Different AI helpers are good at different things. Perch learns when a cheaper or faster helper is still good enough — and it never picks a bad answer just because it costs less. Here is how that works, without the jargon.
You could ask any of them for help. They are all good — just good at different things, at different speeds, for different amounts of pocket money.
Careful and usually gets difficult work right. The safe choice when the job really matters.
Quick, and uses less pocket money. Great for easy jobs — risky for the hard ones.
Notices mistakes Sam sometimes misses — but writes very, very long answers, and long answers cost money too.
Before choosing a helper, Perch asks one question:
If Sam is your chosen helper, then Sam's level of work is the grade. Perch may hand a job to Terry only when there is good evidence that Terry still meets Sam's grade for this kind of work. No evidence? Perch sticks with Sam. Simple.
Terry has done this kind of job well many times before → Perch may pick Terry and save money.
Nobody knows if Terry can do this kind of job → Perch stays with Sam. Saving money never beats meeting the grade.
This idea comes from a company called Thesean, whose “Ship” service lets you name a reference AI and promises a cheaper way to reach the same level. Perch borrows the idea of the grade and the careful measuring — it does not use Ship or pretend to know Ship's secret recipe.
Say Sam has already read your whole 100-page project notebook. Sam remembers it, so the next question is quick and cheap to answer.
Switch to Terry, and Terry has to read all 100 pages from the beginning — even though Terry's price per page is lower, paying for a whole new read can cost more than just asking Sam. AI helpers have the same thing: it's called a cache, and it works like a warm notebook that only one helper holds at a time.
Pretend numbers, real lesson: Terry's lower price per page loses to a 100-page re-read. Perch does this maths before every switch.
So before choosing, Perch counts everything the job will really cost — not just the sticker price:
Some “cheap” helpers write so much that they end up expensive. Some “expensive” helpers finish in one short answer. Only the whole job tells the truth.
The app on your own computer. It picks helpers, does the notebook maths, and keeps a small local scorebook of how each job went.
Keeps: passwords & keys in the Mac's vault — they never leave.
If you choose to help, your Perch sends safe totals only — like “this helper finished this kind of job 42 times.” Never your questions, answers, or names. It sends back signed suggestion cards.
Never sees: your homework. Never answers a question itself.
A Perch you can call over the internet. It must briefly see your request to answer it, passes it straight to the chosen AI, and doesn't save it by default. Your keys sit in a strong locked box with your own lock.
Different: doors, locks & workers from the Scoreboard.
For an important plan, Perch can deliberately spend a little more: ask two helpers separately, then ask a judge to compare their answers — where they agree, where they disagree, and what both of them missed.
Perch shows you who it wants to ask and what it might cost — before asking.
If one helper fails, Perch says so instead of hiding it.
If the judge fails, Perch shows the separate answers instead of faking a combined one.
There's a cheaper twin trick too: ask the same helper twice and keep the better answer. Helpers don't answer identically every time — so two tries from one helper can smooth out an unlucky answer, without paying for a second helper or losing the warm notebook. Perch tests both tricks against just asking the one best helper — the team only wins if it beats them after counting every extra cost.
Perch doesn't experiment on your important work. It starts as an adviser — “here's what I would pick, and why” — and only makes choices by itself at safe moments, like the start of a brand-new chat. For every choice, it can show its working: the grade required, the evidence, the full expected cost, and the way back.
Evidence is fresh and the grade is met → Perch may pick the cheaper helper.
Evidence is old, thin, or a price is unknown → Perch stops being clever and stays put.
Anything looks worse than the grade → straight back to the safe helper. No arguing.
Before Perch trusts a helper, a team of helpers, or a new picking rule, it gets tested in a huge practice lab — hundreds of real coding tasks, each in its own clean room, where the work is checked automatically: the code either passes its tests or it doesn't. No opinions, just results.
The lab (built on open tools like Martian's ARES) lives completely outside your Mac. It never sees your projects or your keys — it only sends back the scores. Only ideas that win in the lab, on problems they've never seen before, get suggested in the app.
Whatever it's doing, Perch decides in this order:
Protect your secrets and obey your rules.Always. Before anything else.
Meet the required grade.A cheap wrong answer is not a bargain.
Count the whole job.Long answers, re-reads, retries, and lost warm notebooks included.
Prefer the simple warm route.Switch only when the new choice clearly wins.
Explain the choice — and keep a way back.Every pick can show its working, and undo is one step away.