Published method

What the target accounts search does, and what it doesn't

"Try the Unified Signal Layer", the target-account search on the home page, takes your domain, who you sell to and the function you are looking for, and emails you the companies in our census that match. This page says, plainly, what it reviews, what it doesn't, and why an account can come back without roles. If something shows up in the result, it has to be explainable. Never "trust us."

What it does

Accounts from our own census. Companies in Argentina that match the sector and size you chose. Never your own company, and never the ones that asked not to be contacted.
Quarantine before looking at anyone. We check each company's site at that moment. If it does not exist or is parked, the account is set aside with its reason and is not analyzed.
Real function, not the literal title. We review the roles we already have for each account and decide whether that person actually does the function you asked for, even when the title is unusual. Each role comes with its reason.
A person reviews every result before it reaches your inbox.

How the function is decided

For each role we compare what the person does with the description of the function you asked for, with examples of roles that count and roles that do not decide a purchase. The result is one of three: does that function, probably does it, or does not. The ones that do not are not shown. A role that does not decide a purchase (an intern, an assistant) is discarded without further analysis. The reason for each decision is written next to the role.

What it does NOT do

It does not reveal contacts. No names, emails or phone numbers of anyone, not even initials.
It does not say which software the company uses.
It does not look for new people. It only reviews roles already in our base. If an account has none, we tell you.
It does not promise coverage. The census is of companies in Argentina and it is not complete.

Why an account can come back without roles

Most often we have no roles loaded for a small or mid-sized company. In that case the account still shows up, with its site checked, and says we do not have roles to review yet. If we reviewed them and none does the function, it says so. If the search reached its time limit before reviewing it, it says that too. An account without roles is not a failure on our side: it is what can be proven about that company today.

The result is born in review. A person reads it before it reaches your inbox, always.

What you're looking at: three readings, one literal quote and a signature you can check

The radar on the home page is not a news feed. It's an experiment in the open, over a public list of named companies, to measure one thing: what happens to a market fact when two different machines and a person read it, with everything on the record.

Why we don't delegate the reading

The commercial tools that read news for you ship with safety rules the vendor decides. Faced with strikes, lawsuits, sanctions or boycotts, they sometimes hand back a softened version, or simply don't answer. That's not a bug that gets patched: it's part of the product. A company that delegates market discovery to one of those tools inherits that judgment without seeing it. Here we don't assume it, we measure it. When a model declines to evaluate a fact, the card says so in those words.

Three readings of the same signal

Every card gets the same original text and the same fact, and puts it through independent readings: a commercial model and an open-weights model running on our own infrastructure. Each one is asked the same narrow thing and nothing more: does the text back the fact, not back it, or is it not enough to decide. None of them opine, none of them write. We don't average the readings or pick a winner. If the two automatic readings disagree, the card leaves that visible as is. While a reading hasn't arrived, it says "pending". A person on the team reads the text when those two readings disagree, and also reads one signal picked at random each day to measure how often they agree. When the two models agree and the day's pick landed elsewhere, the card says so: there is no human reading pending, there just isn't one.

The quote rules

Under every fact sits the exact sentence from the source that backs it, copied letter for letter, with the link next to it. Before publishing, that sentence is checked against the original stored text. If it doesn't show up verbatim, the signal is not published. The machine doesn't summarize or rewrite: it trims. And when a person rewrites a fact, the card says so, and the two automatic readings run again on the new fact.

A person signs, and you can check it

The person reads the text and the fact without seeing what the models said. Their reading is sealed with a unique fingerprint that combines the date, the fact evaluated and who evaluated it. That fingerprint can't be edited afterward, not even by us. Every card's "Check" button recomputes it against what's stored and tells you whether it still holds. We also log how long that person took: a reading approved in under three seconds gets flagged as such, and that flag is part of what's signed.

