Comparability in the public olfactory record: volatile measurement, human perception, and the join between them

1 Introduction

• 84.2% of records contain collision energy in 1,023 distinct ways, while an exact join retains 9.3% of the comparisons retained by a normalised join.Completeness reports 84.2%; the usable comparison figure is 9.3%.Source: §1 · p. 1
• Comparability asks whether two records can be used together, and for what; the paper treats it as relational, graded, and computable.Source: §1 · p. 1

Scope

• The study draws on 180,877 records across eight public repositories, spectral archives, and standards bodies, and classifies 167,331 analysis pairs.It also recovers 206 thermal programmes and 66 reference series from prose and releases 48,356 triples.Source: Scope · p. 2

2 Background

• The authors adopt psychometric measurement invariance because it treats comparability as graded and separately testable, while its statistical machinery does not transfer.Other disciplines independently use related approaches, including measuring known anchors alongside unknowns to place measurements on a common scale.Source: §2 · pp. 2–3

2.1 Measuring the separation

• 12 of 15 terminology pairs share no work; the 28 observed co-occurrences occur in only three pairs, and none cross disciplines.The count measures terminology co-occurrence rather than citation, so it is a lower bound on contact and depends on the chosen terms.Source: §2.1 · p. 3

• The psychometric logic of separately tested levels transfers to gas chromatography, but the factor-analytic apparatus does not.Source: §3 · p. 4
• The four cumulative rungs license identifying the same thing, comparing orderings, comparing calibrated retention indices, and comparing values directly.The records must share a measurand and technique, stationary-phase family, stated reference series, and run conditions for successively higher rungs.Source: §3 · p. 4
• Rung assignment uses only recorded values, takes the highest satisfied predicate, and is reproducible from the same corpus without sampling error.The authors note that uncertainty is definitional, particularly in the stationary-phase and thermal-programme predicates.Source: §3 · p. 4

4.1 A taxonomy of crosswalk operations

• Five crosswalk operations differ by orders of magnitude in reliability, so treating them as one kind of mapping risks confident errors.Reliability and the strength of claims supported by derived values fall down the stack.Source: §4.1 · p. 4, Figure 2 · p. 5
• Registry lookup resolves 5,837 of 5,872 molecules; the remainder are absent from the bridge, not known to be unmeasured.Term binding assigns 2,106 spectra to a modality, while 134,924 are excluded and counted rather than bound.Source: §4.1 · p. 5
• Normalising collision-energy values increases retained comparisons from 197 to 1,724, an 8.7-fold recovery.Source: §4.1 · p. 5
• Anchor inference retains 29 of 3,358 descriptor pairs at 0.544, but supports only a statistical claim that a later method may overturn.In this case it establishes the dominant axis of a perceptual space, not the vocabulary mapping the analysis sought.Source: §4.1 · p. 5

4.2 Provenance and grading of derived values

• Each derived value carries a corroboration level, pipeline attribution, date, and, when extracted from text, its source sentence for reader adjudication.The authors argue that evidence-linked extraction helps readers assess the many small claims made at scale.Source: §4.2 · p. 6

4.3 The error model

• Temperature extraction can capture decoy temperatures or fabricate a hold from an adjacent run-time statement; mitigations reject decoy-governed candidates and exclude run-time durations.Source: §4.3 · p. 6
• Substring matching can confuse chemically distinct phases, while empty joins can make aggregate checks pass vacuously; the authors bound patterns and require checks to fail closed on empty input.Source: §4.3 · p. 6

4.4 Auditing an extraction

• Audits pair extracted values with their source sentences and require a fresh audit after extractor changes; precision intervals remain wide, but the ordering was stable enough to guide release decisions.The thermal extractor improved from 7/12 to 11/12 after tightening and finished at 10/14; the internal-standard extractor returned 4/12 and was withdrawn.Source: §4.4 · p. 6, §4.5 · p. 7

