Why most of Istos's checks are plain code, not AI
An agent that guesses is fine for a first draft and bad for a fact. Why Istos uses ordinary rules for anything that can be checked, where it uses a model, and what that means for the answers you see.
Istos is an agent product, so people assume there's a language model behind every screen. There isn't. A lot of what Istos does, and nearly everything it tells you is true, comes from plain code: ordinary rules that read the data and give the same answer every time.
That's on purpose. This post is about where I draw the line between rules and models, and why.
The problem with asking a model for facts
Language models are very good at writing and at weighing fuzzy things. They're not built to say "I don't know". Ask one whether a company sponsors visas and it will often give you a confident answer, sometimes right, sometimes not, and with no easy way to tell which.
For a job search or a sales pipeline, that's the wrong failure. If Istos tells you a role is verified on the company's site, you'll apply. If it tells you an address is safe to email, you'll send. A wrong answer there costs you an evening, or your sending reputation, or your name in someone's inbox. So the things you act on should come from something that can be checked.
What runs on rules
Here's what Istos does with plain code, and what you see because of it.
Your job filters. When you run Find jobs, roles that miss your rules (location, level, remote, salary floor, companies to skip) are dropped before any model reads them. As the Job filters card puts it, filtering costs nothing.
Vetting every role. Whether a role is still on the company's board, its careers page or a government job bank. How fresh it is. The six red flags for fake listings: an apply link off the company's domain, a free mailbox or chat app, money or ID before an interview, pay far above the norm, a brand-new domain, and no official source. How to spot a fake job posting covers each one.
Visa sponsorship. What the posting says, read by rules that look for the no before the yes, and whether the employer is on the official lists for the UK, the US, Canada or Ireland, matched by name.
Your odds. Strong, fair or low, from freshness, location, visa need, level and skills, each reason marked as a plus, a minus or a note. You can open them and see exactly why.
Who you know at a company. The warm path's matches between your imported connections and the company, and the role's "reports to" line, quoted from the posting.
Why you, requirement by requirement. Each thing a posting asks for, next to the line from your own documents that shows it, marked Strong, Partial or Gap. Matched in plain code against what you wrote. If a model refines it, every piece of evidence still has to be a sentence that exists in your documents.
Your do-not-contact list. Checked before a draft is written, before a paid lookup, in the queue and again before sending.
Email addresses. Found and checked by the finder accounts you connect. Istos never builds an address from a name pattern.
Look-alike companies. Similarity between companies is worked out from their fields, stage, place, size and the roles they hire for, with a line saying which features matched. If a model suggests names, rules still have to find each one in Istos's index or its stored news before it's shown.
What the run wasn't sure about. At the end of a run, one sentence on the single biggest doubt, such as boards that didn't answer or roles a rule couldn't judge. It's built from counts, with no model call.
Where a model helps
Rules can't write a good email, and they're clumsy at weighing a role against a brief written in your own words. That's where a model earns its place:
- Scoring a role out of ten against your brief, with a line on why.
- Drafting an application, a first email to a lead, an investor note or a reply, from your own documents and voice rules.
- Reading the unclear cases that rules couldn't place, such as a reply that isn't obviously a yes, a no or a stop.
Even there, the model works inside rules. Drafts are told to use only facts from your knowledge base and never invent experience, employers or numbers. Where your documents have nothing, the honest output is a gap, not a guess. And every draft waits in your queue, because nothing sends without your approval.
- Rules read the sourcesThe company boards on your watchlist and the job banks.
- Rules apply your filtersRoles that miss your rules are dropped. It costs nothing.
- Rules vet each roleOfficial source, freshness, red flags, visa and your odds.
- A model scores and draftsOnly for roles that passed, in your voice, from your documents.
- You decideEvery draft waits in your queue until you approve it.
What you get from rules
The same answer every time. Run the same check twice and you get the same result. That makes the results something you can learn to trust, or argue with.
Reasons you can read. Every flag, badge and odds rating comes with the rule that produced it, in a plain sentence. "The apply link goes through a link shortener, which hides where your application really goes" is something you can check. "Confidence: 0.82" isn't.
No guessing where a fact is missing. When a registry doesn't publish a domain's age, the new-domain check stays quiet rather than guess. When a company's board doesn't answer, the role says Not verified yet instead of pretending either way.
Cheaper runs. Rules cost almost nothing to run, so the expensive model calls are spent only on what's left after filtering. That's part of how Istos keeps runs affordable, and why reading, editing and approving never count as runs at all.
Why you'll never see a percentage
You'll notice Istos talks in words: strong, fair or low; Strong, Partial or Gap; verified or not verified yet. That's the same principle. A percentage would claim a precision nobody has. "Your odds: fair", with its reasons under it, is honest about what's known and what isn't. It's a guide, not a prediction, and the screen says so.
The trade-off
Rules are strict and a little literal. A real role can trip a red flag because a company uses a form builder for a small hiring round. A sponsor can be missed because its name on the job board doesn't look like its name on the register. When that happens, Istos shows you the reason, so you can see the rule misfired and decide for yourself.
I'd rather give you a strict check with its reasons than a smooth answer you can't question. The judgement at the end, whether to apply, whether to send, is yours anyway. How to find roles that actually fit your level shows how these checks look on a real Jobs page.
Let the agents do the searching
Istos is in private beta. Join early access, or try the free preview on your resume, your website or your raise first. Nothing sends until you approve it.