Every claim has a lineage.
Source provenance is the traceable history of an evidence record: who or what created it, where it originated, how it was acquired, which version was observed, who handled it and what transformations produced the current representation.
Provenance resolves the record. It does not certify the claim.
A complete provenance trail makes an evidence object inspectable and reproducible. It tells us which record exists and how the present representation came to be. Whether that record is accurate, competent or sufficient remains a separate evaluation.
Provenance is the documented continuity between a source event, a captured record, every material transformation and the claim that later uses it.
Its unit is not merely a URL. A URL is a location. Provenance requires stable record identity, origin attribution, acquisition context, temporal and version identity, derivation history and a declared use boundary. When one of those relations is unknown, the uncertainty must remain visible.
Follow the record from source event to claim.
Select a node to inspect its responsibility. The chain preserves a distinction between the event, the raw capture and every derived representation; the branches expose copies that can look independent while sharing one upstream origin.
Same record, stable preservation reference.
New location, shared upstream content lineage.
No recoverable author, version or origin relation.
One record can involve six different actors.
Collapsing these roles into a generic “source” destroys information. The actor registry assigns each responsibility to the entity that actually performed it.
Creator
The person, organization or system that produced the original content, measurement or event record.
Publisher
The entity that made a representation available. Publication does not necessarily establish authorship.
Collector
The actor or process that acquired the record and fixed the query, locale, time and access context.
Custodian
The entity responsible for retaining the record, integrity reference and access history after capture.
Transformer
The code, model or analyst that normalized, extracted, translated, aggregated or classified the record.
Claim author
The actor who mapped the record to a proposition and declared the permissible inference boundary.
Derived is acceptable. Undeclared transformation is not.
A useful provenance record preserves the raw input, names the transformation and exposes the difference. Switch states to inspect how lineage quality changes without pretending that a transformed field was directly observed.
Content lineage and custody lineage must remain distinct.
Content lineage traces intellectual or computational derivation. Custody lineage traces possession, storage and transfer. They intersect at identifiable records, but answering one does not automatically answer the other.
Different records require different provenance fields.
The minimum envelope is stable, but record-specific controls vary. “Required” means the field is necessary to resolve that type; “conditional” means it becomes necessary when transformation, versioning or delegation occurs.
Publisher, canonical URL and authorship when known.
Retrieved time, locale and access path.
Observed version or timestamped snapshot.
Stable archive or content hash.
Extraction, translation or cleanup.
Provider, product, endpoint and task.
Request parameters and acquisition time.
Response version and data period.
Raw payload reference and checksum.
Every normalized or computed field.
Product, model identity when exposed and run.
Prompt, settings, retrieval context and time.
Exact output and run-specific state.
Hash or immutable export reference.
Answer lineage differs from cited-claim lineage.
Input set identities and owners.
Selection boundary and collection state.
Input and output version pairing.
Immutable input references.
Code, parameters and derivation order.
Trace the exact object—not the generic source.
These examples show how provenance changes by evidence form. Each record retains its own identity and makes the inference boundary visible.
Structured search-result response
The provider-returned response is preserved separately from normalized fields and derived classifications. Provider data and analytical interpretation never share one unlabeled layer.
Published page snapshot
The current URL, original publisher, stated author, retrieval time and preserved snapshot are distinct fields. A later edit creates a new version relation rather than silently replacing the observed state.
AI answer with citations
The captured output proves what the system generated in that run. Each factual claim still inherits a separate provenance chain from the cited or retrieved source behind it.
Five URLs, one upstream statement
Republished copies may have different domains, dates and page identities while carrying the same originating report. Provenance collapses apparent multiplicity into one informational lineage.
A chain can fail without the file disappearing.
Broken provenance is not a single condition. The diagnostic must identify the missing relation so later confidence limits are attached to the correct uncertainty.
The record exists, but authorship or producing system cannot be resolved.
The host is known, but the document was copied from an untraced upstream source.
Query, locale, device, authentication state or retrieval time is absent.
The current page is treated as identical to an earlier state without a preserved snapshot.
Extraction, translation, aggregation or classification changed the record without declaration.
A quote or screenshot lacks the surrounding document and recoverable source relation.
