Decentralised News · Agentic Finance · Decision research
By Heath Muchena · Source review 2026-11-11 · Instrument version 1.0.0
What Matters
A citation can point to the right page while the answer changes the conditions of its claim. Date, region, tier, eligibility and uncertainty must survive the journey into prose and machine-readable records. DN’s Condition-Retention Test Set records eight dimensions and separates omitted, changed and unresolved conditions. It measures a defined record, not an AI visibility score.
A buyer asks whether an offer is suitable. The answer names the right supplier and cites the right page, but leaves out the account restriction that makes the offer unavailable to that buyer. The citation is real; the recommendation is still unsupported for the requested use.
Research quality therefore needs a test beyond citation presence. The relevant unit is the claim with its conditions attached. The same test should apply when an answer is exported to a data record and read by another agent, because a correct page can become a misleading field without any broken link.
DN Evidence Block
Sources reviewed: 2026-11-11 · Evidence type: Primary citation-evaluation research and W3C data-publication guidance, followed by an original DN condition-retention protocol.
Scope: Material conditions attached to a specified claim. This is separate from whether the underlying claim is true, fresh or commercially useful.
- ALCE evaluates fluency, correctness and citation quality as distinct properties. Its framework supports examining what a citation establishes rather than counting links alone.
- W3C Data on the Web Best Practices addresses metadata, provenance, quality and version information. DN applies those documentation principles to claim conditions; W3C does not certify this instrument.
Not established by this work: DN has not measured condition-loss rates for named AI systems. The examples and tool statuses are synthetic.
Original DN instrument · version 1.0.0
DN Condition-Retention Test Set
Record condition fidelity for one claim at one test stage. Compare against a preserved source packet and state which answer or export is being assessed.
Enter your assumptions or load a labelled example. Required blank numbers stay missing; unknown evidence remains unknown. Use nonpersonal labels only. This instrument makes no network requests and uses no persistent browser storage.
Use your own inputs
No current result. Complete inputs or load an illustrative example.
Materiality and fidelity require adjudication outside the calculator. Conditions carry unequal decision value. A high proportion does not validate a recommendation, establish source truth or measure production performance.
1. A regional condition becomes a universal claim
A product or term available in one territory becomes a global statement. Preserve the applicable region and distinguish a supplier’s headquarters from the territory where the offer can be used.
Decision check: identify which location controls eligibility: incorporation, account jurisdiction, delivery territory or the user’s current location. A source about one boundary should not silently substitute for another. If the buyer’s relevant region is unknown, preserve that gap.
2. An effective date becomes current availability
The source was read today, but its offer expired yesterday or begins next month. Retrieval time, publication time and effective time answer different questions. Keep the date that governs the claim alongside the observation date.
Decision check: capture the answer’s as-of date and the offer’s effective interval. A current page can describe a future service. Re-fetching the same page cannot repair an answer that ignores the date already present.
3. A version-specific capability becomes universal
Documentation for one release is presented as a capability of another. Preserve the relevant product, API or document version when the distinction affects the decision. A page that updates in place may require a permitted snapshot or a version-specific reference.
Decision check: compare the operator’s actual version with the one named by the source. A feature documented for version three does not establish that a version-two integration can use it. Migration is a separate requirement.
4. A plan-tier entitlement becomes available to every account
A feature included in an enterprise agreement becomes a promise about every account. Record the required tier and whether the cited entitlement depends on a negotiated order form or another condition.
Decision check: record the account tier used by the buyer and any negotiated entitlement. In a fictional fixture, an enterprise-only export feature must not become an included feature for every free account. An unknown tier is not evidence of qualification.
5. Eligibility or permitted use expands beyond the source
A new-customer promotion becomes available to an existing customer, or an internal-use entitlement becomes a resale permission. The condition must remain attached to the recommendation, with unresolved eligibility identified explicitly.
Decision check: distinguish a published rule from evidence that this buyer satisfies it. Account creation, verification, approval and permission to redistribute can be different states. A successful sign-up screen does not establish the entire chain.
6. The number survives but its unit disappears
A per-seat price becomes an organisation price, a monthly rate becomes a total commitment or a capacity figure loses its measurement basis. Preserve quantity, period and the assumptions needed to compare like with like.
Synthetic arithmetic fixture: US$15 per active seat per month with a five-seat minimum means a US$75 minimum monthly base charge. If tax is unspecified, US$75 is not a complete payable quote. Preserving the number fifteen while dropping its unit changes the buying decision.
7. Uncertainty becomes unwarranted certainty
A provisional estimate becomes a verified result, or the absence of evidence becomes a negative finding. Retain the source’s confidence and the researcher’s evidence status. A concise answer can still say that a purchase-changing condition is unconfirmed.
Decision check: a simulated 8–12% saving under stated assumptions must not become a claim that customers achieved 12%. This is an invented example of a meaning change, not an observed model response. Preserve evidence class, uncertainty and material assumptions.
8. The material exclusion disappears from the answer
A broad positive statement survives while the exception that matters to the buyer disappears. Record exclusions near the claim they qualify. A footnote technically present elsewhere may not be available to the downstream agent consuming one field.
Decision check: test the exception that matters to the task. A hypothetical internal-analysis permission with a raw-data redistribution exclusion must not become permission to resell raw records. A short paraphrase can pass if it retains the relevant boundary.
Methodology, limitations and corrections
DN combined citation-quality research with data-documentation principles to propose a claim-level retention test. No current model or vendor was scored. The eight dimensions are a reusable starting set that should be adapted to the actual buying decision. Preserve source packets, exact questions, retrieved material, answers and export versions; adjudicate ambiguous cases separately. Report failure stages and materiality rather than relying on a single aggregate. The synthetic example exists only to demonstrate the calculator’s denominator.
