Evidence¶
Core Idea¶
Evidence is the structural relation between an observable trace and a hypothesis about an unobservable state of affairs: a thing functions as evidence for a claim when it is more probable to occur if the claim is true than if it is false, and when its production is traceable enough that the inference can be challenged and defended. Three structural commitments travel with the relation across substrates. First, evidence is defeasible and weight-bearing: any single item licenses only a graded confidence in the conclusion, not a deduction, and additional information can raise, lower, or wholly defeat its weight. Second, evidence carries provenance: the chain of custody from the underlying event to the trace to its presentation is itself part of the evidence, because breaks in that chain are the standard route to defeating an inference. Third, evidence combines non-arithmetically: two independent corroborating items raise confidence more than one does, but two items from a common source raise it less than the independence picture would suggest, so the combination rule must care about dependence among traces.
What makes this a prime rather than the vocabulary of any one inquiry is that the same role-set — underlying state of affairs leading to a trace-producing mechanism, to a trace, through a chain of custody, via an inference rule, with aggregation across multiple traces and a set of defeaters — structures evidential reasoning in science, law, medicine, history, journalism, intelligence assessment, software debugging, and archaeology. The substrates differ entirely; the skeleton does not. The relation is inherently bound to inquiry practice — it carries an epistemic-normative load (burden of proof, standards of admissibility) and an institutional history in the legal rules of evidence — which is why it reads as framed even though its role-set is medium-neutral.
How would you explain it like I'm…
Clues That Point
Clues You Can Trust
Traces Behind a Claim
Structural Signature¶
the unobservable state of affairs — the trace-producing mechanism — the observable trace — the chain of custody from event to trace — the inference rule from trace to hypothesis — the aggregation rule across multiple traces — the set of defeaters
Evidence is present when each of the following holds:
- An unobservable state (the target hypothesis). A claim about a state of affairs not directly observed; a trace is evidence only relative to a specific such hypothesis.
- A trace-producing mechanism (the generator). The process by which the underlying state leaves an observable mark; its character (study design, taphonomy, instrument) governs how strongly the trace bears on the claim.
- A trace (the observable). The thing that functions as evidence — probative when it is more probable if the claim is true than if false.
- A chain of custody (the provenance relation). The traceable path from the underlying event to the trace to its presentation; breaks in this chain are the standard route to defeating an inference, so provenance is itself part of the evidence.
- An inference rule (the weight-bearing relation). A defeasible mapping from trace to graded confidence in the hypothesis — never a deduction; any single item licenses only weight, which further information can raise, lower, or defeat.
- An aggregation rule (the combination invariant). A rule combining multiple traces non-arithmetically: independent corroboration raises confidence more, common-source traces less, so the rule must care about dependence among traces.
- Defeaters (the defeasibility invariant). Conditions that would undercut the inference — confounders, custody breaks, alternative explanations — whose anticipation is part of evaluating the evidence.
The components compose into a role-set in which any disputed conclusion localizes to one handle — a provenance break, a wrong combination rule, an ignored defeater — making evidential reasoning auditable rather than holistic.
What It Is Not¶
- Not provenance.
provenanceis the chain of custody from event to trace — one component of the evidence relation, the handle that defeats inferences when broken. Evidence is the whole role-set; provenance is the lineage sub-relation within it. - Not the inference rule alone.
inductive_reasoning,abductive_reasoning, andbayesian_updatingare modes of inference from evidence; evidence is the trace-to-hypothesis relation those modes operate on. The rule combines traces; evidence is what is combined. - Not belief.
belief_formationis the resulting doxastic state; evidence is the trace that licenses graded confidence. One can have strong evidence and form no belief, or believe with no evidence. - Not data. Data are traces before a hypothesis is fixed; a trace becomes evidence only relative to a specific claim and inference rule. "Having data" is not "having evidence for a hypothesis" — the prime's central distinction.
- Not proof. Proof is deductive and indefeasible; evidence is defeasible and weight-bearing, licensing only graded confidence that further information can raise, lower, or defeat.
