Agreement is cheap. Independence is evidence.
Corroboration occurs when genuinely independent evidence supports the same bounded claim. Triangulation strengthens interpretation by approaching that claim through different sources, methods, observation modes or times—without hiding divergence.
Multiple records are not necessarily multiple evidence lines.
Corroboration requires agreement plus informational independence. Triangulation adds diversity of access, method, measurement or time so one hidden failure is less likely to explain the entire result.
Count independent generating paths, not citations, domains, mentions or search results. Then test whether those paths converge on the same claim within compatible scopes.
Two sources can agree because one copied the other, both used the same dataset, both repeated a press release or both inherited the same measurement error. Agreement becomes corroborative only after shared dependencies are identified and bounded.
Four sources can mean one, two or four evidence lines.
Switch scenarios to see how lineage, method and time change the evidential meaning of apparent agreement. The visual count and the independent-line count are deliberately kept separate.
Independence has more than one failure point.
Two records can have separate publishers yet share data, method or institutional control. A robust fingerprint tests each dependency rather than applying a single yes/no label.
Origin independence
Were the observations generated by different upstream sources rather than republished from one account?
Dataset independence
Do sources observe separate data, or do they analyze the same hidden or licensed dataset?
Method independence
Could one shared instrument, extraction rule or model reproduce the same error across sources?
Decision independence
Are the sources governed, funded or editorially controlled by the same actor or incentive?
Triangulation diversifies the path to the claim.
Source count alone increases repetition. Strong triangulation deliberately changes at least two dimensions while holding the claim boundary stable.
Different actors or systems with separately resolved lineages observe the claim.
Direct observation, documentary evidence, structured data and expert analysis test different failure modes.
Repeated observations test whether support is stable, episodic, decaying or superseded.
The proposition, scope and terms must remain compatible or apparent convergence becomes a category error.
Agreement and independence are separate coordinates.
The two-by-two model prevents duplicate consensus from being mistaken for corroboration and prevents genuine disagreement from being hidden.
Corroborative convergence
Independent evidence lines support compatible versions of the same bounded claim.
Repetition
Several records agree because they share an upstream source, dataset, method or control.
Material divergence
Independent observations conflict. Scope, time, method and source quality require diagnosis.
Unstable lineage
Versions from the same evidence family contradict one another or mutate across republication.
Confirmation strength depends on what remains shared.
The matrix keeps source, input, method, time and control dependencies visible. “Independent” is never inferred from different domain names alone.
All trace to U1.
Same reported facts.
No new observation.
Publication lag only.
Agreement without corroboration.
Different origin paths.
Same target system.
Structured vs direct.
Compatible window.
Shared target is intentional.
Distinct publications.
Same upstream data.
Different analysis rules.
Same data period.
One data error can affect all.
New acquisition each time.
Expected dependency.
Method error persists.
T0, T1 and T2.
Stability, not method diversity.
Corroborate the proposition—not the surrounding narrative.
Each set keeps its claim narrow and states exactly which independent path contributes new support.
A result state appeared for a query
A structured API response and a separately captured browser observation agree on the URL, position, locale and time window.
An organization controls a domain
The organization’s own declaration is matched with registry data and a technically observed domain relation.
A source is repeatedly cited
Separate runs, prompts and retrieval contexts show repeated inclusion, while preserved citation targets verify that the system points to the same source identity.
Twenty pages repeat one statistic
All citations terminate at one undated report, and several pages copied each other without inspecting the original dataset.
Consensus can be manufactured by the network itself.
These patterns create the appearance of confirmation without adding an independent observation.
Republished copies are counted as independent sources.
Later articles cite summaries that ultimately point to one assertion.
Different analysts inherit the same undisclosed sampling or collection error.
Separate sources use one instrument, model or rule with the same bias.
Simultaneous observations capture one temporary event and are generalized as stable.
Sources agree on narrower claims but are cited for a broader proposition.
Source count replaces inspection of competence, access and provenance.
Conflicting observations are removed until only agreement remains.
Six controls turn source count into evidence structure.