Define the claim that the citation is supposed to support
Freeze a source record and write the exact proposition the answer needs to establish. Include the buyer’s question, the relevant passage and every condition that could change the decision. “The plan costs 20” is incomplete if the source means 20 per seat per month, billed annually, for eligible new customers in one region.
Separate source support from answer fidelity. A faithful answer can repeat an unsupported supplier claim; an unfaithful answer can distort a well-supported source. The protocol here checks preservation of meaning. Verification of the underlying evidence remains a separate task.
Use a claim packet with a stable ID, source URL, captured version or observation time, relevant passage, structured conditions and a rule for unresolved evidence. A reviewer should be able to determine whether the answer’s scope matches the packet without reconstructing the original research from scratch.
Keep competing or contradictory sources visible. If the pricing page and current contract disagree, the appropriate answer may be that the offer is unresolved. Removing the contradiction to produce a cleaner machine-readable field would change the evidence state.
The packet should contain only what the publisher is entitled to preserve and share. A source citation and a full licensed document are different artifacts with different access conditions.
Four conditions that limit where and when a claim applies
For each dimension, decide whether it is material to this claim. Do not attach arbitrary precision. A timeless definition may not need a commercial validity window. A time-limited purchase offer does. The reason for marking a condition inapplicable should be reviewable.
Four conditions that change what the buyer can do
These examples describe potential transformations, not measured prevalence. In a real corpus, the materiality rule should come from the decision. An omitted condition is consequential when it could change selection, eligibility, total commitment or the action the agent is permitted to take.
Test retrieval, answering and machine export separately
Start with a small fixed set of claim packets whose conditions can be independently checked. Include straightforward cases and cases with a material exception. Keep the questions and source versions stable during the comparison.
At retrieval, inspect whether the relevant passage and conditions reached the answering system. If a restriction never entered the evidence packet, the failure differs from a restriction that was retrieved and then omitted in the answer. Preserve that distinction when assigning a remedy.
At answer generation, compare the output with the packet. Record which conditions are retained, changed, omitted or unresolved. A citation can support one sentence while leaving a broader recommendation unsupported; evaluate the actual proposition made.
At export, inspect the machine-readable representation and then read it through the intended downstream route. Conditions must remain linked to the claim, rather than being stranded in unrelated prose or discarded during a schema conversion. Test the final interpreted answer as well as the intermediate file.
| Stage | Evidence to preserve | Typical question |
|---|---|---|
| Source packet | Claim and material conditions | What was supposed to survive? |
| Retrieval | Material actually supplied | Did the restriction reach the system? |
| Answer | Claim, citation and conditions | Did the answer preserve scope? |
| Export and re-ingestion | Record and downstream interpretation | Did another consumer receive the same meaning? |
Use a separate held-out set when changing extraction or answer instructions. Repeatedly fixing the same examples can improve a demonstration without establishing general retention.
Use the DN Condition-Retention Test Set
The instrument records one status for each of the eight dimensions. Mark “not applicable” only when the condition is irrelevant to the specified claim, not when the source is silent about a condition the buyer needs. Silence in a material field remains unresolved.
The retained proportion uses applicable conditions as its denominator. Changed, omitted and unresolved conditions remain visible in the result. If every dimension is inapplicable, the tool reports no rate. It does not award perfect fidelity to an empty assessment.
In an illustrative eight-condition record, five conditions are retained, one is changed, one omitted and one unresolved. Reported retention is 62.5%. That percentage is not a score of the answer’s overall truth or usefulness. The lost condition might be the only one that determines whether the buyer can proceed.
For a maintained research service, retain the condition-level record and the failure stage. That produces an actionable correction: improve retrieval, revise answer construction or repair the export mapping. Counting citations alone would not identify which intervention is needed.
Publish the question set, review date and boundaries when reporting results. A panel tested on selected offers in one region cannot support a claim about every vendor, language or model. The useful commercial asset is a reproducible record of what each answer preserved and what remains unresolved.
For corrections, identify the claim, source and date when contacting Decentralised News. A new record-creation timestamp does not refresh its underlying evidence.
Change log: version1.0.0, 2026-11-11. Initial documentation review, original instrument and labelled examples.
Frequently asked questions
Is a correct citation enough to support an answer?
Only if the cited evidence supports the actual claim with its material conditions. A source can be relevant while the answer overstates its scope.
Does faithful citation prove that the source is true?
No. Fidelity concerns preservation of the source’s meaning. Source quality, factual verification and freshness need their own checks.
How should an absent but necessary condition be recorded?
As unresolved. “Not applicable” is reserved for conditions irrelevant to the claim, not for evidence the researcher could not find.
Primary sources
- Enabling Large Language Models to Generate Text with Citations. Separating answer fluency, correctness and citation quality. Reviewed 2026-11-11.
- W3C Data on the Web Best Practices. Metadata, provenance, quality and version documentation. Reviewed 2026-11-11.
Turn the finding into a decision
Research Desk: Commission a citation-fidelity review for one decision, a defined question set and a permitted source corpus. The deliverable is a condition-level record across retrieval, answer and export, with unresolved claims and specific repair priorities. DN sells research services; the assessment does not promise greater AI visibility, recommendations or revenue.
Related DN research
Disclosure: Analysis and instrument results are editorial. Where labelled affiliate links appear, DN may receive compensation when you sign up or buy through them. Commercial relationships do not determine findings. No tested integration or guaranteed outcome is implied.
