Three citations or three independent sources?

Count evidence origins before counting links

You are preparing a market note. Three articles describe electric cars taking roughly 45% of Singapore's new-car registrations in 2025. They come from different publications and have different bylines. Does that give you three independent checks of the number?

The answer depends on what each article contributes to the particular claim. A publication can repeat an official count, calculate a percentage and add an expert's explanation in the same story. Those are different kinds of evidence, with different dependencies.

Our source-tracing guide follows a market estimate back to its published origin. This guide takes the next step: deciding how much corroboration the resulting sources actually supply. Finding three pages is a starting point, rather than a count of three measurements.

For this case, the explicitly selected sample contains four downstream publications: CNA, HardwareZone, F&L Asia and The Business Times. LTA is a separate primary-origin record. HardwareZone and The Business Times demonstrably share the LTA registration-statistics observation family; CNA's origin is inferred, and F&L's exact inputs remain unresolved. We do not reduce all four to one by assumption.

The preferred leads were CNA, HardwareZone and F&L. Because F&L could be checked only through an indexed excerpt, we added The Business Times to obtain a third fully inspectable claim-bearing article and an explicit LTA-attributed numerical pair. Motorist is a supplementary check of the 45.1% discrepancy. These additions answer named access and numerical questions.

These are different quantities, not a confidence score or a confirmed total family count for every account.

Count evidence origins before counting links.

From this guide

Fix the claim you want to corroborate

A useful claim record is narrow enough to check:

The source is LTA's Annual Vehicle Statistics 2025 table, Total Annual New Registration of Cars by Make. Its first four PDF pages provide fuel-type rows and an annual total. We summed the 2025 Electric rows across both importer types and checked that all fuel categories together reconcile to the printed total.

This claim is about the year's inflow of registered cars. It is not the percentage of all cars already on the road, a measure of every type of electric vehicle, or a monetary EV-market size. Independence becomes a useful question only after the measure is fixed.

Inspect what the selected sources actually say

The table separates four selected downstream accounts from their primary origin:

The table distinguishes membership in the LTA registration-statistics family from use of a particular PDF edition. HardwareZone and The Business Times explicitly identify LTA and print the same 2025 numerator and denominator as the inspected table. That supports grouping two claim appearances inside the selected downstream sample into family F1. The exact historical files remain unestablished.

F&L names LTA, but its 45.1% inputs and exact observation membership remain unresolved. Its access-qualified excerpt is not treated as a fully inspected body.

CNA's matching counts suggest a common lineage without confirming its source. Different bylines do not prove different data collection; identical numbers do not prove copying.

Draw the dependencies at claim level

For the registration-share claim, the supported map is:

1. 01

   LTA registration statistics → HardwareZone and The Business Times: confirmed producer attribution and matching counts

2. 02

   The Straits Times → HardwareZone: confirmed republication credit and exact article link

3. 03

   LTA registration statistics → CNA: inferred common lineage, based on matching counts

4. 04

   LTA attribution → F&L Asia: confirmed in the indexed excerpt; exact statistic inputs unresolved

5. 05

   The inspected LTA annual PDF → New Eruditions calculation: confirmed use of this document

The Straits Times is outside the selected downstream sample. Its credit establishes HardwareZone's reporting dependency; the within-sample family grouping rests on the explicit LTA attribution and matching counts in HardwareZone and The Business Times.

No supported copying arrow connects the selected outlets. January's reporters are not assumed to have used the PDF edition downloaded in October.

CNA also attributes market explanations to Walter Theseira and Benjamin Loo. These support descriptions of their reported views, not another nationwide registration measurement. We did not conduct the interviews or establish exclusive collection. CNA's report

HardwareZone credits the broader story to The Straits Times. Its comments are not newly collected HardwareZone testimony merely because another site publishes them.

Resolve what the arithmetic can resolve

Derived calculation:

Conventional rounding gives 45.0% to one decimal place, or 45% to a whole percentage point.

F&L's 45.1% in the indexed excerpt (also reported by supplementary source Motorist) cannot be obtained by rounding those same counts to one decimal place. That is a reproducible numerical finding. The difference between the two displayed one-decimal figures is 0.1 percentage point.

The fully inspected Motorist account also names LTA and prints 45.1%. It confirms another reported appearance without resolving the inputs.

Different inputs, scope, source editions, calculation choices or an error could explain the difference; none is established. Calling it a typo or a harmless rounding difference would exceed the evidence.

A sensitivity check shows what would have to change. Holding the numerator at 23,684, a whole-number denominator from 52,457 to 52,572 would produce 45.1% under conventional nearest rounding. Holding the denominator at 52,678 would instead require 23,732 to 23,784 Electric registrations. These are calculated input ranges, not alternative observed totals.

