and learns exactly how you want things done. We run your rules live, and then we bake them into your own fine-tuned model, so behaving your way stops being an instruction and becomes its second nature.
Two ways your rules get obeyed
One standard of conduct, delivered twice, because each way does the half the other cannot. The rules keep your bot safe tonight. The fine-tuning is how you stop having to supervise it.
Run live, tonight
Your bot goes to Cotillion. Every correction you make, in ordinary conversation, becomes a rule on your own rules page: visible, editable, yours. The rules are enforced in code before work reaches you and after it comes back. Not remembered. Not vibes. Enforced, on any assistant you use, including the one you already have.
Baked in, for good
When your corrections pile up, we render them into lessons, before and after side by side, and fine-tune your own model with them. The result is a LoRA: a small set of weights fitted to one wearer, encoding your defaults, your voice, your judgment about what happens first and what never happens unprompted. Served only behind your private key. Yours alone.
Watch a correction land
Pick a complaint. On the left, what an untrained bot does with it. On the right, the same bot after its fitting.
A demonstration, with invented examples. The mechanism is the product: a correction goes in, and behavior comes out.
The fitting: how your bot gets trained up
Five steps, and your only job is the first one.
You correct
In ordinary conversation, the way you already do. Shorter. Never on a Friday. The link goes on its own line. That is the entire job, and the pickier you are, the better it works.
We keep every correction
Your Correction Ledger holds each one in your own words, readable by you at any time. Every entry is already a rule your bot obeys tonight.
We make the pairs
Behind the scenes, each correction becomes a lesson with the before and the after side by side. If two of your corrections disagree, they come back to you exactly once, as one decision. Never as homework.
We bake
The lessons train your private LoRA. Then it sits an exam: your own corrections, held back from training, that it has never seen. If the fitted bot does not land closer to your judgment than the untrained one, it does not ship.
You read the Fitting Report
Plain words, no jargon: how much closer your bot lands to your standards than it did before the fitting, graded on your own corrections. The report is the proof. There is no other kind.
Be as picky as you like. It helps.
Only you can tell your staff exactly how you want things. Every other AI product treats an exacting boss as a cost. Here, exacting is the input. Each correction does three jobs at once: it fixes tonight's work, it sharpens every future draft, and it raises the standard on the next exam. The client who corrects most ends up with measurably the best bot, and can watch the measurement. Difficult to please is the whole point.
Booked easy
Tickets, reservations, the car at the curb. Settled and sorted, in your calendar, with the details already fussed over.
They fuss
The follow-ups get chased, the drafts get checked against your rules, the small print gets read. Fussing is the staff's job description.
You decide
What reaches you is a decision, not a project: approve, defer, comment. Casual decisions, made like the boss you are. This is what AI is all about.
Your corrections train your bot. Only yours.
- Your corpus trains your model only. Never pooled, never blended, never used to improve anyone else's bot, and never used for our own marketing or research.
- Your work runs on domestic models by default. No foreign or Chinese model touches it, unless you choose otherwise.
- Want the cheap seats for work that does not matter? Say the word and we will route that work to the inexpensive global models instead. Your call, made knowingly, and changed back any time.
- You can read every lesson we trained on. Nothing about your fitting is a black box.
- Leave whenever you like and take it all with you: the model file, the lessons, the rules, the ledger. Switch assistants, keep your tuning.
- On your word, we delete the lot, backups included, and confirm it in writing.
Fair questions
Does this replace the assistant I already use?
No. The rules run on the assistant you have, tonight, whichever one it is. The fine-tuned model serves your own bots and staff surfaces through your private key. Most clients keep their favorite assistant and add both layers underneath it.
What do you actually need from me?
Corrections. The ones you already give, in your own words, as they occur to you. A short sitting now and then to settle any that contradict each other. Nothing else: no forms about your preferences, no homework.
How is this different from a settings page or a long prompt?
A prompt is a request a model can drift away from. Rules here are enforced in code, and a fitted model does not drift from its defaults, because the defaults are in its weights. The exam on your own held-back corrections is how we prove it, in writing, every fitting.
Start with one correction
Tell us one thing your bot keeps getting wrong. That sentence is the first entry in your Correction Ledger, and the conversation starts there.
Got it.
Your correction is with us, and the reply will come back in writing to your email.