5 The chemical record

• Two of 167,331 analysis pairs reach the rung that licenses comparing values directly, and this is an upper bound.The equality test tolerates one programme stating a run time that the other omits.Source: §5 · pp. 7–8
• The fall from the second rung to the third is approximately a factor of 800 because reference series and index-computation conventions are rarely stated.A retention index alone does not identify whether the Kováts or van den Dool and Kratz convention applies; the correct convention depends on oven operation.Source: §5 · p. 8

6 Recovery from prose

• None of 355 gas chromatography studies has a structured temperature-programme field, although 279 (78.6%) describe one in protocol prose.The extractor reads ordered temperatures, ramp rates, and durations while excluding values tied to decoy terms such as injector or ion source.Source: §6 · p. 8
• 206 ordered programmes (58.0% of studies) were recovered, comprising 529 temperature steps, 370 with hold times, and 135 distinct programmes.The values survive in prose rather than the schema.Source: §6 · p. 8
• Three fresh-draw audits tracked extractor revisions, with the ordered-token version scoring 11 of 12 and then 10 of 14.The remaining failure mode omits a hold rather than inventing one, leaving a visible gap instead of a fabricated measurement.Source: §6 · p. 8
• Retention-index reference series were structured for 3.3% of analyses, but recovered from prose in 66 studies (18.6%), a 5.6-fold increase over the structured baseline.All twelve randomly sampled recovered series were adjudicated correct.Source: §6 · p. 8
• Internal-standard identity extraction was withdrawn after a random audit returned four compounds and eight English fragments among twelve audited candidates.The extractor reported candidates in 138 studies, but nothing it produced was released.Source: §6 · p. 8

7 The perceptual record

• Among eight datasets declaring human odour-character measurement, 42.9% of pairs share no descriptor and mean pairwise Jaccard overlap is 0.061.Across all readable archives the disjoint-pair rate is 86.7%, partly because three archives measure rodents or an ant receptor assay rather than human perception.Source: §7 · p. 9
• A declared measurement-type tag in each archive’s manifest distinguishes human panels from other measurement types, but analyses of data tables alone cannot make that distinction.Stemming left the disjoint-pair count unchanged, and fuzzy matching rescued only two of ninety-four disjoint pairs in the unrestricted comparison.Source: §7 · p. 9

7.1 Agreement between identical descriptors

• Against a permutation-null threshold, 29 of 3,358 descriptor pairs survive, but 24 link atlas descriptors to the modern study’s pleasantness item.Only five of 24 modern descriptors receive any relation.Source: §7.1 · p. 10
• For twelve pairs sharing an identical descriptor word, two clear the threshold and mean correlation is 0.31; sweet against sweet is 0.59.The record cannot determine whether apparent disagreement reflects perception or concentration: the atlas’s high/low field is assigned from molecule names, while matching concentration formats leaves only seven molecules.Source: §7.1 · p. 10

7.2 The unused olfactometric reference

• Across 53 archives, none mentions olfactometry, odour units, dynamic dilution, the standard, or n-butanol as a reference.The search covers curated manifests and tables, not the underlying publications; n-butanol appears as a stimulus in 26 archives and as a reference in none.Source: §7.2 · p. 10

8 The bridge

• The bridge keeps volatility measurements and percept descriptors joined at the molecule, with each side retaining its own provenance and neither treated as evidence for the other.The model can retain disagreement between public datasets about whether a compound has an odour, rather than merging the claims.Source: §8 · p. 10, Figure 7 · p. 11
• 132 molecules link volatility spectra to percepts, backed by 434 spectra from 6 laboratories and 147 descriptors.54 of the 132 molecules have measurements from multiple laboratories, allowing cross-laboratory agreement and a percept to be stated together.Source: §8 · p. 10
• Restricting the join to gas-chromatography volatility measurements reduces an earlier count of 1,710 molecules to 132; liquid-chromatography spectra do not measure volatility.Including those spectra would admit compounds on evidence unrelated to smell.Source: §8 · p. 10

