Skip to main content

Methodology

The analyzed corpus includes 532 citations of scientific studies, collected over the period from March 12, 2025 to June 13, 2026. These citations come from content published under the name “Thérapie Autisme”. It includes posts from the Instagram account “therapietsa” up to June 9, 2026, articles from the “Thérapie Autisme” blog up to June 13, 2026, bibliographic documents associated with tools marketed by “Thérapie Autisme”, and the manuscript titled “Variabilité neuro-architecturale dans l’autisme adulte : vers une cartographie morpho-fonctionnelle”, published on the Skool.com platform “Thérapie TSA - Autisme Adultes”.

Process overview

  • 1. Citation. A cited reference must contain usable author information, a year and a title. Journal, volume, pages and DOI are useful but optional.
  • 2. Extraction. Visible bibliographic elements are separated without reformatting the citation into a standard academic style.
  • 3. M1 search. Crossref, OpenAlex and PubMed are searched from author, year and title.
  • 4. M2 search. Google Scholar is queried through SerpApi when this search is enabled.
  • 5. Final score. The final candidate is selected, each bibliographic element is compared, an overall score is calculated, and the title Jaccard index is recorded.
  • 6. Records. Each record presents the citation, the identified reference, verification links and calculation details.

Corpus construction

Each citation must contain at least a first author, a year and a title present in the published content. Cited elements are preserved even when they differ from the found reference. The calculation measures those differences; it does not complete the cited title using the candidate.

  • Author(s). Must be present in a usable form: single author, multiple authors or “et al.”.
  • Year. Must be present to situate the reference and limit homonymy.
  • Title. Must reproduce the bibliographic title visible in the published content. Generic autism-only titles use the exception described below.
  • Journal, DOI, volume, pages. Useful for assessing citation fidelity, but not required for inclusion in the corpus.

Candidate reference search

  • M1: Crossref, OpenAlex, PubMed. Searches are based on author, year and title. The cited DOI is not used to drive the search, because an incorrect DOI could lead directly to the wrong reference. This method may fail to retrieve some titles or may return incomplete metadata.
  • M2: Google Scholar via SerpApi. Used as a supplementary source when the citation is incomplete, imprecise or phrased differently. Its metadata are less structured, and information such as journal or volume may be missing.
  • Selected reference: M1 or M2, or M3 after review of a community proposal. This is the reference used in the charts and citation records. Ambiguous cases still require human review.

Engine version : 2026-11-17.10-j20

Selection of the reference and score calculation

M1 and M2 are search channels. Their candidate references are compared using one bibliographic score. The same rules apply to the corpus and the citation tester.

Obtaining the M1 and M2 candidates

M1 combines results from Crossref, OpenAlex and PubMed. Crossref uses the title, year and first author; OpenAlex searches the title; PubMed uses title, author and year, with a broader text search if necessary. The cited DOI is not used as a search key when the required title is present.

Within M1, duplicate records are removed. Candidates are ranked by a matching cited DOI, then title Jaccard, similarity of core title words, overall score, author score, year score and presence of a DOI, in that order. This internal ranking is distinct from the final choice between M1 and M2.

M2 uses the first Google Scholar result when that result is usable and the search is enabled. Later results are not automatically promoted in its place. An unavailable search or unusable first result may therefore leave M2 absent.

Existing corpus results are recalculated from recorded candidates without new searches. When a tester input exactly identifies a single public corpus citation, its recorded candidates and accepted M3 are reused. Otherwise, new searches may return different candidates as databases change.

Bibliographic comparison rules

  • The comparison uses the available M1 and M2 candidates and their recorded metadata.
  • The final score uses five equal weights of 20%: title, combined authors, year, journal and DOI. Elements that cannot be assessed are omitted from the denominator.
  • Authors are compared in the same position. Full family names must match after normalizing accents and separators; a shared fragment is insufficient. Given-name initials must be compatible when available on both sides.
  • If M1 and M2 share the same nonempty DOI, journal, volume and pages are taken from M1 before recalculating the final score.
  • If the candidate score and title Jaccard are both equal, M1 is retained by convention. This is not evidence that the references are identical or validated.

A surname alone does not establish a person's identity; different people can share it.