- Not a signal in the detection sense.
signal_detection_theoryconcerns discriminating signal from noise at a sensor; evidence is the broader relation in which a trace bears on an unobservable hypothesis through a custody chain and an aggregation rule. - Common misclassification. Gathering data, then finding the claim it best supports, and treating the fit as evidential weight (HARKing). Catch it by asking whether the hypothesis was specified before the trace was examined: if the claim was chosen to fit the data, the trace was never a likelihood-discriminating test of it and its weight is illusory.
Broad Use¶
Evidence is the working unit of every discipline that reasons from observable traces to unobservable states. In law, physical exhibits, testimony, documents, and expert opinion must be both probative and admissible, and juries weigh them against an explicit standard of proof; the Federal Rules of Evidence operationalize the structural commitments directly. In science, data are evidence for or against hypotheses, with strength depending on study design (the trace-producing mechanism), replication (independent corroboration), and freedom from confounders (defeaters). In medicine, signs and symptoms are evidence for diagnoses, and differential diagnosis is structurally a Bayesian-style competition among hypotheses, each finding shifting posterior weight. In history, primary sources are evidence for past events, with provenance, internal-versus-external criticism, and triangulation across independent sources playing the chain-of-custody and combination roles. In journalism and intelligence, sources, documents, and intercepts function as evidence for assessments, and tradecraft is precisely the discipline of provenance, source reliability, and corroboration. In software debugging, a stack trace, log entry, or reproducer is evidence for a hypothesis about a defect, with bisection and instrumentation as evidence-production strategies and flaky reproducers as defeaters. And in archaeology, physical remains and stratigraphic context are evidence for inferences about past life, with taphonomy supplying the trace-producing-mechanism analysis.
Clarity¶
Naming the structural relation moves the analysis from "we have lots of data" to a specific set of questions: evidence for what claim, by what trace-producing mechanism, with what chain of custody, in combination with what other traces, vulnerable to what defeaters? It clarifies the common confusion between having data and having evidence for a hypothesis — data become evidence only relative to a specific claim and a specific inference rule. It also separates the probative question (how strongly does this trace bear on the claim?) from the admissibility question (does our procedure permit us to use it?), a distinction law makes explicit and other domains often blur. The clarifying force is to dissolve a vague mass of "findings" into a small set of named handles, each of which can be examined and contested on its own, so that an argument about a conclusion becomes a set of sharper arguments about specific evidential components.
Manages Complexity¶
The evidence relation factors any inference into a small set of analytically separable handles: the trace-producing mechanism, the provenance chain, the inference rule, the combination rule, and the candidate defeaters. When a downstream conclusion is in dispute, the dispute localizes to one of those handles, which compresses the argumentative space dramatically. Forensic, scientific, and historical disputes that look like disagreements about the conclusion almost always turn out to be disagreements about one specific evidential handle — a provenance break, a wrong combination rule, an ignored defeater — and isolating which handle is at issue converts an unbounded quarrel into a tractable one. The reduction also makes evidential reasoning teachable and auditable: rather than cultivating an inarticulate sense of "what the evidence shows," a practitioner can walk each handle in turn, which is exactly what legal procedure, scientific peer review, and intelligence tradecraft institutionalize. The complexity managed is the combinatorial tangle of many traces bearing on many hypotheses; the handles give it a fixed, inspectable structure.
Abstract Reasoning¶
Evidence relations support reasoning under the explicit assumption that all inference from traces is defeasible and weight-bearing, which unlocks several second-order moves. Reasoning about the optimal next evidence to collect — experimental design, discovery requests, source cultivation, active learning — is uniformly an instance of asking which trace would most discriminate between the live hypotheses given the current posterior. Reasoning about evidence of absence versus absence of evidence turns on whether a positive finding was likely: a failed search where success was probable is evidence against the claim, while a failed search where success was improbable is not, and treating these symmetrically is a recurring error. Reasoning about correlated traces warns that two traces sharing a producing mechanism do not constitute independent corroboration, a pattern that recurs in news (outlets citing one source), in clinical findings (two tests on the same biomarker), and in intelligence (two reports from one asset). And reasoning about defeaters as predictions notes that a hypothesis which systematically generates defeaters of its own apparent evidence is in worse shape than one that does not, even before any single defeater is realized — a structural feature that adversarial testing deliberately exploits.
