CompanyGraph
CompanyGraph is not trying to make more claims. It is trying to stop claims from becoming stronger than their evidence.
Most financial tools turn data into scores, labels, or summaries. CompanyGraph takes a slower route: it asks what each claim is allowed to say.
CompanyGraph is built around claim discipline.
A number is not yet a claim.
A claim is not yet evidence.
Evidence is not yet trust.
Trust is not yet a recommendation.
And a verified observation is not automatically something a reader should see.
Before CompanyGraph shows a reader-facing statement, the system tries to keep that statement inside the limits of its origin, its evidence, its trust level, its placement, and its wording. This page describes what that means in practice, and the failure modes it is designed to resist.
What counts as a claim
A claim is any statement the product makes to a reader. It can be a sentence, a chart label, a report line, a score, a badge, a comparison, or a visual cue. A green mark is a claim. A prominent position on a page is a claim. A name is a claim.
Claim discipline means asking, for each of them:
- Where did this claim come from?
- What supports it?
- How strong is that support?
- What does it allow us to say?
- What does it not allow us to imply?
- Where should the claim appear, if anywhere?
- Could the wording accidentally sound like prediction, advice, or investment merit?
We do not only ask whether a number is correct. We ask what that number is allowed to mean.
This is why CompanyGraph keeps data, quality, claim, proof, evidence, trust, placement, and surface separate. Each is a different question, and answering one does not answer the others.
The path a claim travels
Inside CompanyGraph, a reader-facing statement is the end of a chain, not the beginning — and the chain closes back on itself:
In plain terms: data becomes usable. Usable data becomes a claim. The claim is checked. The check becomes evidence. Evidence becomes trust. Trust guides placement. Placement shapes the surface a reader sees. The surface is used. Use creates feedback. Feedback drives steering. And steering changes the system.
No layer claims truth. Each layer only says how strongly the next layer may speak. A claim that fails a check is not argued for; it gets weaker wording, a lower trust label, a less prominent place, or no place at all.
To see how a number moves through the system before it becomes a reader-facing claim, read From data to checked claims.
Common ways financial claims become too strong
These are not accusations against any specific product. They are general failure modes that can appear wherever financial data is turned into reader-facing meaning, including here, which is why the system is built to watch for them. For each one: what happens, why it can mislead, and what CompanyGraph tries to do instead.
Borrowed trust
A weak or uncertain claim becomes more believable because it appears inside a careful-looking system. The claim did not earn that trust itself; it borrowed credibility from the interface around it.
Instead: keep support levels visible per claim, and do not let evidence or trust language near one claim strengthen an unrelated one.
Origin drift
A claim changes source as it moves through the system. "We interpret this" quietly becomes "the company says this." Readers weigh company testimony, model interpretation, and calculated observation differently, so a drifted origin changes what they believe.
Instead: every claim carries an origin. Generated interpretations are labelled as CompanyGraph interpretations, never as company statements.
Evidence stretch
The evidence supports a narrow fact, but the wording implies a broader conclusion. "Revenue grew" can become "strong growth company," and then quietly imply quality or safety.
Instead: keep the wording inside the evidence. Say the narrow thing, unless the broader thing is separately supported.
Verdict language
A descriptive observation becomes a judgment: healthy, distressed, cheap, expensive, safe, strong, weak. Verdicts feel like conclusions, even when the system only measured one limited pattern.
Instead: prefer descriptive language over verdict language. An observation is named for what its formula measures, not for an investment label.
Advice leakage
A description starts to sound like a recommendation. Even without the words buy or sell, wording can imply what the reader should do.
Instead: avoid buy, sell, and should language. CompanyGraph stays descriptive, not advisory.
Freshness illusion
Old or stale data is shown as if it were current. A true historical number can become misleading when it is presented without its age or context.
Instead: show as-of dates, fiscal periods, freshness notes, and stale-data warnings where they matter.
Score compression
A complex company situation is collapsed into one score, color, badge, or simple verdict. The reader may treat the score as the answer, while the uncertainty and trade-offs disappear.
Instead: expose the underlying observation and its limits. A simple visual is not allowed to carry more meaning than the evidence supports.
