Consulting runs on claims. Most practices deliver them unlabeled, so a number somebody measured sits in the same sentence, with the same confidence, as a number somebody guessed. We label every factual claim we publish and deliver with one of six evidence classes, and this page is the public definition of the system: what each class means, what it takes to earn it, and how to apply the ladder to any vendor claim in about a minute. It is written so you can adopt it, cite it, and hold us to it. The Evidence Ladder is published under the Creative Commons Attribution 4.0 International license (CC BY 4.0): you may use, adapt and republish it, provided you credit Delta Arc and link to this page.
What does each rung of the ladder mean?
Observed. We examined the primary thing itself: the original study document, the raw export, the call log, the contract, the timestamped record. Observed does not mean true in some cosmic sense; it means the shortest possible chain between the claim and the evidence, with no intermediary's summary in between. "The 2007 study reports a 100x change in contact odds [Observed]" means we read the study, not a blog post about the study. Observed claims always carry a link or an exhibit and a date.
Corroborated. At least two genuinely independent sources support the claim. Independence is the hard part and the whole point: two blog posts citing the same infographic are one source wearing two hats, a pattern we dissected in our census of the fabricated follow-up statistics. Corroboration requires separate chains of evidence: a primary document plus an independent reconstruction, or two measurements taken by parties who did not copy each other.
Derived. Computed from disclosed inputs by a method we show. A Derived claim is exactly as good as its inputs and its arithmetic, and labeling it Derived invites you to check both. "Your coverage multiple is 3.4x [Derived]" comes with the export it was computed from and the formula, as in our pipeline coverage method. When a derivation rests on provisional inputs, the label says so, as our CRM audit worry bands do.
Inferred. Our professional judgment from incomplete evidence: pattern recognition, experience, reasoning past a gap. Inferred is not an insult; most decisions that matter are made partly on inference, and a practice that claims otherwise is mislabeling. The discipline is admitting which parts are inference, so you know which parts of a recommendation would change if better evidence arrived.
Bounded absence. A defined search did not locate a record. It states the search, not a conclusion: "No primary study found [Bounded absence: searched X, Y and Z on November 9, 2026]" does not prove the study never existed, and the label refuses to let anyone round it up to that. Bounded absence carries the search scope and date so a reader can rerun it, and it is upgraded the day someone produces the record.
User-supplied. Context a client or reader gave us that we have not independently verified. It is preserved because it often matters, and it is labeled so it is never silently upgraded to verification. "Your team estimates quotes go out within a day [User-supplied]" stays exactly that until a timestamp export turns it into Observed or contradicts it.
The six classes map to the full names on our methodology page: direct observation, corroborated conclusion, derived estimate, working inference, bounded absence, and user-supplied. The short tags are the inline forms.
Why do the labels go inline instead of in a methods appendix?
Because appendixes are where honesty goes to be skipped. An inline tag sits inside the sentence doing the persuading, at the exact moment the reader is deciding whether to believe it. It also disciplines the writer: it is uncomfortable to type [Inferred] after a sentence you were hoping would land as fact, and that discomfort is the quality control. Across a document, the tags form a visible texture: a report that is wall-to-wall [Inferred] is announcing its own weight class, which is precisely the information an unlabeled report conceals.
How is this different from just citing sources?
A citation says where a claim came from. An evidence class says what kind of support it has, which is a different and usually more decision-relevant fact. The statistics we autopsy in this series were almost all cited, immaculately, link after link, and the links formed a circle with no primary source inside it. A fabricated number with twelve citations is still fabricated; the ladder grades the chain's anchor, not its length. Citation answers "who says?" The ladder answers "how do they know?"
How do you apply the ladder to a vendor claim in 60 seconds?
A vendor slide says: "Businesses like yours lose 40% of revenue to X." Run four questions, about fifteen seconds each, and you will land on one of the six classes:
- Can they show the primary source? Not a blog post, the study or dataset itself. If yes, and it says what the slide says, the claim can be treated as Observed, and you can proceed to checking its sample, date, and funder. If the trail ends at a roundup page or an infographic, mark it no better than Inferred and move on.
- Does anyone independent agree? One independent confirmation with its own chain of evidence upgrades toward Corroborated. Ten copies of the same origin do not; count origins, not links.
- Is it a computation, and are the inputs shown? If the number multiplies assumptions (an assumed job value, an assumed miss rate, an assumed close rate), it is Derived at best, and it is your right to see every input, a pattern we walked through with contractor missed-call math.
- What is left? If they searched for support and found none, that is Bounded absence, stated with its search. If the number came from a customer anecdote, it is User-supplied. Whatever else cannot clear the first three questions is Inferred: possibly honest, possibly useful, but a judgment, and it should be priced like one. The final move: ask what the vendor would accept as a test against your own data. Confidence invites measurement; theater resists it.
Sixty seconds will not tell you the claim is false. It will tell you what the claim weighs, which is usually enough to make the decision correctly.
What are the rules for using the ladder on your own work?
Four rules, and they are the whole method. Label at the claim level, not the document level, because real documents mix several classes in a single paragraph. Grade the anchor, not the chain: a claim inherits the class of its weakest necessary link. Date everything, since an Observed claim about 2012 is not an Observed claim about now, a decay this whole series documents. And when new evidence regrades a claim, change the label publicly rather than silently, the way our registry logs its changes. If you adopt the ladder, link this page as the definition so the labels mean the same thing across documents; a shared convention beats a private one.
Frequently asked questions
Can a claim hold two labels?
Yes, and precision improves when you split it. "Response audits found 23% of companies never responded [Observed], so some of your leads are probably dying unanswered [Inferred]" is two claims, honestly weighted, where one fused sentence would smuggle the inference in under the observation's badge.
Is Observed always stronger than Inferred?
No. An Observed claim from a tiny, stale, or conflicted sample can deserve less weight than a careful inference from broad experience. The ladder labels the kind of support, not a final score; it exists so the weighing can happen in the open.
Who checks you?
You do. Every labeled claim carries enough of a trail, a link, an export, a formula, a dated access note, that a reader can retrace it. When we get one wrong, the regrade is logged, not buried.
May we use this system internally or publicly?
Yes, under CC BY 4.0, and attribution is required: credit "The Evidence Ladder, Delta Arc" with a link to this page, and say if you changed any definition. The convention works better the more teams share it, and the credit keeps every copy traceable to one definition.