Several publications refer to each other while none exposes an independent origin.
The record is traceable, but the exact proposition it is used to support is unstated.
Store the lineage as first-class evidence data.
The envelope separates observed source fields from research-system fields. It is deliberately explicit: unknown values remain unknown, transformations name their inputs and outputs, and claim use is always bounded.
{
"record_id": "evd-r3-017",
"record_type": "structured_serp_observation",
"source": {
"creator_or_system": "resolved entity",
"publisher_or_provider": "structured-search-provider",
"product_context": "declared endpoint / task"
},
"acquisition": {
"captured_at": "ISO-8601 timestamp",
"request_context": "query + locale + device",
"collector": "identified process"
},
"integrity": {
"raw_reference": "immutable object id",
"sha256": "7bf1…9d42"
},
"derivation": [{
"input": "raw:v01",
"operation": "normalize",
"version": "2.4",
"output": "derived:v03"
}],
"claim_use": {
"claim_id": "claim-042",
"support_scope": "recorded result state",
"does_not_establish": "ranking cause"
}
}Correction creates a new version and a declared relation.
Use an explicit unresolved state until the entity boundary is repaired.
Extraction, translation and classification are evidence events.
Traceability without a use boundary cannot justify an inference.
Returned observations remain distinct from derived analysis.
A citation points somewhere. Provenance explains what happened.
A defensible source record connects identity, origin, acquisition, version, integrity, transformation and claim use. Each field solves a different failure mode; together they make the evidence recoverable and interpretable.
URL, title, language and document identifier distinguish the page from the domain that hosts it.
OBJECT ≠ WEBSITENamed author, publishing entity and original publication location are recorded separately.
CREATOR ≠ HOSTTimestamp, access path, locale, authentication state and rendered or source representation.
HOW WAS IT SEEN?Published, modified and retrieved times are distinct; a later live page cannot replace the captured state.
FIX THE VERSIONSnapshot, archive identifier, content digest or exact retained copy protects the observed wording.
VERIFY THE RECORDQuoted passages, translations, summaries and structured fields link back to the exact page version.
TRACE EVERY CHANGETask, endpoint, product and returned object identifiers separate one observation from another.
PAIR INPUT + OUTPUTThe service and underlying observation product are named without confusing the analyst with the producer.
NAME THE ACTUAL PRODUCERQuery, market, language, device, filters and collection time define the environment represented.
CONTEXT IS EVIDENCEResponse schema, index state and covered period may differ from retrieval time.
THREE CLOCKS MAY EXISTAn immutable copy or digest permits later comparison with normalized fields.
RAW PRECEDES MODELRenamed, aggregated or computed values cite their raw inputs and transformation version.
OBSERVED ≠ COMPUTEDTitle, persistent identifier, edition and file identity establish which dataset was used.
RELEASE-SPECIFICData collector, sponsor, repository and publisher may be different entities with different roles.
SEPARATE RESPONSIBILITYPopulation, sampling frame, inclusion rules, instrument and field period define the observations.
HOW DID ROWS ENTER?Corrections, appended waves and revised weights create new derivation states.
DO NOT SILENTLY UPDATEChecksums, repository version and retained codebook bind analysis to exact bytes and meanings.
DATA + CODEBOOKFilters, exclusions, recoding, joins and weighting are recorded as a reproducible path.
SUBSET HAS LINEAGEProduct, conversation or run identifier and exact response distinguish one generation event.
OUTPUT IS THE OBJECTModel identity when exposed, interface and retrieval context belong to the origin record.
MODEL STATE MATTERSPrompt, language, account state, settings, region and time define the observable run.
PRESERVE INPUT CONDITIONSA repeated prompt can produce a different answer as indexes, models and policies change.
RUN ≠ PERMANENT STATEFull output and cited targets are retained rather than relying on a later regenerated answer.
CAPTURE, DO NOT RECREATEThe output lineage and the factual lineage behind each assertion are traced separately.
ANSWER ≠ SOURCE PROOFFollow the evidence until the first recoverable observation.
These cases show how a provenance trail changes with the evidence object. Each example identifies what must be preserved, where lineage commonly breaks and which claim remains defensible after the audit.
