Most commercial intelligence tools rest on a fragile premise: they rent someone else’s judgment. They send the news to a closed commercial model, get a summary back and fire an alert.
That has two problems. The first is that it treats reading the market as a part you buy and plug in, when it is really a capability you grow, measure and correct. The second is that it inherits someone else’s decisions. Commercial models are tuned to be cautious, and that caution can soften exactly what matters: a labor dispute, a lawsuit, a reputation crisis.
The REVsign Market Radar adds a new source: an open record of the world’s press. It reads it from the opposite angle: an environment we control end to end, open-weights models measured every day and a person who signs before anything becomes a signal.
Reading is grown, not assembled
If the model doing the reading changes without notice, the yardstick for reality changes with it. That is why the exact version of the model is pinned, and if it is not pinned the radar does not start. The instructions it receives do not change from one day to the next either.
The daily reader is an open-weights model running on a dedicated machine. It switches on once a day, reads everything that came in and switches off.
No model is the best in advance. Every day, a fixed sample of 100 articles is also read by two other open-weights models and two commercial models. Each reading is compared against a person’s signature. We measure whether the category matches, how well it separates what matters from what does not, whether the quote is exact, whether it finds the right organizations, how long it takes and what it costs. The other two open-weights models only read when their machine is already on, so the comparison never switches a machine on just for itself.
The hypothesis we measure. A model tuned to be agreeable tends to soften a strike or blur who is responsible in a lawsuit. An open-weights model, without those commercial layers, should return the role as it is: who is striking and who is affected. We do not take it for granted. That is exactly what the daily comparison against the human signature will confirm or refute.
The six mandates, in the radar
Each of the six mandates becomes a concrete rule of the radar.
Function Dictates Reality, Not the Label
The radar does not use the events the source already summarizes. Every 15 minutes it reads the record of each article, with its themes, places and organizations, and checks it against seven kinds of fact: labor dispute, regulatory change, litigation, environment and corporate governance, consumer boycott, supply chain and reputation crisis. A company enters the radar because a fact places it there, in Argentina or on the list of companies we follow, not because of how it presents itself on its website.
Absolute Operational Silence
A false positive is worse than a missed signal, so the radar discards a lot before it reads. Social networks are left out at the source, and so are articles where noise weighs more than the fact. The full text of each article is fetched by an extraction service, with a cap of 1,000 articles a day. That cap exists twice, in spend control and in the queue record: if one fails, the other stops it. And the reading has nowhere to store a person’s name. The radar follows companies and organizations, never individuals.
Non-Negotiable Traceability
Every reading carries a quote from the article. Before saving it, the system looks for that quote in the original text, after evening out capitals, accents, quotation marks and spaces. If the model paraphrased it, summarized it or added a single word, the whole reading is discarded.
Pure Engineering over Discursive Noise
An article does not get through because it mentions a word. The filter weighs the themes of the seven kinds of fact against a noise list: elections and campaigns, common crime, homicides, kidnappings, terrorism, armed conflict, sports, entertainment and weather. If the noise outweighs the business facts, the article is discarded. Where the theme appears also matters: a strike in the first paragraph is not worth the same as a passing mention. With that, and with the match against the companies we follow, the queue is sorted by score before a single extraction is spent.
A Single Layer of Truth
The radar has parts written in two different programming languages, and the risk is that each part understands a reading its own way. That is why the shape of a reading is written three times, and an automatic test checks that all three say the same thing before any change is accepted. If the same batch of readings arrives twice, it is saved only once.
Internal Rigidity to Master External Chaos
The world’s press changes every 15 minutes. What processes it does not. The reader’s machine switches on in a single window a day and works in batches of fixed size. Each part of the radar shuts itself off if a credential is missing. If the extraction service fails on its side, the article goes back to the queue, with a limit on attempts. If the article’s own website is what fails, it is marked as failed and the rest carries on. And an alert goes off at 800 of the day’s 1,000 extractions.
From reading to signal: the signature
No reading turns into a signal on its own. Everything ends on an internal screen where a person reviews one article at a time, with the structure already laid out. An illustrative example:
| Field | Reading |
|---|---|
| Category | Labor dispute |
| Kind of fact | National strike |
| Organizations | A transport union (union) and an e-commerce company (affected company) |
| Summary | A national transport strike halted the distribution of goods across the country. |
| Relevant to companies | Yes |
| Quote | The exact sentence from the article, as it appeared |
| Language | Spanish |
Only when that person signs does the reading move into the Unified Signal Layer, linked to the right company.
In three lines
- Open-weights reading, not rented opinion. The models run on a dedicated machine, with a cost known in advance and without a third party’s filter.
- Rules that are not negotiable, instead of asking the model not to make things up. The shape of the reading is fixed, there is no room for personal data and every quote is checked against the original article.
- The machine cleans, the person decides. Automation ends where commercial judgment begins. A signature turns the data into a signal.
$ curl revsign.os/docs/open-weights-reading
→ {
title: "Open-Weights Reading: How the Market Radar Reads the World's Press",
answers: "How do you read the world's press with AI without depending on a commercial model's judgment?",
schema: [TechArticle]
}