Knowledge Transfer¶
Because evidence is a role-set rather than a domain, an intervention discovered in one inquiry practice transfers to any other by re-identifying the roles, and the prime's reach lies in how routinely this works. Chain-of-custody discipline from law transfers to data lineage in engineering: the legal rule that admissibility depends on documented custody from collection to courtroom becomes the engineering rule that traces whose lineage is unrecorded cannot ground a downstream claim, and the intervention is identical — instrument the pipeline and attest each hand-off. Triangulation in history, corroboration in intelligence, and replication in science are the same move under three names — produce a structurally independent trace and check for convergence — so a historian's source-criticism and a scientist's replication study are recognizably the same upgrade to confidence. The legal notion of burden and standard of proof transfers to the statistical significance threshold and to the policy bar-for-action: choosing a threshold for acting on an evidence-weighted conclusion is present in every evidential domain, and the legal habit of making that threshold explicit and contestable is exportable to fields that usually leave it implicit. Adversarial examination in law transfers to red-teaming in intelligence and pre-mortems in organizations, all generic disciplines for surfacing defeaters before acting. And Bayesian updating supplies the canonical aggregation rule, so that recognizing an evidential problem suggests the Bayesian form even where practitioners lack the vocabulary — while the prime remains broader than any one rule, since Dempster-Shafer weighing, legal balancing, and narrative coherence in history all instantiate the same role-set with different aggregation functions. In every transfer the practitioner runs the same diagnosis — identify the trace, its producing mechanism, its custody chain, the inference and combination rules, and the live defeaters — and the transfer holds because none of these roles is tied to a substrate: a bridge inspector reading a hairline crack and an intelligence analyst weighing two intercepts are performing the same structural inference, distinguished only by the material of the trace.
Examples¶
Formal/abstract¶
A diagnostic medical test interpreted by Bayes' rule is the cleanest formal instance, instantiating the inference and aggregation roles quantitatively. The unobservable state is whether the patient has a disease \(D\); the trace-producing mechanism is the assay, characterized by its sensitivity \(P(+\mid D)\) and specificity \(P(-\mid \neg D)\); the trace is the positive result. The inference rule is Bayes' theorem: the posterior \(P(D\mid +) = \frac{P(+\mid D)\,P(D)}{P(+)}\), which makes the prime's defeasibility/weight-bearing invariant exact — a positive test does not deduce disease, it shifts a prior to a graded posterior. The famous base-rate result is the prime's claim made vivid: for a rare disease (low prior \(P(D)\)), even a highly sensitive and specific test yields a posterior well below certainty, because most positives are false positives drawn from the large healthy population. The aggregation rule's non-arithmetic, dependence-sensitive character is precise here too: a second independent test (different mechanism, e.g. imaging versus blood assay) multiplies the likelihood ratios and sharply raises the posterior, whereas re-running the same assay — a common-source trace — adds little, because its errors are correlated with the first. The defeaters are explicit: a confounder that inflates the marker, a mislabeled sample (a custody break), an alternative condition producing the same trace. The reasoning-about-next-evidence move is the clinician's: order the test whose likelihood ratio most discriminates between the live hypotheses given the current posterior.
Mapped back: The diagnostic test instantiates every role — target hypothesis (disease), trace-producing mechanism (the assay with its sensitivity/specificity), trace (the result), defeasible inference (Bayesian update), dependence-aware aggregation (independent confirmatory test versus correlated re-test), and defeaters — and the base-rate and correlated-trace results are the prime's weight-bearing and non-arithmetic-combination invariants in exact form.
