What the model may do, and what not
At Schwung, AI is a measuring instrument with distrust built in. The model may read, organise and write. It may not count, not judge, and not claim anything without a source. This page is the accountability behind our way of working: how we check what a model says, and where the limits are.
Six rules that are not settings
They apply in every engine and every analysis, and they live in the code, not in a manual.
The model does not count
All numbers in an analysis come from a deterministic counting step. The model may only quote them. A number check rejects every figure, every percentage and every "most" that is not in the counts. An external figure carries its source in the same sentence.
A second model tries to refute the analysis
The writer and the checker are different models. The checker looks for what does not hold: axes that do not spread, distinction that is not real, a conclusion wider than the evidence. If the analysis fails, it is rewritten with the criticism included. What is still weak after four rounds is stated in the piece.
Every statement carries a label
What has been measured is called a fact. What has been seen but not counted is an observation. Our reading is an interpretation. An explanation is a hypothesis, and cause and effect are always a hypothesis. What we would need to know to decide is stated as a question.
The outside world only with source and date
Legislation, policy, funding, labour market, demographics: every finding carries a URL, a date, a status (in force, decided, proposal, forecast) and a quote found literally in the source. A follow-up check discards what is not literally in the source. News and press releases are a signal, never the source of a rule.
Second reading is standard
Every piece is checked by a second reader before it goes to a client or online: first the profiles against the source, then the context, then the synthesis per section. For the Schwung Merkbeeld, the second reading of a larger group found dozens of corrections before they went live. The interpretation itself is human work and is read by Schwung before distribution.
What we do not know is stated
Unknown never counts as no. An engine reads what organisations publish and can say what someone claims and how they prove it, not whether it is true or who is best. A group drawn from client runs is regional and small. Every publication says so again, and every analysis ends with the questions only the client can answer.
The engine owns the rules, the model delivers the content
In every engine the build is the same. The engine owns the template, the rules and the counts. The model reads, organises and writes, and delivers its work in a fixed form. A step without a model turns that into the analysis, the page or the file.
That way a model can never send a conclusion, a figure or a page into the world on its own. What the reader sees has passed through a rule that is not the model's. And because every step keeps its intermediate result, we can redo one organisation, one rule or one section without touching the rest.
Honest about the coherence: the engines are separate tools, not a shared system. What one delivers, we carry into the next step. The chain is one coherent process, not an automatic link, and we keep it that way until a link is demonstrably better than a person who carries forward what they have learned.
The same division of roles applies to imagery. Where we create imagery with AI, people decide the concept, the choice and the direction, and we never present generated people as real patients, clients or employees.
How we test whether it works
- The Brand Scan on our own brand. Two methods, the instant scan and the conversation, run independently on the same site, arrived at the same diagnosis and the same first step. We published that, including where the instant scan fell short.
- The Schwung Merkbeeld with an open method. What we read, which grounds a count carries, what high, medium and low mean and when the version changes, is on the method page, with a change log. Every edition publishes the dataset.
- Instability in the accountability. If a coding step gives different outcomes per run, we run it three times and take the majority, and we put that spread in the piece.
- The limit is documented too. Three runs and a second model once let an announced project stand as existing, because all models read the same future tense as completed. Repetition catches inconsistency, not a shared error of thought. That is why external source checking is a separate step, and why a human reading remains the last.
What the engines read and what they do not share
The engines read public sources: websites, annual reports, strategy documents, legislation and open data. What a client shares with us in confidence stays outside the shared research data: the boundary sits in the first step of the chain, not in an agreement afterwards.
Published sector research therefore never contains client information. In the register a client is an organisation like any other. Models are not trained on client work.
Frequently asked questions about our AI method
Why is the model not allowed to count at Schwung?
Because a language model can invent numbers that sound credible. That is why all numbers in our analyses come from a counting step that uses no model. The model may only quote those numbers. Any other number, percentage or 'most' without a count is rejected before the piece is finished.
What does the second reader do?
A second model, on a different basis than the writer, tries to refute every analysis: are the axes right, is the distinction real, is the white space really there. If the analysis fails, it is rewritten with the criticism included, up to four rounds. What is still weak after that we report in the piece instead of smoothing it over.
How do I know whether a statement is a fact or an interpretation?
Every statement in our analyses carries a label: fact, observation, interpretation, hypothesis or question. A cause-and-effect relation is always a hypothesis. So you see at a glance what has been measured and what is our reading.
Does the model use its own memory for facts about the market?
No. Legislation, policy, figures and developments come from external sources, with URL, date and a quote that has been found literally in the source. What the model knows from memory we do not use as a source.
How do you test whether the method works?
We measure ourselves. We ran the Brand Scan on our own brand and published the outcome. The method of the Schwung Merkbeeld is open, with a change log. And if a control step gives different outcomes per run, we run it three times and take the majority, and we put that uncertainty in the accountability.
What does the method not know?
A lot. Our engines read what organisations publish; they can say what someone claims and how they prove it, not whether it is true or who is best. A group drawn from client runs is regional and small. We say that in every publication, every time.
Start with the free Brand Mirror.
In ten minutes you see what the method makes of your brand, with the core tension we see from the outside and the questions we cannot answer without you.