The Seams
Honest systems disagree.
The disagreement has a structure, and the structure can be measured.
A study of why measurement systems inside one organization report different values for the same reality, where that disagreement sits, and what it means to agree.
Ongoing research · Amoeba AI Labs
Abstract
Every organization measures itself more than once. Marketing, finance and sales each hold a defensible account of the same quarter, and the accounts rarely match. The usual response treats the mismatch as error, to be cleaned away until one number remains. We argue that much of it is structural: it comes from the definitions applied where one system hands data to another, and some of it cannot be removed by better data at all.
We model an organization's systems as a cellular sheaf. Systems are nodes, handoffs are edges, and the definitions applied at each handoff are restriction maps. Established results in the spectral theory of sheaves then give three things organizations currently lack: a measure of disagreement that can be located edge by edge, a test for whether a consistent account exists at all, and a reconciliation that records whether agreement was reached by changing the evidence or by changing a definition.
We test the framing against documented public cases in national accounting, public health reporting and digital advertising, and identify three types of seam: source, definitional and causal. Adjacent fields, notably sensor fusion and risk analysis, already use these instruments. We found little published work applying them to an organization's measurement of itself. Worked examples use a hypothetical company, Acme. Its figures are illustrative.
Figure 1. Three Acme systems, marketing automation, CRM and ERP, report revenue-linked values that should agree. They do not, and 69% of the total disagreement sits at a single handoff, from marketing automation to CRM.
(a) Each system holds a value; restriction maps send values into the shared space on each edge, and the coboundary δx measures the mismatch there. (b) Total disagreement E(x) decomposes by edge. (c) Sheaf diffusion drives the values to the nearest consistent state, the orthogonal projection of the data onto the global sections H⁰. Values in $M after documented definitions are applied. Acme is hypothetical; figures are illustrative.
Findings
Four primary results
1.1
Consistent systems can be wrong together
When systems share a flawed definition, every consistency check passes. Uber paused $100 million of $150 million in annual advertising spend and observed almost no change in app installs, as reported by its former head of performance marketing. Agreement between systems was not evidence about the world. Only an intervention revealed the error.
1.2
Plurality of measurement is structural
Independent, honest measurement of the same quantity persistently diverges, and greater precision does not close the gap. US output measured by expenditure and by income has differed by an average of 0.8 percent of GDP since 1947 (BLS), and revisions do not, on average, shrink the discrepancy (Federal Reserve Bank of Cleveland).
1.3
Disagreement has a location
The sheaf Laplacian decomposes total disagreement by edge, so a mismatch can be attributed to a specific handoff and the definition applied there. Its kernel is the space of globally consistent assignments. When the definitions admit no consistent assignment compatible with what each system observes, the conflict is structural: no amount of data cleaning resolves it.
1.4
Agreement can be produced
In the discourse-sheaf model of Hansen and Ghrist, a network reaches public agreement while private states remain far apart, and translations between parties often move faster than beliefs. Ghrist calls the result “an expressed external consensus: a truce.” An agreed number may record a settlement rather than an observation.
Implications
What it means
If disagreement is structural, located, and sometimes invisible to consistency checks, then organizational knowledge depends on instruments that sit between systems as well as within them.
A shared number is a settlement until shown otherwise
Every reconciliation changes either the evidence or a definition. Keeping only the final figure discards which one it was, and with it the means to audit the agreement later.
A quiet meeting is weak evidence of alignment
Agreement in expression is compatible with large disagreement in the underlying systems. Measuring alignment by the absence of argument measures the wrong quantity.
The seams are the object of study
Most investment improves systems in isolation. Whether systems can agree is decided at the edges between them: their definitions, their owners and their history.
Some conflicts are decisions
When definitions admit no consistent assignment, more data cannot settle the question. Resolution requires a choice between the owners of those definitions, and that choice should be recorded as such.
Consistency must be tested against the world
Systems that share a flawed rule pass every internal check. Holdouts, pauses and natural experiments are the only instruments that test agreement against reality.
Models propose; mathematics adjudicates
Language models are well suited to generating interpretations of evidence. They are not well suited to judging them. Auditable mathematics determines what the evidence supports, and recorded outcomes correct both.
The series
The Seams, in four parts
The truce in the numbers
Introduces the discourse-sheaf model of Hansen and Ghrist and maps its two mechanisms, adjusted translation and withdrawn communication, onto the functions of an organization.
The math came first
Traces the model to the spectral theory of cellular sheaves and derives four instruments: a located measure of disagreement, a structural test for consistency, an auditable reconciliation, and a method for learning definitions that were never written down.
Honest systems, different numbers
Examines documented cases from national accounts, public health reporting and digital advertising, and classifies each as a source, definitional or causal seam.
Where the gap actually is
Surveys applications of sheaf methods in sensor fusion, enterprise data, risk analysis and accounting theory, and locates where the published work stops.
Statement of scope
The mathematics in this series is published and well established; none of it is new. What is proposed here is its application to an organization's measurement of itself. Public cases are drawn from the sources cited in each part, some of them secondary or vendor-reported. Acme is a hypothetical company used to show what each instrument reports, and its figures are illustrative. Every instrument described here measures consistency between systems. None of them measures truth. That limit is deliberate: it is what keeps the instruments honest, and it is why the program ends in experiments.
Questions
Questions this work addresses
Why do two honest systems report different values for the same quantity?
Because each applies its own definitions at the point where it receives data from another system. The difference is often structural.
Can an organization have a single source of truth?
It can have a single agreed value. Whether that value reflects shared evidence or a negotiated change in definitions is a separate question, and one most organizations do not record.
Does more or better data resolve disagreement between systems?
Not always. When definitions admit no consistent assignment, the conflict persists regardless of data quality and requires a decision between the owners of those definitions.
If all systems agree, are they right?
Not necessarily. Systems that share a flawed definition pass every consistency check. Only intervention, such as a holdout or a pause, tests agreement against the world.
Ongoing research
This work continues
The Seams is the first study in a research program on a question older than any of our systems: how an organization decides what to believe enough to act on. The next studies extend from measurement to belief, examining how disagreement between systems becomes disagreement between people, and how either can be resolved without hiding what was given up.
We are collecting documented instances of internal measurement disagreement. If two of your systems report different values for what should be the same quantity, we would like to study the seam.
Contribute a casetitle = {The Seams},
author = {{Amoeba AI Labs}},
year = {2026},
howpublished = {\url{https://amoeb.ai/the-seams}},
note = {Ongoing research}
}