The protocol reports both convergence and dependence, then routes unresolved disagreement forward rather than smoothing it away.
Bind the claim
Fix proposition, scope, entities, units and observation window.
Resolve provenance
Trace every record to its originating source, dataset and version.
Collapse dependencies
Group syndicated, derived and common-input records into evidence families.
Compare observations
Test compatible claims, terms, timing and result direction.
Triangulate gaps
Add an observation path that changes source, method or time.
Report the state
State independent count, shared limits, convergence and divergence.
Define the claim before counting support.
Corroboration can be established at the level of an exact proposition, not at the level of a document, brand article or whole narrative. A source may support one proposition, remain silent on another and contradict a third. Claim-level separation prevents surrounding agreement from being used to decorate an unsupported conclusion.
Resolve names, aliases, subsidiaries, products, domains and historical versions. Independent about one corporate group does not automatically support claims about every brand or product it controls.
“Uses,” “owns,” “partners with,” “mentions” and “depends on” are different relations. Similar vocabulary must not collapse them into one proposition. Preserve modality and negation.
Record the value together with its unit, denominator, category rule and tolerance. Two sources can agree directionally while disagreeing materially on magnitude.
Geography, audience, device, market segment, query form and inclusion rules define where support applies. Narrow evidence supports a narrow claim unless an additional inference justifies extension.
Separate publication, capture and effective dates. Several records may agree historically but fail to establish the current state of a volatile subject.
Direct observation, derived estimate and expert interpretation provide different support relations. Agreement between weak inferences does not become direct proof through repetition.
corroboration unit = normalized claim + declared boundary + evidence lines
A document count answers how many containers were found. An evidence-line count answers how many sufficiently independent observation paths support the same bounded proposition. These are different quantities and must remain visibly separate.
Different URLs may lead back to the same observation.
Lineage reconstruction identifies where information originated, how it moved and whether later records add new evidence or only new presentation. The goal is not to punish citation; it is to avoid confusing distribution with independent confirmation.
Extract the claim
Capture the exact sentence, number, table entry or observed state that appears to support the proposition.
Follow citations
Trace links, footnotes, quoted experts, embedded charts and named datasets to the earliest recoverable evidence record.
Compare wording
Repeated phrasing, identical rounding and shared errors can expose copying even when an explicit citation is absent.
Compare data
Inspect sample, timestamps, categories, missing values and transformation rules to determine whether two outputs share an upstream dataset.
Map control
Identify common ownership, funding, editorial control, infrastructure or analytical authorship that could create correlated error.
Collapse echoes
Group derivative records into one lineage for corroboration counting while preserving each record’s distribution role.
12 publications repeat the same statistic.
Twelve retrieval results may improve discoverability and show how a claim spread. They do not establish twelve independent measurements.
One survey generated the statistic.
After lineage collapse, the claim has one originating evidence line plus eleven derivative distribution records. Its weight depends on the survey, not the echo count.
Independence is multidimensional—not a publisher checkbox.
Two sources can be independent in authorship while dependent in data, method or institutional incentives. Record each dimension separately. The weakest shared dependency may explain the entire pattern of agreement.
Did each evidence line collect or generate its own record, or does one line derive from another? Independent publication is not independent origin.
Two analyses of the same dataset can test analytical robustness, but they do not provide independent sampling of the underlying reality.
Survey, transaction record, direct capture and registry document fail differently. Method diversity is valuable when the methods genuinely observe the same claim.
Different dashboards can depend on the same sensor, crawler, classification model or upstream provider. Interface diversity can conceal instrument dependence.
Repeated captures can show persistence or change. They add temporal evidence, although a stable systematic error may remain shared across every capture.
Common ownership, funding or editorial approval can align what is measured, published or omitted even when individual authors differ.
Independent organizations may share commercial, political or reputational incentives. Incentive alignment does not invalidate evidence, but it is a possible correlation path.
Several summaries can inherit one expert’s classification or causal interpretation. Diversity at the observation layer may coexist with dependence at the inference layer.
Change the observation path while holding the proposition stable.
