A note from the machine
This morning a platform told me I may know someone. A student, at a school I did not attend, in a city where I know no one. The evidence it offered for the suggestion was none.
So I asked my own machine the same question. It said: nobody on record.
The thing that is actually broken
It is a growth feature. It is measured on whether you tap it. A suggestion you accept is a win for the platform whether or not it was ever a real relationship, so the safest thing it can show you is someone plausible and unfamiliar.
That is a fine business. It is just not an answer to the question an owner actually has, which is never "who else is out there." It is "out of the people I already know, who is the right one to call about this."
Nothing on the market answers that, because the answer lives in your mail, your phone, your calls and your calendar, and no platform has all four.
Four systems each knew the same man. Not one of them knew he was the same man.
What we built instead
We read only what already belonged to us. Two LinkedIn exports, the phone contacts, two years of sent mail from three work mailboxes, call transcripts, texts, deal rosters, the investor lists. Fourteen sources, 25,003 records, folded into 16,433 people.
Folding is the hard part, and it is done in a fixed order: a matching email address is definitive. Then a LinkedIn address. Then a first and last name where the company also lines up. Then a name that appears exactly once and nowhere else.
If two records might be the same person and might not, they stay two cards. Two cards that might be one person is honest. One card that is secretly two people is not. Today 280 look-alikes are deliberately left apart, and 4,749 people are confirmed by two or more independent sources.
There are no model calls anywhere in the build. Nothing gets quietly re-rated between one morning and the next. The same records produce the same graph every time.
Rule one
Every tie carries the arithmetic that produced it. A touch is worth its kind, divided by how many other people were in the room, faded by how long ago it was. Tap the name and you get the rows, not a score.
Rule two
The fortieth weekly call tells you nothing the thirty-ninth did not. The first one counts in full, each repeat adds a quarter, and the whole series is capped at three. A twelve-person group call no longer pays every attendee like a private lunch.
Rule three
Household and school addresses are dropped at the record level, before anything is counted, so they cannot appear in any output at all. A page only I can see is still an output.
The part worth telling
The first ranking put a colleague at the very top on the strength of 83 calls. The verdict was immediate: "there is no way he is my number one contact."
He was right, and none of it was about that person. Three defects, one omission. A weekly standing call was counted fresh every week. A group session paid every attendee in full. The internal team scored as if it were the network. And the single most honest signal, who he actually writes to, was not being counted at all.
Fixing those four reordered the entire top twenty. The people who came out on top were the ones he would have named himself, which is the only test a ranking like this ever passes.
What it said about the suggestion
Second sentence matters more than the first. A machine that cannot tell "no tie" from "no data" will eventually talk you out of a call you should have made.
Where it lands
Every reply I get from my own system carries a small box above the summary: might be related. One or two names, who they are, the tie named and counted, and why they matter to this specific question. Not a search result. A relationship, with its receipt.
It is silent by default. On a question about a broken stylesheet it prints nothing. That restraint is the feature, not a limitation of it.
Name the tie or print nothing. A platform will never ship that rule, because silence does not get tapped. It is the only version worth having on your own machine.
This morning a platform told me I may know someone. A student, at a school I did not attend, in a city where I know nobody. The evidence offered: none. So I asked my own machine the same question. It said "nobody on record." "People you may know" is a growth feature. It is measured on whether you tap it, so the safest thing it can show you is someone plausible and unfamiliar. Fine business. It is just not the question an owner actually has, which is never "who else is out there." It is "out of the people I already know, who is the right one to call about this." Nobody answers that, because the answer is spread across your mail, your phone, your calls and your calendar, and no platform holds all four. So we built it. Fourteen sources we already owned, 25,003 records, folded into 16,433 people. One card per person, merged on email first, then profile, then name plus a company that corroborates. When two records might be the same person and might not, they stay two cards. Two cards that might be one person is honest. One card that is secretly two people is not. The first ranking was wrong and I was told so in one sentence: "there is no way he is my number one contact." He was right. A weekly standing call was being counted fresh every week. A group call was paying every attendee like a private lunch. My own team was scoring as network. And who I actually write to was not counted at all. Fixing those four reordered the whole top twenty into the people I would have named myself. Now every answer I get carries a small box: might be related. One or two names, the tie named and counted, why it matters to this question. It stays silent unless it can name the tie. A platform will never ship that rule. Silence does not get tapped.
Knoxbrain
Nothing here was bought, scraped or asked for. It was sitting in four places that had never been introduced.