Luceria research: methods, data and results
Small, reproducible studies about the mechanics of job applications, built so that anyone can read the method and download the files; the parser benchmark's measures can be recomputed from its published extracted text.
To see what a parser reads from your own file, run the parser preview in your browser.
Measured studies
Method v1.0.0 · measured 4 November 2026
Parser benchmark: what Luceria's extractor reads from nine resume layouts
A reproducible benchmark of Luceria's extractor on nine synthetic resume layouts: text recall, reading order, bullets, headings and contact details.
- Fixtures measured9
- Mean token recall99.5%
- Mean reading-order score85.9%
- Fixtures with all four headings recognised9 of 9
Method v1.0.0 · measured 4 November 2026
Document repair lab: five one-change fixes, measured before and after
Five resume layout fixes, each changing one thing, measured before and after with Luceria's parser: reading order, contact details, headers and text layers.
- Fixtures measured12
- Repairs measured5
- Reading order, two columns to one column (PDF)54.5% → 95.5%
- Reading order, layout table to one column (Word)54.5% → 95.5%
Defined, not yet measured
These studies have a published method and no results. They are listed here, without their own pages, until the data and conditions exist. Nothing below is a finding.
Pending · method v0.1.0
Where applications stall: stage patterns from consented tracker data
This study is defined but not yet run. It would count, in aggregate and with no individual-level data, the final stage recorded for applications in Luceria's tracker by role family and market. Results will be published only when the aggregation, disclosure and consent conditions below are met and the sample is large enough to describe honestly.
Method and limitations
Aggregation and disclosure rules
Counts would be computed server-side over applications whose owners have the analytics consent record and are grouped by role family (derived from the job title field with a fixed taxonomy) and market (UK, US, other). No cell with fewer than 50 applications or 20 distinct accounts would be published, and no free text, employer name or user identifier would be read or stored by the analysis.
Measures and bias
Final tracker stage (one of the twelve stages from Saved to Archived) as a share of applications per group, with the share of applications still open stated separately. Users who track in Luceria are not a representative sample of job seekers, and people stop updating stages after an outcome, so results would be reported as patterns in this sample with those caveats, never as hiring odds.
Limitations
- Self-selected sample of Luceria tracker users; not representative of the labour market or of any employer's process.
- Stage updates depend on users recording outcomes; silence and abandoned records bias final-stage shares.
- No individual, employer or hiring-outcome inference is possible or intended; the study describes tracker records, not employer decisions.
Pending · method v0.1.0
How job adverts phrase requirements: essential, desirable and unlabelled
This study is defined but not yet run. It would apply the published requirement-extraction rules from Luceria's free worksheet tool to a documented sample of public adverts drawn from the job sources Luceria already indexes, and report how requirements are labelled. Results will be published with the sample definition, date window and extraction version, or not at all.
Method and limitations
Sample and extraction
A dated sample of adverts from Luceria's supported public job feeds, deduplicated by employer and title, with the market recorded. Each advert's text would be processed by lib/tools/requirement-evidence.ts to count requirement lines and their essential, desirable or unlabelled kind. Advert text would be processed in memory and not republished.
Reporting
Shares of adverts with explicit essential or desirable headings, median requirement lines per advert and the distribution of kinds, by market and broad sector, with the extraction version and its known limits stated.
Limitations
- The sample reflects the sources Luceria indexes, not all adverts; feeds over-represent some sectors and markets.
- Extraction follows fixed rules that miss unconventional wording; the study measures what those rules find.
- Says nothing about how employers actually score applications or make hiring decisions.
How Luceria publishes research
- Every published number comes from code and data you can download and run; the write-up is not the record, the data file is.
- Fixtures are synthetic and labelled; no study reads or republishes a real person's documents or applications.
- A study is listed before it has results only as a method. Charts and figures appear when the measurement has been made, never before.
- Each study states what it does not test. Luceria measures its own software; it does not test employer systems or predict hiring outcomes.
- Corrections are recorded on the study page with the date, and methods are versioned so earlier results stay comparable.
Journalists, researchers and educators may reuse the published data files with attribution to Luceria and a link to the study page. Questions and corrections: contact Luceria. Related free tools: parser preview and resume check.