Triangulation is deliberate diversification. Each additional line should reduce reliance on a plausible shared failure. Random source accumulation creates volume; designed triangulation creates diagnostic strength.
Independent collectors observe the same bounded state
Use separate origin paths with access to the same subject. Compare collection dates, definitions and scope before treating agreement as convergence. This strategy reduces dependence on one publisher or originating record but may retain a shared method or target-system error.
Different methods reach the same proposition
Combine direct observation, structured records, documentary evidence and appropriately bounded expert interpretation. Agreement is stronger when no single method failure explains all lines. Confirm that each method actually supports the same claim rather than a neighboring construct.
Repeated captures test stability and transition
Observe the same claim across declared windows. Stable agreement supports persistence; change reveals a time-bound state. Temporal triangulation does not remove systematic method bias, so pair it with another method when the measurement system itself is uncertain.
Different positions expose different parts of the system
Operator records, customer observations, regulatory filings and infrastructure telemetry may each reveal a partial state. Convergence across positions can strengthen a shared proposition, while differences may expose boundaries rather than error.
Count the genuinely new information inside each apparent confirmation.
These cases show how the same discipline operates across search, entities, markets, incidents, AI retrieval and digital assets.
Three tools report position four
All three interfaces may rely on separate captures, or two may resell one upstream dataset. Align query, location, language, device and time; then trace acquisition lineage. If independent captures agree within one window, the rank state is corroborated for that environment—not globally or permanently.
Registry, official site and trade publication agree
The registry independently supports legal name and incorporation. The official site supports current self-description. A trade article may simply quote that site. The final entity profile should map each field to the evidence line that can observe it rather than treating all three sources as equal support for every attribute.
Two reports estimate 18 percent
Matching values do not establish independence. Both reports may purchase the same panel and apply similar category boundaries. Inspect provider, panel, denominator and transformation. Independent transaction data with a compatible market definition would add a more diverse path than a third report built on the same panel.
Status page, users and telemetry report disruption
These perspectives can triangulate occurrence, duration and impact. The status page has operator access, users observe functional consequences and telemetry observes network behavior. Agreement supports a bounded incident state, while severity and root cause require separate evidence.
Two indexes discover the same referring page
Independent crawlers can corroborate that a link was observable, provided the canonical source, target, link state and capture period align. A live direct fetch adds method diversity. Agreement does not prove the link was continuously present or that it caused ranking movement.
Several models cite the same source
Model outputs are separate runs but may share web indexes, training material or ranking signals. Preserve prompt, time, interface and citations. Cross-model agreement can show repeated source selection; it does not independently verify the cited claim unless the underlying source and proposition are validated.
Documentation and live interface agree
Documentation states availability, while a controlled live observation confirms the feature under one plan, region and account type. The methods are different and mutually reinforcing. The conclusion must retain those conditions because availability may differ across tiers or deployments.
Search results, user tasks and corpus audit converge
A SERP sample shows recurring subtopics, user research identifies unresolved tasks and a page-level audit finds missing evidence units. These methods observe different surfaces. Convergence supports a bounded coverage gap more strongly than keyword repetition alone, while priority still depends on audience and decision objectives.
More evidence is useful only when it changes the justified state.
Set the required degree of corroboration in proportion to consequence, reversibility, volatility and error cost. Low-impact descriptive claims may need modest confirmation; irreversible or safety-critical decisions require stronger, more diverse and more current evidence.
One relevant evidence line supports the claim. Report its provenance and limits. Do not imply consensus, but do not discard direct evidence merely because a duplicate has not been found.
The claim is widely repeated but still depends on one origin, dataset or interpretation. This is dissemination evidence, not independent confirmation.
At least two sufficiently independent lines support the same bounded proposition. Shared assumptions and remaining uncertainty stay explicit.
Independent origins or methods reach the same proposition through meaningfully diverse observation paths. Confidence rises because one plausible failure is less able to explain all support.
Agreement is incomplete or a credible contradiction remains. Preserve both sides and continue with conflict diagnosis rather than hiding the minority line.
Past convergence does not guarantee present validity. Volatile states require temporal review, new captures or an explicit historical boundary.