9 Determinants of field completeness

• The interpretive field ranges from 100% for seismology’s instrument response to approximately 0% for volatile chemistry’s thermal programme.Metrology’s measurement technique is 53.8%, while meteorology’s sampling strategy is 16.6%; all four fields are optional.Source: §9 · p. 11
• A field counts as complete when the primary consumer cannot complete the primary task without it; comparison may be a later consumer’s task.A chromatogram still opens without its thermal programme, although comparison requires the programme.Source: §9 · p. 12
• Despite mandatory international metadata policy, meteorological sampling strategy appears in 16.6% of records, and sample treatment appears in 6.6%, all marked unknown.Instrument height is the best-populated field at 62.1%.Source: §9 · p. 12

10 Learnability of the derived corpus

• The evaluation used 3,522 molecules with 113 binary odour labels, twenty structure-string character counts, ridge regression, and five-fold cross-validation.The transfer test trained on this corpus and tested on a separately curated corpus of 4,565 molecules.Source: §10 · p. 12
• Mean AUC was 0.816 against a shuffled-label null of 0.481, with 45 of 51 labels above 0.70; transfer mean AUC was 0.772.All 50 testable labels in the separately curated corpus were above 0.60.Source: §10 · p. 12
• The authors describe the result as a floor, not a frontier: the weak structural features establish signal in public data, not performance approaching published odour prediction.They state that a learned representation would improve on these features.Source: §10 · p. 12
• Aggregate checks must fail closed on empty input, because a join returning no records can otherwise satisfy a verdict and appear to pass.The authors report the same empty-input fault in their conformance layers and in a completeness table computed after a failed network fetch.Source: §10 · p. 12

11 The conformance ledger

• The ledger reports machine-readable, versioned open findings with severity, count, and the condition that would clear each one.The count becomes a tracked quantity; repositories audited currently report holdings but not defects.Source: §11 · p. 13
• The released graph has zero structural violations and 531 open warnings, all flagging percepts without concentration as incomparable across panels.The warning remains deliberate because perceived quality can change or reverse with concentration.Source: §11 · p. 13, §12 · p. 14

12 The released corpus

• The derived graphs are released so results can be inspected and queried without re-harvesting the public sources.Reproduction otherwise requires re-harvesting several repositories, one of which declined further automated access.Source: §12 · p. 13
• The release includes 33,115 derived-comparison triples, 10,514 bridge triples for 132 molecules, and 3,633 recovered thermal-programme triples.It also contains 444 perception-provenance triples, 360 recovered-reference-series triples, and 290 inferred-descriptor-relation triples.Source: §12 · p. 13
• The largest component, 86,226 interlaboratory-comparison triples, is withheld because no source licence was declared and the publisher declined further automated access.Withholding it costs the deposit roughly two thirds of its triples.Source: §12 · pp. 13–14
• Each recovered value retains its source sentence as evidence and is attributed to the extracting pipeline, not the original depositor.This keeps an extracted claim about what the text says distinct from a statement made by its author.Source: §12 · p. 14

13 Limitations

• The comparability ladder transfers graded, separately tested levels across domains, but not psychometrics’ factor-analytic apparatus; the scalar rung is an upper bound.The factor-analytic apparatus is not used in this paper.Source: §13 · p. 14
• Only 16 of 53 perception archives carry a compared vocabulary, and spelling normalisation merges variants but not synonyms.Some remaining archives are reference compilations rather than experiments, making overlap conservative on spelling but optimistic on meaning.Source: §13 · p. 14
• The bridge treats percepts as descriptors asserted for molecules, not intensity ratings at stated concentrations; resolution through one registry makes 132 a floor.Source: §13 · p. 14
• The literature comparison counts terminology co-occurrence, not citations, so it is a lower bound on contact rather than evidence of isolation.A citation-level analysis was not performed.Source: §13 · p. 14
• The metrology corpus is a contiguous slice of the identifier space, not a sample, so proportions should not be computed from it.The publisher declined further automated requests partway through the permitted crawl.Source: §13 · p. 14

14 Data and code availability

• The corpus is deposited with versioned access, seven OWL modules, six SHACL layers, and 69 regression tests; graphs are regenerable where source licensing is unclear.Headline numbers are computed both set-based and by graph query, with scripts exiting non-zero on disagreement.Source: §14 · p. 15