A page appears in position four for a defined query
The provenance object begins with the query environment: exact query, country or market, language, device, result type, observation time and the returned result record. The raw response or faithful capture remains distinct from the normalized position, canonical host and later category labels. A stored screenshot may preserve the visible page, while a structured response preserves machine-readable fields; neither representation should silently replace the other.
For the claim that a URL occupied a reported position in that environment, the chain can be direct and recoverable. For the claim that the URL “ranks fourth everywhere,” it is insufficient because location, personalization, index variation and time are omitted. For the claim that a particular factor caused the ranking, the observation is contextual rather than causal.
A repeated capture should receive a new record identity even when the query is unchanged. Aggregating daily observations into an average creates a derived dataset whose lineage points to every included capture and names the missing-data and weighting rules.
A company states that a product launched on a specific date
The company page is primary documentary evidence of the declaration. Provenance records the publishing entity, stated author or department, canonical location, publication date, retrieved version and exact wording. If the announcement is syndicated through media outlets, those pages become distribution records connected to one upstream statement rather than independent origins.
The announcement directly supports the claim that the company publicly said the launch occurred. It does not automatically verify product availability, market coverage, technical capability or customer access. Those claims require transaction records, independent observation, product tests or other appropriate evidence classes.
If the live page is updated from “will launch” to “launched,” a later screenshot cannot prove the earlier wording. Both states need their own preserved representations and a supersession relation. A reliable trail makes it possible to say not only what the page says now, but what was actually observed at each point in time.
Nineteen articles repeat the same 62 percent figure
A search finds nineteen publications with identical or slightly rounded wording. Provenance work follows citations backward, compares publication dates, inspects footnotes and searches distinctive phrases until the earliest recoverable dataset or report is identified. The nineteen URLs may demonstrate reach and repetition, but they do not establish nineteen independent measurements.
The originating survey record should expose the sample, questionnaire, field dates, denominator, weighting and sponsor. If only a press release survives, the lineage is partially resolved: the statement’s publisher is known, while the underlying measurement remains unavailable. The correct status is not “verified by nineteen sources,” but “widely repeated from one partially recoverable origin.”
When two publications use the same panel provider or purchased dataset, they may remain dependent even if they perform separate analyses. Provenance must trace both documentary citation and data lineage. Independence is a property of the information path, not the number of authors or domains.
Two indexes report the same referring page
Each index produces its own discovery record, including source URL, target URL, anchor, first-seen and last-seen fields. Yet the underlying web link is one documentary object. Provenance separates the referring page from the discovery systems and records whether the canonical page was inspected directly.
If the link disappears, historical index data can still support a claim that it was previously observed, provided the observation time and provider-specific field meaning are retained. It cannot prove that the link remained continuously present between two crawls. A cached or archived source page can strengthen the historical trail, while a current live page answers a different time-bound question.
Metrics derived from the backlink—authority scores, estimated value or toxicity labels—form separate computational records. Their lineage must name the index state and metric version. Combining raw link existence with a proprietary score in one field would obscure which part was observed and which part was modeled.
A status page changes after a service disruption
During an incident, the first status entry may say “investigating,” a later entry may identify an affected service and the final report may revise the duration or cause. Each update is a distinct documentary version connected through one incident identity. Preserving only the final page removes the chronology through which public understanding changed.
Operational telemetry, customer reports and status updates create parallel content lineages. They may converge on occurrence while disagreeing about start time, severity or cause. Custody lineage records who collected and retained each object; content lineage records how each later conclusion derived from those records.
A post-incident report may supersede an early causal hypothesis without making the initial update fraudulent. Provenance allows both to coexist with explicit version and correction relations. This prevents a current explanation from being projected backward as though it had been known at the first observation.
A generated answer cites a report that cites another article
The output record proves that the system produced the answer and displayed the citation in a specific run. The citation target is then inspected. If that page does not contain the supported claim, the output has a citation alignment failure even when the link is real. If the page quotes another report, lineage continues upstream until the originating observation or an explicit unresolved boundary is reached.