Model laundering
A language model's output is shown as if it were direct reality. Generated interpretation can sound like fact when its source is not visible.
Instead: label model-generated interpretation as interpretation. Generated text stays downstream of evidence, inside stated boundaries.
Placement overreach
A verified number is shown so prominently that readers assume it matters more than it does. "Verified" gets mistaken for important, safe, or actionable.
Instead: keep trust and placement separate. A reproducible observation can still be a supporting detail, or not belong in a report at all.
Context stripping
A number is shown without the context needed to understand it: period, unit, currency, industry, peer range, denominator, or calculation trail. The number may be technically correct and still practically misleading.
Instead: show the context where it is needed: period, units, peer bands, calculation trail, source, and support level.
Name overclaim
The name, title, or caption says more than the formula measured. Readers trust the name, not the implementation.
Instead: treat names and captions as part of the claim. If a formula measures operating income growth, its label does not say margin trend.
Proof confusion
A claim being reproducible is treated as if that made it true, important, useful, or suitable. Proof checks whether the claim can be reproduced from stored inputs; it does not establish investment merit.
Instead: keep proof, trust, suitability, and reader placement as separate questions with separate answers.
Surface confidence
Visual design makes a claim feel more certain than the underlying support allows. Green badges, strong icons, prominent placement, and confident prose can all imply certainty.
Instead: match the visual register to the support level. Trust notes exist to make uncertainty legible, not to impress the reader.
What CompanyGraph tries to do differently
CompanyGraph tries to keep claims bounded. That means:
- Every claim needs an origin.
- Every evidence-backed claim needs a support path.
- Every trust label needs a reason.
- Every reader-facing claim needs a suitable place, and some have none.
- Every strong visual cue needs enough support to justify its register.
- Every correction should make the same class of mistake harder to repeat.
This does not mean CompanyGraph is perfect. It means the system is designed to notice when a claim has become too strong, route the problem to the layer that owns it, and tighten the guard so the same slip is harder next time.
Discipline does not mean perfection
CompanyGraph does not claim that every past or future statement is perfect.
We have found cases where a claim said more than it should have, where wording carried the wrong origin, where stale data needed clearer surfacing, where a name implied more than its formula measured, and where a verified observation still was not suitable for a reader-facing report.
Those cases are part of how the system improves. The goal is not to pretend mistakes cannot happen. The goal is to make them visible, correctable, and harder to repeat.
Examples of tightening
Each of these reflects a real adjustment made while building the product:
- Generated structural readings should not be phrased as company statements.
- Verified observations should not automatically become report headlines.
- Stale figures need visible as-of context.
- Provider ratios need plausibility checks before they become observations.
- Formula names and reader-facing captions must match what the formula actually measures.
- A lower honest trust level is better than fake certainty.
Companies are not only financial statements
CompanyGraph uses balance sheets, income statements, cash-flow statements, and historical financial data because they can reveal useful traces of how a company works: scale, margin structure, capital intensity, cash generation, leverage, cyclicality, deterioration, consistency, and change over time.
Financial statements are traces of a company, not the company itself.
A company is also a role inside the economy. It uses inputs, depends on infrastructure, serves customers, and sits inside supply chains. It may be exposed to regulation, energy, logistics, commodities, labor, platforms, capital markets, or physical bottlenecks. Other companies, households, institutions, or systems may depend on it. Those relationships can matter as much as the numbers, especially for long-term understanding.
So CompanyGraph tries to look at both the financial trace and the operating position. The financial trace asks: "What happened in the accounts?" The operating position asks: "What kind of system is this company part of?"
This does not make CompanyGraph an oracle. It does not mean the system can see every internal decision, hidden risk, cultural problem, strategy change, fraud, future shock, or qualitative development. It means CompanyGraph treats companies as real operating systems, not just financial instruments.
Where the claim stops
CompanyGraph is built for people who want to inspect companies without being pushed toward a verdict.
It does not try to replace judgment. It tries to make the material for judgment cleaner: where the claim came from, what supports it, how strong that support is, and where the claim stops.
The product is not a precision machine. It is a claim-discipline machine.