Applied/industry¶
Software debugging shows the identical role-set in an engineering substrate, with no epistemological vocabulary required. The unobservable state is the hypothesis about a defect — say, "a race condition in the cache-eviction path corrupts the record." The trace-producing mechanisms are the running system and the engineer's instrumentation: logging, a debugger, a stress harness. The traces are a stack trace, a log line, a failing reproducer. The chain of custody matters exactly as the prime says: a stack trace from an optimized build with inlined frames, or a log from a machine whose clock was skewed, has a broken provenance that defeats the inference — which is why engineers pin builds and record environments. The inference rule is defeasible: a reproducer raises confidence in the race hypothesis but does not prove it; a flaky reproducer (the prime's canonical defeater) that fails to reproduce on retry undercuts the weight. The aggregation invariant appears in the warning against correlated traces: two failures observed on the same misconfigured CI node are not independent corroboration, just as two news outlets citing one source are not. The reasoning-about-next-evidence move is bisection — choosing the commit or the instrumentation point that most discriminates between the live defect hypotheses, the active-learning instance of "collect the trace that best splits the posterior." The transfer the prime promises is literal: chain-of-custody discipline imported from law becomes data-lineage instrumentation (attest each hand-off in the pipeline), and adversarial examination becomes the pre-mortem or red-team that surfaces defeaters before the fix is shipped.
Mapped back: Debugging runs the prime end-to-end — defect hypothesis, instrumentation as trace-producing mechanism, traces with auditable custody (pinned builds), defeasible inference, dependence-aware aggregation (correlated CI failures), and defeaters (flaky reproducers) — and demonstrates the cross-domain transfer: the bridge inspector, the clinician, and the engineer run one structural inference from trace to hidden state, distinguished only by the material of the trace.
Structural Tensions¶
T1 — Data versus Evidence-for-a-Hypothesis (Relational Scope). A trace is evidence only relative to a specific hypothesis; "having data" is not "having evidence." The tension is that traces are collected before the hypothesis is fixed, tempting their reuse for whatever claim is convenient. The failure mode is hypothesis drift: gathering data, then finding the claim it best supports — the structure of HARKing and post-hoc storytelling — so the same trace appears to confirm a hypothesis it was never a likelihood-discriminating test of. Diagnostic: ask whether the hypothesis was specified before the trace was examined; if the claim was chosen to fit the data, the trace's evidential weight for it is illusory, because it was never more probable under that claim than its rivals.
T2 — Probative Weight versus Admissibility (Two Gates). Evidence must clear two distinct gates — how strongly it bears on the claim (probative) and whether the procedure permits its use (admissible) — and they are routinely blurred. The failure mode is conflating force with permission: discarding a highly probative trace because of a procedural defect as if it were uninformative, or admitting a procedurally clean trace as if admissibility implied weight. Diagnostic: ask the two questions separately — "how much does this shift my confidence?" and "may I use it under my rules?"; a trace can be strong but inadmissible, or admissible but weak, and treating either gate as the other mis-weights the conclusion.
T3 — Independent versus Correlated Traces (Aggregation Dependence). Evidence combines non-arithmetically: independent corroboration raises confidence sharply, but common-source traces raise it far less than their count suggests. The failure mode is double-counting a single source: treating two outlets citing one wire report, two tests of one biomarker, or two CI failures on one bad node as independent confirmation, inflating confidence on what is structurally one trace. Diagnostic: ask whether the traces share a producing mechanism or upstream source; if they do, their errors are correlated and the aggregation rule must discount them — counting correlated traces as independent is the most common route to false certainty.
T4 — Absence of Evidence versus Evidence of Absence (Sign Asymmetry). A failed search is evidence against a claim only when a positive finding was likely had the claim been true; otherwise it is uninformative. The tension is the asymmetry between these, which intuition collapses. The failure mode is symmetric treatment: reading "we found nothing" as disconfirmation regardless of whether the search could have found something, or dismissing a genuinely powerful null result as "just absence of evidence." Diagnostic: ask how probable the positive trace was under the hypothesis given the search's power; a null from a search that would likely have detected the effect is real evidence against, while a null from an underpowered search carries almost no weight, and the two must not be equated.
T5 — Chain of Custody versus Trace Content (Provenance Coupling). The trace's content and its provenance are separable handles, and a probative trace with a broken custody chain is defeated regardless of what it shows. The tension is that attention flows to the vivid content while the unglamorous lineage is where inferences actually break. The failure mode is content fixation: reasoning from what a trace appears to show while ignoring that its path from event to presentation is unrecorded or compromised (a skewed clock, an inlined build, an unverified source). Diagnostic: before weighing what the trace says, audit how it reached you; if any hand-off from the underlying event to the presented trace is undocumented or breakable, the provenance defeater applies and the content's apparent force is suspended until the chain is repaired.