Different models may cite the same page because they share a retrieval environment or because the page dominates public distribution. Their agreement is therefore not automatically independent. Record each run, prompt, time and cited target, then group outputs by underlying source family before assessing corroboration.
A generated summary can transform the source through compression, inference or wording changes. The exact source passage and the model’s claim should remain side by side so the transformation is inspectable. Provenance does not declare the answer false or true; it reveals whether the asserted support can be followed.
A foreign-language document is translated and summarized
The original document remains the source record. The translation is a derived representation with its own creator, date, method and version. A later summary derives from the translation or, preferably, from the original plus a documented translation. Provenance must show which path was actually used.
Material terms, uncertainty markers, legal definitions and numerical qualifiers can change meaning in translation. Preserving sentence-level alignment allows reviewers to inspect whether “may,” “must,” “estimated” or “confirmed” survived the transformation. Machine-assisted and human-reviewed translations can both be legitimate when their process and limitations are declared.
Citing only the translated summary can make the final claim appear closer to the origin than it is. The defensible chain links final wording to summarized passage, translated passage and exact original text. When no competent review is available, language uncertainty should remain attached to the claim.
Three exports become one analytical table
A merged table is not a neutral container. Its lineage includes three exact input versions, schema mappings, join keys, duplicate handling, exclusions, timezone conversion and the code or procedure that produced the output. Each derived row should be traceable to one or more source rows when the research design permits.
Joining by domain name can collapse separate entities; joining by display name can create false matches; choosing the newest value can erase conflicting observations. Provenance records these operations so a surprising result can be diagnosed at the source, mapping or transformation layer.
If one input is later corrected, the derived table receives a new version. The previous analysis remains tied to the old input and should not silently inherit the correction. Reproducibility requires a frozen relationship between input identities, transformation version and output identity—not merely access to the newest files.
Resolve lineage in five passes. Stop where the evidence stops.
The audit moves from the exact object to its permissible claim. Each pass produces a recoverable output, allowing unresolved fields to become visible research tasks rather than hidden assumptions.
Fix the object
Identify the exact page, response, file, passage, image, row or output used. Assign a stable local identifier and preserve enough context to recover it.
Resolve origin
Separate creator, publisher, host, sponsor and collector. Follow citations and repeated wording toward the earliest recoverable production event.
Freeze context
Record acquisition method, parameters, access conditions, locale, time and version. Preserve the observed state rather than relying on a changing location.
Trace changes
Document extraction, cleaning, translation, normalization, joins, models and summaries. Bind each output to named inputs and transformation rules.
Bound the claim
State exactly what the record can support, which inference remains indirect and which unknowns propagate into confidence or require new collection.
“The plan costs €99 per month”
The live pricing page is captured with currency, billing period, tax treatment, market, plan name and retrieval time. Structured markup, visible text and checkout may disagree, so each becomes a separate observation. The defensible claim is tied to the state and customer context actually observed.
If a comparison article quotes the amount, it remains a secondary representation unless it provides an independent transaction record. A later price change creates a historical relationship; it does not invalidate a properly time-bounded statement about the earlier capture.
Two company names appear to describe one organization
Provenance traces registry identifiers, official domains, legal names, brand names, redirects and acquisition history. Similar wording or shared contact details may suggest a relationship but should not silently merge the entities.
Every assertion—same legal entity, parent-subsidiary relation, renamed brand or shared operator—requires the record that establishes it. The entity map remains explicit about resolved identity, inferred association and unresolved ambiguity.
Ask these before the evidence enters synthesis.
A provenance review is complete enough only when another reader can follow the record backward, understand every material change and see why it was permitted to support the stated claim.
What is the exact evidence object?
A domain, organization or database is too broad. Name the recoverable unit used: response, page version, paragraph, dataset release, row, chart, model output or interview segment.
PASS / OBJECT HAS A STABLE IDENTITYWho created it?
Distinguish original creator from host, publisher, distributor, sponsor and collector. When authorship is unknown, preserve that uncertainty instead of assigning the hosting domain as creator.
PASS / ACTOR ROLES ARE SEPARATEDWhere did the information originate?
Follow citations, identical phrasing, embedded charts, named datasets and earlier versions. Stop at the first recoverable observation or declare that the upstream origin remains unresolved.