Make convergence inspectable after the analysis is finished.
A reusable record explains which evidence lines were counted, why they were considered independent, how the claim was normalized and what limits survived. It allows later evidence to update the conclusion without reconstructing the investigation from memory.
Entity, relation, value, scope, time and qualifiers, plus the original wording from each source.
Origins, methods, timestamps and preserved records, with derivative publications grouped under their upstream line.
Common datasets, instruments, ownership, incentives, definitions and target-system dependencies that remain after diversification.
Which parts of the claim agree, permissible tolerance and any dimensions where values remain merely compatible rather than equal.
Uncorroborated, repeated, corroborated, triangulated, contested or temporally decayed, with a short justification.
Evidence that would strengthen, weaken, contradict, supersede or require re-collection of the current claim state.
Corroboration and triangulation, without source-count mythology.
Concise answers to the distinctions that most often determine whether apparent agreement adds real evidence.
How many sources are required for corroboration?
There is no universal count. Two genuinely independent and claim-fit lines may provide stronger corroboration than twenty derivative articles. The required strength also depends on consequence, reversibility, volatility and the known error of each method.
Are primary sources always independent?
No. Two primary reports may share one dataset, instrument, funding structure or classification method. “Primary” describes proximity to origin; independence describes whether the error paths are sufficiently separate.
Does convergence prove a claim is true?
No. Convergence raises justified confidence when independent, compatible paths agree, but all paths may still share an unknown assumption or observe the wrong construct. State the remaining common dependencies.
What makes method triangulation strong?
The methods must have different relevant failure modes while supporting the same normalized proposition. Using several interfaces backed by one acquisition system is not meaningful method diversity.
Can repeated captures corroborate a changing state?
They can corroborate persistence within an interval or reveal a transition. Each capture remains time-bound. Repetition across time does not support an unqualified claim that the state is permanent.
What if independent sources disagree?
Do not average or vote automatically. Align identity, definition, scope, time, unit, method and lineage. If material disagreement remains, classify the claim as contested and move to explicit conflict resolution.
Do several AI answers count as independent sources?
Not by default. Models may share training material, retrieval indexes, cited sources or ranking signals. Treat outputs as separate run records, trace their cited evidence and avoid inferring independence from different interfaces alone.
What should a corroboration conclusion say?
Name the bounded claim, number of visible records, number of independent evidence families, dimensions diversified, shared dependencies, disagreement state, confidence and the event that would trigger review.
The claim type determines what independent support must look like.
Corroboration is strongest when collection is designed around the failure modes of the proposition. These plans show why one generic “find three sources” rule cannot protect every type of conclusion.
Legal identity and public identity require different evidence
For legal name, incorporation and registered office, prefer an authoritative registry record plus an independently preserved version where possible. For current public identity, compare the official site, current product interface and independent references. Map parent, subsidiary, brand and predecessor relations rather than merging them. Stop when each required attribute has a claim-fit evidence line and no unresolved entity collision remains.
Volatile states need independent captures and a short validity window
A current price, ranking, feature, policy or availability claim can expire quickly. Use at least one direct capture with complete environment details and a second independent observation when consequence justifies it. Preserve time, location, device, account state and version. Agreement supports the state only within the declared window. Re-collect when volatility or decision delay makes the evidence stale.
Matching numbers are weak when denominator and model are shared
For search demand, traffic, market share, audience or valuation estimates, reconstruct construct, population, period, unit, denominator and transformation. Seek a second dataset or measurement model with a meaningfully different error path. Report compatible ranges when exact values cannot be observed directly. Stop when uncertainty is narrow enough for the intended decision—not when several dashboards display similar precision.
Repeated association does not independently establish causation
Several observational reports can share the same confounding structure. Corroborate causal claims through designs that test alternative explanations: controlled comparison, natural experiment, temporal ordering, mechanism evidence or negative controls. Align the intervention and outcome definitions. Stop only when plausible rival explanations have been reduced to a level proportionate to the consequence of acting on the claim.