How each bibliographic element is scored

  • Authors — The scores of the authors provided in the citation are averaged, in their stated order, up to the first three. For example, two matching authors out of three give 66.7/100. An author missing from the candidate does not match a cited author.
  • When given-name initials are available in both the citation and the candidate, incompatible initials give 0/100 for that author, even if the family name matches. Compatible initials and a matching full family name give 100/100. Missing initials on either side are not penalized; an additional middle initial on only one side is not a contradiction.
  • Initials explicitly present in the author part of the raw citation are recovered when its structured author fields retain only the family name. Existing structured initials are preserved. No author is added, reordered or inferred from the candidate or the title. This rule applies to the corpus and the citation tester, before averaging the author scores and selecting M1 or M2.
  • Year — The same year gives 100/100, a difference of one year gives 70/100, and a larger difference gives 0/100.
  • Journal — The journal name is decisive, including recognized abbreviations and aliases. If the name does not match, the journal component is 0/100 even when the volume and issue match. If the name matches, the component is 100/100 when comparable volume/issue values agree, and 70/100 when the volume, issue, or both differ. These differences produce a single partial score; penalties are not added together.
  • A volume or issue absent on either side is not treated as a contradiction. An ambiguous format is not penalized. A journal score of 100/100 concerns the comparable information; it does not certify that the record is complete. If the journal name itself is missing from the citation or the candidate, this component is omitted from the denominator.
  • Standard journal word abbreviations are expanded before comparison: for example, J Am Acad Child Adolesc Psychiatry matches Journal of the American Academy of Child and Adolescent Psychiatry. All significant words must correspond; a truncated name or a different journal is not made equivalent by this rule.
  • DOI — When present on both sides, identical normalized DOIs give 100/100 and different DOIs give 0/100.

Each calculation displays the cited and found values together, followed by their status, score and weight. Volume and issue belong to the journal component; they are not additional weighted elements. The title uses the Jaccard index and the subtitle tolerance described below.

The title component is the Jaccard percentage rounded to one decimal place. The global score is the weighted sum of the assessable component scores divided by their total weight, rounded to one decimal place. A cited author absent from the candidate receives zero; a missing candidate year, journal or DOI is not assessable and is omitted.

Character similarity alone never makes different journal names equivalent. For example, Neuropsychologia and Neuropsychology receive a journal concordance of 0/100. Recognized abbreviations and explicitly verified title variants, such as historical titles of the Royal Society biological B series, remain accepted.

Title typography and significant words

Exception for generic autism-only titles: when both titles consist solely of Autism, Autistic, ASD, ASC or an equivalent autism spectrum disorder/condition expression, they are normalized to the single token autism. Their intersection and union both contain one token, so title Jaccard is 100%. A title with additional meaningful words does not qualify for this exception. For other comparisons, the usual exclusion of generic domain words remains in force.

A Jaccard of 100% for a generic title establishes lexical equivalence under this rule, not publication identity. Different articles can have the same title; authors, year, journal and DOI are still compared separately.

A hyphenated compound already shared by both titles is kept as one word. If a compound differs and one of its significant parts occurs as a separate word in the other title, it is split for the comparison. All its parts are then subject to the usual word exclusions; unmatched words and negation prefixes are not silently discarded. Word hyphen variants are normalized, while long dashes remain punctuation separators.

For example, gender-informed can match gender informed, and children-what can contribute the word children while keeping what as an additional word. The calculation shows which compounds were split. The displayed titles are preserved.

This tolerance does not treat every hyphen as a subtitle boundary. The existing subtitle tolerance applies to an entire matching colon-separated segment; a missing question or other words still affect Jaccard outside that rule. A change in title Jaccard can also change candidate selection or whether the final rejection threshold is crossed.

Identification and differences in the citation

A score measures concordance; the result label separately describes identification and differences in the evaluated elements. The same interpretation is used for the corpus and the citation tester.

  • DOI resolved — cited title contradicts the reference — The DOI leads to a publication, but its title does not confirm the title supplied in the citation. This is not a validated identification of the cited title.
  • Publication identified — differences found — A publication is identified, but at least one evaluated element differs or only partially matches. The fields concerned are listed, including coauthors, year, title, journal and DOI where applicable.
  • Compared elements concordant — The evaluated elements agree under the stated normalization rules. This label does not certify missing or unevaluated fields, scientific quality or the validity of the conclusions.
  • The labels specify the measured differences: matching title and authors with a different year or journal, differences in volume/issue, or low/partial title correspondence with its percentage.
  • The cited and found year, journal, volume/issue and DOI are shown when they differ. Information unavailable in the candidate record is distinguished from an element not provided in the citation.
  • Full title correspondence can result from normalization, subtitle tolerance or a generic title. It does not by itself confirm publication identity. More precise labels do not change scores, candidate selection, distributions or archived calculation verdicts. Excluded candidates and searches with no reference remain distinct.

