Sorting a prospect list by headcount tells you who can pay. It tells you nothing about whether your problem lives there. This reads each company's footprint, asks four narrow questions, and says how sure it is about every answer.
| Company | What's at stake | Has the team | Already solved | |
|---|---|---|---|---|
| 1 | Bizom retail intelligence platform, ai agents for business operations |
3.12 | 3.86 | 1.00 |
| 2 | DoubleTick whatsapp business api platform, ai chatbots, crm integrations |
3.07 | 3.94 | 1.02 |
| 3 | FieldEZ field service management software · hiring: Sales Development Representative |
2.99 | 2.01 | 1.97 |
| 4 | Script All DNA Technologies healthcare technology, radiology, pacs, vna, ehr |
2.32 | 3.40 | 0.39 |
The problem
Same industry, same headcount, same region, same job titles. One handles the problem in-house and is perfectly content. The other is drowning in it. Firmographics describe what a company is. The problem comes from what it's doing.
How it works
Claude reads your business and writes four questions, once. Jev answers them for every company on your list — returning typed scores and probabilities rather than prose, which is why it costs pennies instead of hundreds.
01
Or give it a domain and it reads their site, pricing and review pages and drafts it for you. Anything it inferred is flagged for you to correct.
02
Four questions written for your business, in a file you can read and edit. A linter rejects any question the data can't actually answer.
03
One call per company, four judgments in parallel, each with a probability distribution and a confidence. Two thousand in four minutes.
business context ──► Claude ──► rubric pack ──┐
(what you sell, (four questions, │
who's already solved it) yours to edit) │
▼
companies.csv ──► state builder ──► 12 fields ──► Jev
bucket the numbers, │
drop the noise │
▼
score · probability · confidence
│
▼
composition — plain code, no model
│
┌──────────────────┬───────────────┬─────────┴────────┬──────────────────┐
▼ ▼ ▼ ▼
work these check these ruled out can't judgeThe judgment
Ask a model to rate an account out of ten and it silently averages four unrelated things. Split them and you can see which one disagrees with you.
Is the team, process or system your problem needs even present? If not, nothing else matters — however attractive the company looks.
Given it could exist here, how much does it cost them when it bites? This is what orders your list.
If they already run a product that does what you do, you're late. An adjacent tool is not a solution — getting that distinction wrong throws out your best accounts.
A sparse record gets set aside, not quietly scored as a zero and buried.
The part nobody else shows you
Bar length is the score. The solid part is how much of it the model stands behind. Two accounts can share a number and mean completely different things — and in every other tool they look identical.
Click any row in the app and it opens: the level it landed on in the rubric's own words, the full probability distribution, and the raw evidence it read.
The output
A missing field is not a zero. Most tools treat it as one, which quietly pushes companies with thin data to the bottom — and thin data correlates with nothing useful. Separating we looked and the answer is no from we couldn't see is the difference between a ranking you can trust and one that's deleting good accounts.
Cleared every threshold, ordered by what's at stake. This is your week.
Cleared, but the model isn't confident in the ranking. Twenty seconds each.
Excluded, with the reason on every row. Read a few — it's the most informative list.
Not enough in the record to say anything honest. Set aside, not buried.
Measured, not estimated
Which matters, because you will rewrite it. The first rubric is always wrong somewhere, and the only way to find out is to run it and read the excluded list.
What happens to your list
Sent
What the company does, its description, staff by department and seniority, the tools on its site, roles it is hiring for, and a few supporting fields. These go to TypeSafe to be graded.
Never sent
Company emails and phone numbers are in a typical export and are never included in what the model sees. Neither is revenue or any financial column — those measure who can pay, not whether the problem exists.
Kept
Your uploaded file is swept after thirty days, and only your three most recent lists are held at any time. Your results live in your own browser, not on the server — clearing site data removes them.
We keep a usage log, and the maker can read your graded results. That means the company names on your list and the scores the model gave them — the same view you see. It exists so the rubrics can be improved by looking at how they actually performed, rather than guessing.
What is not readable: your uploaded file itself, and the email and phone columns in it. Those are never sent to the model and never stored alongside a result. If your list is confidential, use a sample or a list you are comfortable sharing.
If you use Research a company, the business description is written by Claude from that company's public pages. Your uploaded list is never part of that, and never reaches Anthropic.
Being straight with you
Every one of these is a real gap, not a roadmap item dressed up as one.
Timing comes from events — funding, a new hire, a leadership change. Those are barely present in a static export.
You get a ranked list and the evidence behind each grade, not a line that proves you looked at them specifically.
If your export was already filtered, those filters are baked in. Widening who's eligible means pulling a new list, not re-scoring this one.
Run thirty accounts you closed through the same rubric. If your best customers don't land near the top, it's measuring who could pay. Do this before you trust it.
Point it at a business and a company list. Ten cents and four minutes later you'll know which accounts have your problem — and which ones it couldn't tell.