TL;DR
- Findings in organizations travel through a chain of summaries, and each hop tends to remove qualifiers, scope limits and sample caveats before it removes the conclusion.
- The result is a one-way ratchet: confidence at the top of the chain tends to be higher than confidence at the source, even though no one at any step intended to overstate anything.
- Summaries are not the problem. The problem is that the evidence trail is rarely attached to them, so the reader cannot tell a verified claim from a repeated one.
- The useful diagnostic is a single question about any decision input: what is the earliest record of this, and who has looked at it since?
The pattern: confidence travels faster than caveats
A finding begins life with its limits attached. An analyst notes that a trend holds for one region, over one quarter, among customers who responded to a survey. That sentence of limits is longer and less quotable than the conclusion, so the first summary shortens it. The next summary, written for a busier reader, shortens the shortened version. By the time the finding reaches a planning meeting, it reads as a plain fact: "customers prefer the new option."
Nobody in that chain lied. Each person compressed accurately by the standards of their own reader. What changed is the ratio between claim and support. The claim stayed intact, or grew slightly stronger through rounding, while the support dwindled to a reference nobody opens.
This is a specific mechanism inside what WBA has described elsewhere as epistemic debt: "The gap between what an organization believes to be true and what is actually true, compounded over time by the cost of acting on that gap." The summary chain explains one way that gap opens even when the original work was sound.
Why qualifiers go first
Three properties of ordinary communication produce the asymmetry.
Qualifiers are costly to read and cheap to omit. A conclusion can be acted on. A limit can only be weighed, and weighing takes attention. Executive summaries, slide titles and status lines are optimized for action, so the structure of the format selects against hedges.
Qualifiers belong to the author, conclusions belong to the audience. The person who ran the analysis knows where it is thin. The person who repeats it knows only what was said to them. A caveat that lives in the author's head cannot be forwarded.
Repetition reads as corroboration. When the same claim appears in a team update, a leadership deck and a vendor conversation, it feels like three sources. These can easily be one source, echoed. The organization experiences agreement where it actually has duplication.
Three signatures of an inflated summary chain
The inflation tends to take three recognizable forms.
| Signature | What it looks like | What was lost |
|---|---|---|
| Scope creep | A result about one site, product or period is stated as a general truth | The boundary of where it was measured |
| Tense drift | "Was true when checked" becomes "is true" | The date of the observation |
| Certainty promotion | "Likely," "appears," "in most cases" become unqualified statements | The author's own estimate of confidence |
Tense drift deserves particular attention because it is invisible from the inside. A statement that a process "handles exceptions automatically" may have been accurate when someone last tested it. Systems, staffing and suppliers change underneath it. The sentence persists in documentation and in conversation long after the observation behind it has gone stale, and nothing marks it as old. A related problem, the absence of expiry dates on decisions, is examined in Decisions Without Expiry Dates; claims suffer the same condition as the decisions built on them.
The ratchet has no reverse gear
Compression is easy in one direction. Restoring a dropped qualifier requires going back to the source, and the source is usually the least accessible part of the chain. The original analysis may sit in a folder no one has opened since, or live in the memory of someone who has moved to another role.
This creates an asymmetry in who bears the cost of checking. Questioning a widely repeated claim means interrupting a meeting, implying a colleague was careless, and doing archaeology that no one has budgeted for. Accepting it costs nothing today. Organizations therefore tend to accept, and the cost of that acceptance is deferred into whichever decision eventually relies on the claim and fails.
The deferred cost is also hard to attribute. When a plan built on an inflated claim underperforms, the post-mortem looks at execution, because execution is where the visible events happened. The claim that quietly set the plan's assumptions is rarely the subject. This is part of why the same errors recur: the review examines the layer that is documented rather than the layer that was wrong.
What changes the dynamic
The pattern points toward structural remedies rather than exhortations to be more careful. Careful people still work inside formats that reward brevity. Several structural changes target the mechanics directly:
- Summaries that carry a pointer to the primary record, so that any reader can find the earliest version of a claim without asking.
- A visible observation date on operational claims, treated as part of the claim itself rather than as metadata.
- A separate field for confidence in reports that feed decisions, so that "how sure" travels alongside "what."
- A distinction between "reported" and "verified" in status language, so that a repeated claim and a checked one do not look identical.
None of these requires new technology. They require agreeing that a summary without a trail is an assertion, and labeling it as one. Organizations that add AI-generated summaries to their workflows face a sharper version of the same issue, since automated compression is fast, fluent and equally indifferent to which sentences carried the caveats.
What to take from it
The summary chain is not a failure of honesty. It is a predictable result of compressing information for people with limited time, in formats that favor conclusions. The cost appears later, as decisions made with more confidence than the evidence supported and no record of where that confidence came from.
The measure of an organization's knowledge is not how firmly its claims are stated but how quickly someone can walk one back to its source. Where that walk takes minutes, confidence is probably earned. Where it takes an investigation, the confidence is likely borrowed.
Research questions on this pattern are welcome through Inquiries.