Research Publications

Everything we have published, including the corrections.

Eight full reports and seven shorter outputs. Papers carry a report number and a version; when a number changes we say so in the paper rather than replacing the file quietly. Correspondence on any of it goes to research@aiworklab.com.

2026

AWL-2026-06 The comprehension gap: 200 engineers, 12 months, and what agentic coding did to recall M. Bhardwaj · AI Work Lab Research We instrumented the workflow of 200 working engineers across 14 companies for a year and measured two things that are usually reported separately: how much they… Paper14 July 2026

AWL-2026-05 Explain-to-merge: where one beat of friction changes what sticks M. Bhardwaj · AI Work Lab Research Friction that teaches and friction that annoys are separated by about ninety seconds and one design decision about timing. We ran four placements of the same… Paper26 June 2026

AWL-2026-04 Lines accepted predicts nothing: six productivity proxies against six-month capability M. Bhardwaj · AI Work Lab Research Every AI coding dashboard reports acceptance rate. We correlated six commonly-reported proxies against measured engineer capability six months later, on 168… Paper2 June 2026

AWL-2026-03 A concept tagger that does not lie: calibration across 41 repositories M. Bhardwaj · AI Work Lab Research The skill graph is only as honest as the classifier that populates it. We measured accuracy and calibration for four tagger designs across 41 repositories, and… Paper12 May 2026

AWL-2026-02 Spaced retrieval for code: FSRS against 18,000 review events M. Bhardwaj · AI Work Lab Research Spaced repetition was built for vocabulary, where an item is atomic and a lapse is cheap. Code concepts are neither. We fitted FSRS to 18,240 code review events and… Paper21 April 2026

AWL-2026-01 Model-agnostic by construction: teaching quality across nine models M. Bhardwaj · AI Work Lab Research Bring-your-own-model is a permanent stance, not a launch limitation, which obliges us to show that the teaching layer works on models we do not control. We measured… Paper18 March 2026

AWT-2026-04 Gating question generation behind a calibrated tagger M. Bhardwaj · AI Work Lab Technical Note Why we stopped asking a question for every diff, and what the 0.85 threshold costs. Technical report2 July 2026

· Comprehension checks as desirable difficulty in production developer tools M. Bhardwaj · Workshop on Learning at Scale for Software Engineering Short paper. Extends the placement result in AWL-2026-05. Workshop paperMay 2026

· concept-taxonomy-v2 and prerequisite graph M. Bhardwaj · Public release, CC BY 4.0 214 concepts, three levels of granularity, 3,410 curated prerequisite edges. DatasetApril 2026

· What we owe an engineer we are measuring M. Bhardwaj · FOSDEM, Brussels Slides and transcript published. 24 minutes. TalkFebruary 2026

2025

AWL-2025-02 Bus factor as a leading indicator: 14 organisations, nine months M. Bhardwaj · AI Work Lab Research Concentration of understanding is easy to measure and easy to ignore until someone resigns. We tracked module-level comprehension concentration across 14… Paper14 October 2025

AWL-2025-01 What an org dashboard should refuse to show: aggregation limits on individual skill data M. Bhardwaj · AI Work Lab Research We built a skill graph accurate enough to be dangerous. This paper sets out the re-identification results that determined which views we ship, which we deliberately… Paper20 August 2025

AWT-2025-03 Normalising acceptance telemetry across four vendor definitions M. Bhardwaj · AI Work Lab Technical Note Four tools, four definitions of accepted. The mapping, and what it cannot fix. Technical reportNovember 2025

· Comprehension-weighted ownership as a maintainability signal M. Bhardwaj · Preprint, under review Extends the bus-factor result in AWL-2025-02 to defect prediction. Comments wanted. PreprintSeptember 2025

· Against per-engineer skill scorecards M. Bhardwaj · AI Work Lab Research Why we will not ship an individual comprehension score to management at any tier. Position paperJuly 2025

Citation

BibTeX for the eight reports.

If you are citing a technical note, a dataset or a talk instead, write to us and we will send you the entry rather than have you guess at it.

@techreport{bhardwaj2026comprehension,
  author    = {Bhardwaj, Mohit},
  title     = {The comprehension gap: 200 engineers, 12 months, and what agentic coding did to recall},
  year      = {2026},
  number    = {AWL-2026-06},
  institution = {AI Work Lab},
  url       = {https://aiworklab.com/research-1.html}
}

@techreport{bhardwaj2026explain,
  author    = {Bhardwaj, Mohit},
  title     = {Explain-to-merge: where one beat of friction changes what sticks},
  year      = {2026},
  number    = {AWL-2026-05},
  institution = {AI Work Lab},
  url       = {https://aiworklab.com/research-2.html}
}

@techreport{bhardwaj2026lines,
  author    = {Bhardwaj, Mohit},
  title     = {Lines accepted predicts nothing: six productivity proxies against six-month capability},
  year      = {2026},
  number    = {AWL-2026-04},
  institution = {AI Work Lab},
  url       = {https://aiworklab.com/research-3.html}
}

@techreport{bhardwaj2026concept,
  author    = {Bhardwaj, Mohit},
  title     = {A concept tagger that does not lie: calibration across 41 repositories},
  year      = {2026},
  number    = {AWL-2026-03},
  institution = {AI Work Lab},
  url       = {https://aiworklab.com/research-4.html}
}

@techreport{bhardwaj2026spaced,
  author    = {Bhardwaj, Mohit},
  title     = {Spaced retrieval for code: FSRS against 18,000 review events},
  year      = {2026},
  number    = {AWL-2026-02},
  institution = {AI Work Lab},
  url       = {https://aiworklab.com/research-5.html}
}

@techreport{bhardwaj2026model,
  author    = {Bhardwaj, Mohit},
  title     = {Model-agnostic by construction: teaching quality across nine models},
  year      = {2026},
  number    = {AWL-2026-01},
  institution = {AI Work Lab},
  url       = {https://aiworklab.com/research-6.html}
}

@techreport{bhardwaj2025bus,
  author    = {Bhardwaj, Mohit},
  title     = {Bus factor as a leading indicator: 14 organisations, nine months},
  year      = {2025},
  number    = {AWL-2025-02},
  institution = {AI Work Lab},
  url       = {https://aiworklab.com/research-7.html}
}

@techreport{bhardwaj2025org,
  author    = {Bhardwaj, Mohit},
  title     = {What an org dashboard should refuse to show: aggregation limits on individual skill data},
  year      = {2025},
  number    = {AWL-2025-01},
  institution = {AI Work Lab},
  url       = {https://aiworklab.com/research-8.html}
}

Access

What you can have, and what we hold back.

Open

Instruments and code

Rubrics, taxonomies, the prerequisite graph, normalisation code and every analysis notebook are published outright. Reproducing our figures should not require our permission.

On request

Labelled corpora

The 12,400-span labelled corpus contains code from private repositories, so it moves under a data agreement rather than a download link.

Never

Individual records

No per-engineer comprehension record leaves the product, to us or to anyone else. The reasoning is set out in AWL-2025-01, along with the requests we declined.