PASS / ORIGIN OR BOUNDARY IS EXPLICITHow was it acquired?
Record request parameters, access path, instrument, query environment, selection logic and collection time. Acquisition context determines which environment the record represents.
PASS / CAPTURE CAN BE UNDERSTOODWhich version was observed?
Separate publication, modification, observation and retrieval times. Preserve a stable representation so later edits cannot silently rewrite the evidence used in the analysis.
PASS / VERSION IS FIXEDCan integrity be checked?
Use retained originals, archives, checksums, immutable identifiers or controlled exports according to the evidence type. Integrity confirms continuity of the record, not truth of its claims.
PASS / MATERIAL CHANGES ARE DETECTABLEWhat transformed the record?
Name every extraction, translation, cleanup, classification, aggregation, model and summary. Record parameters and input-output relationships at the level necessary to reproduce the result.
PASS / DERIVATION IS INSPECTABLEIs the path independent?
Different pages may share one report, dataset, crawler or editorial source. Group dependent records into lineages before counting corroboration or source diversity.
PASS / ORIGIN FAMILIES ARE KNOWNWhat can this record support?
Map the evidence object to an exact proposition. State whether support is direct, indirect or contextual, and name the inference that would be required to claim more.
PASS / USE DOES NOT EXCEED LINEAGESource provenance, answered precisely.
These answers resolve the distinctions that most often determine whether digital evidence can be traced, repeated and defended.
What is source provenance?
Source provenance is the documented history of an evidence record from its creation or originating event through acquisition, preservation, transformation and use in a claim. It identifies the object, actors, versions, custody and derivation relations.
Provenance makes evidence traceable. It does not by itself prove that the source is accurate, unbiased or sufficient.
Is a URL enough to establish provenance?
No. A URL identifies a location, and the content at that location may change, redirect or disappear. Provenance requires the exact observed version, retrieval time, publisher or creator, capture context and a stable representation when the wording matters.
The URL remains useful as one identity field, but it should not be confused with the preserved evidence object.
What is the difference between provenance and source quality?
Provenance asks where the record came from and how it reached its present form. Source quality asks whether the relevant source or process is competent, transparent, reliable and appropriate for the claim.
A perfectly traceable record can still be wrong. A potentially credible statement with unknown origin remains difficult to verify. Both controls are necessary, but they answer different questions.
What is the difference between content lineage and chain of custody?
Content lineage traces intellectual or computational derivation: which observation, document or dataset produced the current representation. Chain of custody traces possession, storage, access and transfer after acquisition.
A file can have well-documented custody while its original creator is unknown. Conversely, a document’s origin may be clear while preservation history is incomplete.
How should changing web pages be handled?
Capture the observed version and record retrieval time, publication time and modification time where available. Material changes should create new version records connected by revision or supersession relations.
Do not use the current live page as silent evidence of an earlier state. Historical claims require the historical representation actually observed or a trustworthy archive.
Does a screenshot prove provenance?
A screenshot preserves visible appearance but may omit the canonical URL, hidden data, interaction state, response headers and surrounding context. It is one documentary representation, not a complete provenance trail.
Pair it with capture metadata and, where relevant, source files, structured responses or archived pages. The required envelope depends on the claim.
How are copied articles treated?
Each article has its own publication identity, but copied content shares an upstream informational lineage. Provenance should connect republished, quoted or paraphrased records to the earliest recoverable source.
Copies can demonstrate distribution and persistence. They do not add independent corroboration of the underlying claim unless they contribute a genuinely separate observation.
How should translated evidence be cited?
Preserve and cite the original record, then identify the translation as a derived representation. Record translator or method, date, version and any review. For consequential claims, retain aligned original and translated passages.
If a conclusion relies on an uncertain translation, that uncertainty belongs in the evidence record and confidence statement.
What provenance is needed for AI-generated output?
Record the product or interface, model identity when exposed, prompt, relevant settings, retrieval context, time and exact output. Treat repeated runs as separate observations.
Then trace each factual assertion through its cited or retrieved sources. The provenance of the generated answer is different from the provenance of the facts it claims.