Choosing the reference

  • The highest candidate score wins. Title Jaccard breaks a score tie. DOI contributes to the score but has no absolute priority.
  • If both the candidate score and title Jaccard are equal, M1 is retained by convention.

Selection compares each candidate with its own metadata, before the possible addition of M1 metadata. The final score is then recalculated. The score used for selection may therefore differ from the final displayed score.

Verified manual additions to candidate metadata

A candidate can also be replaced manually after verification in a publisher or primary bibliographic record. The changed fields are identified as manual corrections, with their evidence and date; the previous candidate is retained in provenance. This does not change the cited reference or impose the final choice between M1 and M2.

Some candidate records returned by the search engines have missing or truncated metadata. When the publication is identified without ambiguity, the missing information is completed manually from verified publisher or bibliographic records. This applies only to elements present in the cited reference. The cited reference itself is preserved.

The Status column identifies each manually added or corrected element, its original value when applicable, the verification date and a link to the evidence. The same verified metadata are used by both the corpus and the citation tester, before comparing M1 and M2.

An existing candidate DOI can also be corrected when a publisher-page or file suffix was included in the identifier. These corrections use exact DOI pairs and publication titles verified in primary records; arbitrary suffixes are never removed. Unlike adding missing metadata, this cleanup also applies when the citation does not provide a DOI. The cited DOI is preserved, and an uncited DOI does not acquire a score or weight. The original candidate identifier and verification source remain visible in Status.

Complete saved citation tests are recalculated offline from their recorded candidates when the calculation rules change. Incomplete old inputs remain unevaluated. Published change-history snapshots and previous moderation decisions preserve their original scores and calculation versions.

  • Volume and issue are checked separately from article numbers. Explicit labels, leading zeros, combined issues and typographical variants are normalized; unlabelled volume and issue numbers are never swapped automatically. An ambiguous format is not used to impose a mismatch.
  • When the publisher gives a volume and an article number without a separate issue, only the verified volume is added. Real differences in the cited volume or issue remain visible and affect the journal comparison.
  • Pages are not included in these manual additions: they do not enter the score calculation. A missing author or journal is not invented when the found publication does not have that element.

Inspect citations whose M1/M2 candidates contain verified manual metadata

Final rejection rule for a very weak title

After selection and metadata consolidation, a title Jaccard strictly below 20% combined with zero or one other concordant element reduces the final score to zero. This indicates a high risk of an incorrect match; it does not establish that the cited publication does not exist.

  • The combined authors, journal or DOI must score 100/100 to count as concordant. An individually matching author is insufficient if the combined author score is partial.
  • The year counts at 100/100 or with the existing one-year tolerance (70/100). Other partial matches do not count as concordant elements for this rule.
  • The rule applies with zero or one concordant element outside the title. Removing an author match therefore cannot restore a small score merely because no concordant element remains. The rule does not apply with two or more such elements, or when title Jaccard cannot be calculated. Equality with the threshold does not trigger rejection.
  • The calculation shows the provisional score crossed out, followed by zero and the reason. Candidate selection and measured component scores are preserved; the distributions use the final score after this rule.

Existing title normalization and subtitle tolerance apply before this rule: when the cited title matches an entire colon-separated segment of the candidate title, the omitted segment is tolerated and title Jaccard is 100%. This does not cover arbitrary omissions or the reverse situation.

Outside this rule, a weak title or an author mismatch alone does not force the score to zero. The measured scores remain visible, with warnings and a separate validation status. The score does not establish scientific validity.

The small colored bars show title Jaccard within each score group. Their counts and percentages refer to that group, and their links combine its filters. Gray represents Jaccard values that cannot be calculated.

Community contributions and reference versions

Adding a citation means reporting a citation that may have been overlooked when compiling the corpus and is present in the Thérapie Autisme content studied, including blog posts, Instagram posts, tool bibliographies and study pages.