For the defined claim and inspected annual table, use approximately 45.0%, accompanied by the counts and source vintage. Retain 45.1% in the discrepancy record as an attributed report whose inputs remain unestablished. Do not average the two figures into a compromise percentage.

Separate data independence from analytical independence

For practical research, ask which layer a source adds:

Recalculation checks extraction, denominator choice and rounding. It cannot detect every error in the underlying register. A second analytical check still shares the data dependency. Overlapping inputs can also make separately collected evidence partly dependent; independence guarantees neither accuracy nor comparability.

In evidence synthesis, the Cochrane Handbook distinguishes studies from their multiple reports and recommends linking those reports. The corresponding discipline here is to retain additional information without counting shared observations twice.

Decide what this case supports

For the defined national registration share, this check supplies one inspected statistical origin. Within the four-publication sample, HardwareZone and The Business Times share family F1, also represented by the separate LTA origin record. CNA's exact numerical origin remains unresolved, with a shared LTA lineage inferred. F&L names LTA in its indexed excerpt, but its differing share has unestablished inputs. It should not be credited as a confirmed independent measurement merely because it appears on another domain.

This bounded selection establishes no second independent nationwide measurement. It neither confirms all exact source routes nor groups every sentence into one family.

A defensible note could say:

LTA's 2025 annual car-registration table supports 23,684 Electric registrations out of 52,678 cars, approximately 45.0%. These counts include COE and tax-exempt cars and exclude taxis. HardwareZone and The Business Times explicitly attribute matching counts to LTA; HardwareZone also credits a Straits Times republication. CNA reports matching counts without visible producer attribution; an indexed F&L excerpt names LTA but says 45.1%. These accounts do not disclose additional independent national measurements. A later reported 45.1% share cannot be reproduced from the inspected counts; its derivation remains unresolved. Public sources checked 1 October 2026.

Cite LTA for the published rows and this guide for the dependency assessment, if useful. Cite the news accounts for their reporting or attributed comments. Let each citation do the job its evidence can support.

Reuse the independence check

1. 01

   Freeze the claim. Record measure, unit, period, geography, category, population and status.

2. 02

   Separate the passages. A count, calculation, forecast and interview statement may have different origins inside one article.

3. 03

   Find the disclosed inputs. Follow exact links, credits, dataset identifiers and methods. Record the document and its version separately from the producer.

4. 04

   Test the relationships. Mark confirmed attribution, inferred lineage and unresolved routes. Similarity is a lead to investigate.

5. 05

   Group shared inputs. Put reports of the same evidence in a provisional family for that claim. Keep uncertain assignments visible.

6. 06

   Check the transformations. Reproduce calculations and compare denominators, categories and rounding. A disagreement can reveal a transmission or scope problem without proving its cause.

7. 07

   Look for genuinely distinct evidence. Ask what was separately collected, by whom, how, and with what overlap. A fresh interpretation of the same data has value, but solves a different validation problem.

8. 08

   Write the conclusion at the supported level. State what each source adds, where dependence remains, and what evidence would change the assessment.

Before calling a claim corroborated, ask:

1. 01

   Am I counting publications, evidence-producing processes or actual observations?

2. 02

   Can I identify the input behind each citation?

3. 03

   Is the claimed relationship explicit or inferred?

4. 04

   Have I mistaken a fresh calculation for fresh data?

5. 05

   Do the sources measure the same thing?

6. 06

   Have I preserved unresolved discrepancies and missing source routes?

7. 07

   What would remain untested even if all these accounts agreed?

Use the accompanying claim-support and relationship templates. The worked market-size comparison addresses comparability. Independence alone does not establish that two measures can be compared. See the research guides.

Method, limits and corrections

This guide used public source discovery, inspected reporting and a downloaded LTA annual PDF. We checked PDF layout and scope notes, extracted its first four pages, reproduced fuel totals and calculated the share. The research files preserve the source URL, access date and file hash. The PDF is labelled Annual Vehicle Statistics 2025; its exact publisher publication and revision dates were not established. A current file is not proof of the exact edition available to January's reporters.

This is a bounded source-dependency check. We did not audit raw registration records, measure the EV market independently, conduct interviews or use proprietary data. F&L evidence in the extended ledger is access-qualified. Additional sources or a dated correction could change an edge or resolve the 45.1% discrepancy.

For a factual correction or broken source link, use our contact page with the article address and the issue. A material correction should identify the affected claim, preserve the earlier record and carry its actual update date. Routine formatting or rebuilds should not refresh publication or research-check dates.

Reuse the method

1. 01

   Reader checklist

2. 02

   Blank claim-support ledger

3. 03

   Blank family summary

4. 04

   Source dossier

5. 05

   Populated claim-support ledger

6. 06

   Dependency graph and text equivalent

7. 07

   Official-row calculation inputs

8. 08

   Calculation script and instructions