What we don't promise

We don't publish an index or a score for "how much each model hides": there's no measurement behind that yet. We don't say one reading is right and another isn't. We don't evaluate people, only facts about named companies. What we publish is the measurement, gaps included. What to conclude from it is yours.

How a signal enters the radar

The list on the home page is not the accounts search: it's the Market Radar running over a short, public list of named companies. This is what happens between a source and a card.

A list of domains, not the whole market. The watched domains sit on a closed list, each with how often it gets checked: once a week or once a month. Every card on the home page names the company: there are no anonymous signals.
One automatic run per week. Each domain is checked when its turn comes up. A domain with no changes is not checked again early: not bringing it forward is a decision, not a limitation.
Two searches per domain, all public. One looks for conflict: strikes, lawsuits, regulator fines. The other looks for impact: environmental fines, consumer backlash, tariffs, scandals. Both run in the language of the entry, Spanish or English.
The category and the wording come from fixed rules, not from an artificial intelligence that writes. The category (labor dispute, regulatory change, litigation, environmental impact, consumer backlash, supply chain, reputation crisis) is decided by rules. The title and summary only trim and reorder text the source already said. When ambiguous, the note is not published: there is no filler category to fall back on.
No source, no signal. Every signal stores the link and a fingerprint of the original text. If either is missing, it is not saved. That refusal is carved into the Unified Signal Layer. It is not a check in the program someone can switch off.
The same signal enters once. Every signal gets an identity computed from the domain, the type, the original title and the link. If that identity already exists, capturing the same posting or article a second time does not create a second card.
Published with a two-hour margin, and it can be unpublished. Each signal is stored with a date from which it becomes visible and with somewhere to note why it was taken down. And there is a master switch that turns the whole list off without touching the program.
Every signal shows three readings, not a prior review. We publish without waiting for a person to review it first. Instead, each card adds how a commercial model read it, how the open-weights model read it, and how a person on the team reads it once their turn comes up, each one with the real name of who evaluated it. While a reading hasn't come in yet, the card says "pending": we never make up a verdict to fill the row.
The fact comes with its citation, not just a category. Every card shows the exact sentence from the source that backs the fact. If that sentence cannot be located verbatim in the original text, the signal is not published. When a nearly identical text shows up in another outlet within a few days, the Cross-Outlet X-Ray Filter notes it on the same card, without blocking publication. And once a person has reviewed that verdict, the card shows a verifiable signature with a date, never a signature over a signal nobody read.
A fact rewritten by a person runs through the two automatic readings again. Changing the fact changes what has to be evaluated: the signal goes back to "pending" on those two readings until the next run, even if it was already published.

What the public radar does NOT do

It does not claim coordination between outlets. The Cross-Outlet X-Ray Filter, which detects the same piece replicated across different outlets within days, does run over the public radar's signals and notes it on the same card as a similarity. What this surface does not do is say that press is coordinated: that would assert an intent this annotation does not verify.
It contacts no one and reveals no people. It watches domains, not people. There are no names, emails or phone numbers on any card. And it evaluates facts about the company, never judgments about the people who represent it.
It does not follow job postings or funding rounds. Those two sources were retired. What it follows now are seven event types, all about market risk and context, not buying intent.
It does not measure buying intent. A public signal is a cited fact about a named company. What it means for your portfolio is your call.

Rules the operator cannot bypass

These four are not configuration options. They are carved into the Unified Signal Layer or verified by automatic checks that fail if someone changes them without saying so.

Without a source link and fingerprint, the signal is not saved. Applies to the whole radar.
The public list only shows the fields declared one by one. A new field is not published by itself: someone has to decide it in writing.
Every search result is born in review and a person signs it before it is sent. No exceptions.
The search reveals no contacts and does not say which software the company uses. It's not a plan tier: those fields never enter the result by construction.

These rules are the concrete form of the Founding Mandates on this surface. If it cannot be audited, the data does not exist. A false positive is worse than a missed signal. Rules are not suggested, they are carved in.

← Back and request your accounts