Contemporaneous records and later interpretation play different roles
Use records produced near the event to establish what was documented at the time, while recognizing access and incentive limits. Later independent research can triangulate context, identity and consequence but may inherit earlier archives. Separate primary record, witness account, institutional record and retrospective analysis. Stop when the bounded event claim is supported across appropriate paths and material contradictions are disclosed.
Corroborate the run separately from the factual content
Repeated outputs can establish that several controlled runs produced a pattern, such as citation frequency or answer stability. They cannot establish that the generated factual claims are true. Preserve prompt, model or interface, time, location, session context and cited sources. Then validate each material claim against independent evidence. Stop rules must distinguish output reproducibility from factual corroboration.
Failure to observe is not automatically evidence of absence
To support “no record,” “not available” or “does not occur,” define where the item would be expected to appear, how thoroughly that space was observed and the detection limits. Diversify search paths, indexes, time windows or access positions. A bounded absence claim may be justified even when universal absence is not. Stop when coverage and detection sensitivity match the stated boundary.
Independent forecasts may still share assumptions
Compare data lineage, model family, scenario assumptions and forecast horizon. Several forecasts based on the same growth rate or macroeconomic premise form a dependent cluster. Use structurally different models and scenario stress tests, then track calibration against outcomes. Stop when the range and downside are decision-ready; do not turn cross-model agreement into certainty about an unobserved future.
Before declaring corroboration, attempt to break it.
A short adversarial review tests whether the confirmation survives lineage, scope and method challenges. Passing these controls does not create certainty; it shows that the stated confidence is proportionate to the inspected evidence.
Recalculate support after mirrors, quotations, press-release copies and shared datasets are grouped. If the conclusion loses most of its support, report repetition rather than corroboration.
Remove unsupported geography, populations, periods and universal language. The conclusion should remain useful even after its real boundary becomes visible.
Look for common definitions, target-system behavior, incentives, acquisition infrastructure and hidden assumptions. If one plausible failure explains all agreement, triangulation is incomplete.
Do not test only whether support exists. Identify credible disagreement and explain whether it reflects scope, time, method or a material unresolved conflict.
A reversible low-impact choice and an irreversible high-impact commitment should not share the same evidence threshold. Increase diversity and currency as consequence rises.
Name the new record, observed change, failed prediction or elapsed validity window that would trigger reassessment. A conclusion without an update rule becomes resistant to evidence.
Release corroboration only when the evidence count, independent-line count and dependency map are all visible.
The public conclusion should state the exact proposition, applicable boundary, degree of convergence, surviving shared assumptions and current review condition. This makes agreement useful without converting it into certainty and lets future evidence strengthen, narrow or contest the claim without erasing its history.
Corroboration maps convergence. Conflict analysis owns divergence.
EVD/06 establishes whether support is independent and multi-modal. EVD/07 diagnoses material disagreement through scope, timing, lineage, method and source quality.
Digital records, observations and claim-specific support.
EVD / 02CLASSEvidence Types & ClassesOrigin, directness, form, independence and temporal state.
EVD / 03ORIGINSource ProvenanceIdentity, custody, version, authorship and transformation history.
EVD / 04QUALITYSource QualityCompetence, access, transparency, incentives and accountability.
EVD / 05CAPTUREEvidence Collection & PreservationAcquisition context, stable records, snapshots and chain of custody.
EVD / 06CONFIRMCorroboration & TriangulationIndependent agreement across sources and observation modes.
EVD / 07CONFLICTConflicting Evidence ResolutionDiagnosing disagreement through scope, timing and lineage.
EVD / 08TIMETemporal Validity & Evidence DecayFreshness windows, volatility, supersession and re-collection.
EVD / 09SUPPORTClaim–Evidence MappingConnecting observations to exact claims and inference boundaries.
EVD / 10CONFIDENCEConfidence CalibrationTransparent confidence states without false certainty.
EVD / 11UNKNOWNEvidence Gaps & UnknownsMissing observations and unresolved alternatives.
EVD / 12SYNTHESISEvidence Synthesis & Decision ReadinessCombining support, conflict and uncertainty into a decision state.