Why preserve raw data after normalization?
Normalization can rename fields, change units, merge categories, remove records or compute new values. Retaining raw input allows later reviewers to determine whether a finding came from the observed state or the transformation.
The normalized dataset remains useful, but its lineage should point to immutable inputs and a named transformation version.
What if the original source cannot be found?
Mark the lineage boundary as unresolved. Record the earliest recoverable copy, citation path, search attempts and missing relation. The evidence may still be used for discovery, distribution analysis or carefully qualified context.
Do not promote a repeated but untraceable statement into verified fact merely because it appears on many pages.
What is the minimum provenance record?
At minimum: evidence object identity, creator or producing system when known, publication or origin location, acquisition method and time, observed version, integrity reference, transformation history and exact claim use.
Unknown fields should be explicit. The objective is not administrative completeness; it is enough continuity for another reader to recover the record and understand every material step.
“Came from” is too vague. Name the exact relation.
Provenance becomes computationally useful when connections express what actually occurred. These relations distinguish creation, publication, capture, transformation, copying, revision and claim use without collapsing them into one generic source link.
Connect the record to its producing actor
“Created by” identifies the person, organization, instrument or system that produced the original evidence object. For a measurement, this may be the observation process rather than the organization hosting the export. For a written document, creator and publisher may be the same entity, but that equality should be established rather than assumed.
Use this relation when authorship or production is supported by the record, metadata or authoritative documentation. When the creator cannot be resolved, preserve “unknown creator” and retain the known publisher or host separately. Substituting the host for a missing creator manufactures certainty and can conceal copied material.
Connect availability to the responsible publisher
“Published by” records the entity that made a representation publicly or institutionally available. It answers who controlled that publication state, not necessarily who wrote the content or generated the underlying data. A journal may publish an article written by researchers; a news outlet may publish a licensed wire report; a repository may host a dataset created elsewhere.
This distinction matters when evaluating corrections, editorial responsibility and version history. The publishing entity can authenticate its own publication event while remaining unable to validate every upstream observation contained within the record.
Bind acquisition to an observed environment
“Captured from” links a retained record to the location, endpoint, device, instrument or environment from which it was acquired. The relation carries parameters and time because the same target can return different states under different conditions. A search page, authenticated dashboard or localized price has no complete acquisition meaning without those boundaries.
The collector should also be named. Two teams can capture the same location through different processes, producing records with distinct custody and possibly distinct observational limitations. Acquisition provenance is therefore more than a download timestamp.
Connect every transformed output to its inputs
“Derived from” states that an evidence object was produced by transforming one or more earlier objects. The relation should identify the transformation: extraction, translation, cleaning, normalization, classification, aggregation, statistical model or human synthesis. When parameters materially affect the output, they belong to the derivation record.
A derived object is not inherently weaker. Its value may come precisely from controlled transformation. The critical requirement is that the final representation does not masquerade as directly observed and that reviewers can locate the inputs and understand how the output was produced.
Preserve change without erasing earlier evidence
“Revision of” connects a new version to a prior version while retaining both identities. It applies to corrected reports, updated pages, amended datasets, recalculated metrics and edited generated summaries. The relation should describe what changed, why it changed and whether earlier conclusions are affected.
Revision differs from replacement. A replacement operation may overwrite the accessible file, while a provenance system still needs to represent the historical sequence. If the earlier bytes cannot be retained, record at least the known version identity, change event and limitation.
End the chain at a bounded proposition
“Supports claim” connects an evidence object to the exact proposition for which it is being used. The relation records whether support is direct, indirect or contextual, together with population, entity, variable, time and scope. This prevents a valid record from being stretched into a broader assertion than its lineage permits.
One object can support several claims differently. A status page directly supports what an operator reported, indirectly supports the full user impact and may only contextualize root cause. Recording separate relations makes synthesis and confidence calibration explainable instead of attaching one undifferentiated credibility label to the source.
Provenance resolves the history. Quality evaluates the source.
Continue from the record’s identity and lineage to source quality, preservation, corroboration, conflict, temporal fitness and its bounded role in a decision.
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 QualityAuthority, competence, transparency, incentives and reliability.
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.