View the content studied for this corpus

Anyone can propose a better reference by providing its DOI, or report a citation overlooked when compiling the corpus from the studied content. The original citation must contain a first author, a year and a title, and the Thérapie Autisme content must have been published between 12 March 2025 and 13 June 2026, inclusive. The publication date of the content is distinct from the year of the scientific article.

Testing a proposal and submitting it do not change the corpus. Only the administrator can accept or refuse it, with a public reason in French and English. For a replacement, the proposed final score must exceed the currently selected overall-concordance score at review time. A higher score alone is insufficient: the identity and metadata of the publication must also be checked.

An accepted community reference becomes M3 and is selected after administrative review. The original M1 and M2 candidates and their automatic selection remain available for comparison. M3 is a reference origin, not an additional scoring formula. The reviewed reference uses the existing bibliographic calculation rules.

The proposed DOI identifies the candidate; it is never added to the cited fields. An uncited DOI therefore does not gain weight, and an incorrect cited DOI remains a disagreement. Verified corrections of metadata for an already identified DOI can be accepted separately even when completing the record decreases its score.

Every decision preserves its previous reference, scores and calculation version. Pending or refused proposals do not affect statistics. Accepted additions complete the citations extracted from the studied content, while preserving the corpus's content scope and publication window. Comments remain private; public forms require an arithmetic check and accept short plain-text comments only. No contributor email address is collected.

Improve the corpus · Proposal history · How to contribute

Change history — Previous references, scores and reasons are preserved when a change is published. Tests and pending proposals do not change the corpus.

General score

The bibliographic concordance score is the weighted mean of assessable components. A year, journal or DOI missing from the candidate is omitted from the denominator, as is an uncited element. A cited author missing from the candidate, however, receives zero in the author average. A non-comparable title is not treated as a zero match. The mean is rounded to one decimal place.

  • Title: 20%. Compared using significant words and controlled tolerances.
  • Authors: 20%. Positional comparison of cited authors in the order provided.
  • Year: 20%. Exact match or near match, depending on the case.
  • Journal: 20%. Journal name, including volume or issue when cited.
  • DOI: 20%. Exact match if the DOI is cited.

Missing DOI example. If title, authors, year and journal are assessable but DOI is absent, the denominator is reduced to 80%. A citation scoring 80 for title and 100 for the other three available elements therefore receives: (80×0.20 + 100×0.20 + 100×0.20 + 100×0.20) / 0.80 = 95/100.

Title fidelity: Jaccard index

The Jaccard index compares the significant words in the cited title with those in the identified title. It is calculated as the number of shared words divided by the total number of significant words appearing in either title.

  • Normalization. Case, accents, punctuation, compound words and singular/plural variants are normalized.
  • Acronyms. ASD or ASC may correspond to Autism Spectrum Disorder(s) or Condition(s).
  • Language variants. Variants such as behavior and behaviour are treated as equivalent.
  • Overly generic words. Terms such as autism, spectrum, disorder(s), autistic, ASD, ASC and condition(s) are excluded, except when both titles consist solely of autism-domain expressions, in which case the exception above applies.
  • Incomplete titles. A title segment appearing before or after a colon may be tolerated only if all cited words correspond to the remaining segment.

Special rules and edge cases

  • Cited but incorrect DOI. The DOI does not drive the search. It is compared afterwards and may penalize the DOI element.
  • Abbreviated journal title. Standard acronyms or abbreviations may be accepted, for example JADD for Journal of Autism and Developmental Disorders.
  • Volume only. A citation such as 41 may be considered compatible with 41(9) if the issue number was not cited.
  • Different volume. For 41(2) versus 40(3), both numbers differ: the journal component is 70/100 if the journal name matches, and zero otherwise.
  • Missing element. An uncited element is omitted. A year, journal or DOI missing from the candidate is also not assessable; a cited author missing from the candidate instead receives zero in the author average.
  • Approximate citation. The procedure allows certain variations to determine whether a real and coherent reference exists, while still reporting identified divergences.

Verification and corrections

  • Verification links. Records include the citation source, the selected bibliographic source and a Google Scholar search generated from the citation.
  • Reading records. Detailed comparisons show authors, year, title, journal, DOI and Jaccard index to support manual verification.
  • Corrections. Possible errors or useful corrections can be reported through the AutiHub contact page.