CompanyGraph · how the system works
From data to checked claims.
One rule runs through everything Smallbox built into CompanyGraph, the stock-analysis platform: a number does not become a claim in one jump. It crosses a row of stages — each answering a different question, none allowed to borrow certainty from the others — and what readers do with the result flows back and changes the earlier stages. This page draws that pipeline — the system in time. The companion page, System anatomy, draws the same system in space: the parts, the boundaries between them, and the shapes that cross each line.
Data is not meaning.
Most systems jump straight from data to confident prose. A figure appears and a sentence tells you what it means — with no visible line between the number, the check, and the interpretation. CompanyGraph refuses that jump. Raw data, a specific claim, its proof, how far the claim can be trusted, and where a reader may meet it are kept as separate stages. Each answers a different question, and none of them is allowed to borrow certainty from the others.
What each stage does.
Each stage answers one question and prevents one kind of false confidence.
What the company reported — raw statements and figures, each tagged with its date.
What did the company actually report?
Whether a figure is usable at all. Units and sanity are checked before anything is computed with it.
Can we use this number?
A single, specific observation — one measured pattern, described plainly. Never a verdict or a forecast.
What does this figure show, exactly?
The claim is recomputed from the company’s own statements to see whether it reproduces.
If we check it ourselves, does it hold?
The check, written down — a durable record of what was verified, the arithmetic, the source, the date.
What is the standing record of the check?
How firmly the system can stand behind it — a level that reflects which check passed, not whether the company is “good”.
How strongly may we speak?
Where, if anywhere, a reader meets it — a body claim, a supporting figure, or not shown as a claim at all.
What claim does this become, and where?
What actually appears on the page — the stock page, a report section, a note — each line labelled so its status is clear.
What does the reader see, and in what register?
A person reads, compares, asks a question.
What happens on contact with a reader?
What comes back — a wrong label, a stale figure, a check that disagrees, a sentence that confused someone.
What did we learn from being used?
The system adjusts. Feedback changes the earlier stages — better data checks, corrected formulas, clearer captions, different placement.
What do we change so the next pass is better?
Checked, but not always shown.
A figure can pass every check and still not appear as a claim — because its name would promise more than the check proves. “Can we rely on this number?” and “What should a reader be told, and where?” are different questions, so they are answered by different stages. That separation is what lets the surface stay honest: trust is earned by proof; being shown is earned by relevance and plain, accurate wording.
Steering feeds back.
The last stage is a control step. When feedback arrives — a wrong label, a stale figure, a check that disagrees — steering changes the stages that produced it: the data checks, the formula, the wording, the placement. The next pass through the path behaves differently. Nothing is frozen; the system moves toward being well-calibrated, not toward saying more.
You can always tell what kind of thing a sentence is.
Because the stages stay separate, every sentence carries its kind on its face:
- a figure the company reported
- an observation the system checked itself
- a structural reading, offered openly as interpretation
- a gap it cannot yet see
A system that blends these into one confident voice hides the difference. Here the difference stays visible by design — trust is something you can inspect, not something you are asked to grant.
A report is one surface, not the system.
A report is one place this output is read. It consumes checked observations; it does not decide what is trusted. Trust is decided upstream, before the report tries to use the claim — so a report can only show what the earlier stages already stand behind. The loop is the system. A report is one of its surfaces, and reporting — a reader telling it something is off — is one of its feedback signals.
No stage claims truth. Each stage only says how firmly the next one may speak.
The goal is not more output. The goal is output with boundaries.
You do not have to take this page’s word for it. CompanyGraph publishes this method in its own voice, on the product: the claim loop and how it handles claims.
This is one system Smallbox built.
The same discipline — one rule holding from the database all the way to the public copy — is what Smallbox brings to a codebase.