T6 — Defeasible Weight versus Threshold for Action (Standard-of-Proof Locus). Evidence only ever licenses graded confidence, but action requires crossing a threshold, and the threshold is a separate, often implicit, choice. The tension is between the continuous weight evidence supplies and the binary decision it must feed. The failure mode is implicit-bar reasoning: treating "the evidence points this way" as automatically sufficient to act, with no explicit standard of proof, so the threshold floats to whatever the advocate prefers. Diagnostic: ask what confidence level is required to act here and whether the evidence clears it — make the standard explicit and contestable (as law does with "beyond reasonable doubt" versus "preponderance"); a conclusion presented as action-ready without a named threshold has hidden the most consequential choice in the inference.
Structural–Framed Character¶
Evidence sits on the framed side of the structural–framed spectrum, with a frontmatter aggregate of 0.6 driven above the midline by its inseparability from inquiry as a practice. The relational core is real and richly articulated — a target hypothesis, a trace-producing mechanism, a trace, a custody chain, an inference rule, an aggregation rule, and a set of defeaters — and that role-set is genuinely medium-neutral, which is why a clinician, a bridge inspector, and a software engineer run the same trace-to-hypothesis inference. But what tips the prime framed is that the whole relation only makes sense relative to an agent conducting inquiry.
The decisive criterion is human-practice-boundedness, scored the full 1.0: evidence is inherently an inquiry-practice category. A trace is evidence only relative to a hypothesis being entertained and a question being asked, so the relation presupposes an inquirer — there is no evidence-for-a-claim in a physical substrate that asks nothing. That single maxed criterion is the heaviest contributor to the 0.6. The remaining four each sit at 0.5, recording genuine but partial framing. The vocabulary half-travels (vocab_travels 0.5): the role-set ports from law to science to debugging, but the epistemic idiom — probative, defeasible, burden of proof — travels with it. It carries epistemic-normative load (evaluative_weight 0.5): standards of proof, admissibility, and burden are normatively charged, though the bare likelihood relation (more probable if the claim is true than if false) is neutral. Its origin is partly institutional (institutional_origin 0.5): epistemology is formal, but the legal rules of evidence are a concrete institutional history the term carries. And invoking it half-imports (import_vs_recognize 0.5): one partly recognizes a trace-to-hypothesis relation, partly imports the inquiry frame of weighing and contesting.
The honest reading is a prime whose role-set is substrate-neutral enough to support real cross-domain transfer, but whose existence presupposes the practice of inquiry — the maxed human-practice score, with the four supporting halves, correctly places it on the framed side without pushing it to the institutional extreme of certification.
Substrate Independence¶
Evidence is a strongly substrate-independent prime — composite 4 / 5 on the substrate-independence scale. Its breadth is at the ceiling (domain breadth 5): the trace-to-hypothesis relation — with its provenance, defeasibility, and non-arithmetic combination — recurs with the same force across law (exhibits, testimony, standards of proof), science (data weighed by study design and replication), medicine (signs and symptoms in a differential), history (primary sources with internal and external criticism), journalism and intelligence (sources and corroboration as tradecraft), software debugging (stack traces and reproducers, with flaky reproducers as defeaters), and archaeology (remains in stratigraphic context). Transfer evidence is likewise at the ceiling (5): the role-set ports so cleanly that a clinician, a bridge inspector, and a software engineer run the same trace-to-hypothesis inference, and the chain-of-custody, defeater, and aggregation handles carry unchanged. Structural abstraction is a notch lower at 4 because the relation, though medium-neutral in its role-set, is inherently an inquiry-practice category: a trace is evidence only relative to a hypothesis being entertained and a question being asked, so it presupposes an inquirer and has no purely physical substrate that asks nothing. The composite of 4 records a role-set recognized across nearly every discipline that reasons from traces, held just short of 5 by its dependence on the practice of inquiry — the same human-practice centroid that places it on the framed side.
- Composite substrate independence — 4 / 5
- Domain breadth — 5 / 5
- Structural abstraction — 4 / 5
- Transfer evidence — 5 / 5
Relationships to Other Primes¶
Foundational — no parent edges in the catalog.
Children (1) — more specific cases that build on this
-
Provenance decompose Evidence
Provenance (the documented custody chain) is one COMPONENT of the evidence relation — the handle that defeats inferences when broken. Provenance is broader than evidence (art, data lineage), so this is a part-of, not a reparent of provenance.
Neighborhood in Abstraction Space¶
Evidence sits among the more crowded primes in the catalog (15th percentile for distinctiveness): several abstractions describe nearly the same structure, so a description that fits it will tend to fit its neighbors too — transporting it usually means disambiguating within this family rather than landing on it exactly.
Family — Inference & Evidence (26 primes)
Nearest neighbors
- Proof By Contradiction — 0.74
- Falsifiability — 0.74
- Provenance — 0.74
- Bijectivity — 0.73
- Eventual Consistency — 0.73
Computed from structural-signature embeddings · 2026-06-14
Not to Be Confused With¶
The nearest neighbor (similarity 0.93) is provenance, and the relation is part/whole: provenance is one component of the evidence relation, not a synonym for it. Provenance is the chain of custody — the traceable path from the underlying event to the trace to its presentation — and it is the handle that defeats an inference when broken (a skewed clock, an inlined build, an unverified source). But evidence is the full role-set: target hypothesis, trace-producing mechanism, trace, inference rule, aggregation rule, and defeaters, of which provenance is just the custody sub-relation. A trace can have impeccable provenance and still be weak evidence (it bears little on the hypothesis), or strong content yet be defeated by a custody break. Conflating the two collapses the separable "how strongly does this bear on the claim?" question into the "where did this come from?" question, when the prime's whole value is keeping the handles distinct so a dispute can localize to one of them. Provenance answers lineage; evidence answers the entire trace-to-hypothesis inference, lineage included.
A second genuine confusion is with the inductive_reasoning / abductive_reasoning / bayesian_updating family — the inference rules that operate on evidence. These name how one moves from traces to graded confidence: induction generalizes from instances, abduction infers the best explanation, Bayesian updating multiplies likelihood ratios into a posterior. Evidence is not any of these rules; it is the relation between trace and hypothesis that the rules consume and combine. The distinction matters because the same evidence can be processed by different rules (legal balancing, Dempster-Shafer weighing, Bayesian updating, narrative coherence) yielding different combination behavior, and because errors localize differently: a wrong inference rule is a different defect from a defeated trace or a broken custody chain. Treating evidence as identical to Bayesian updating, in particular, over-narrows the prime — evidence is broader than any one aggregation function, and the role-set survives even where practitioners lack the Bayesian vocabulary. The rule is the operation; evidence is the operand.
A third confusion is with belief_formation. Belief is the doxastic output — the state of holding a hypothesis as true; evidence is the input that licenses graded confidence in it. The two are routinely run together ("the evidence" treated as identical to "what we believe"), but they dissociate in both directions: one can possess strong evidence and withhold belief (suspending judgment pending a defeater check), or hold a belief with no evidence at all (faith, bias, motivated reasoning). The prime insists that evidence only ever licenses weight, never compels belief, and that crossing from weight to a held conclusion requires a separate, often implicit, threshold choice. Confusing evidence with belief hides that threshold — the most consequential and most contestable step — by treating the trace's weight as if it automatically settled the doxastic question. Evidence shifts confidence; belief is the state confidence may or may not be allowed to produce.
For a practitioner these distinctions make inference auditable. Confusing evidence with provenance reduces a multi-handle relation to a custody check, missing weak-but-clean and strong-but-defeated traces. Confusing it with the inference rules conflates the operand with the operation, mislocating defects. Confusing it with belief hides the standard-of-proof choice that turns graded weight into action. The unifying discipline is the prime's handle-by-handle audit: name the hypothesis, the trace-producing mechanism, the custody chain, the inference and combination rules, the live defeaters, and the threshold — and contest each separately rather than arguing about "the evidence" as an undifferentiated whole.
Solution Archetypes¶
No catalogued solution archetypes reference this prime yet.