Fellowship · HQ-W27 · Upcoming
Evidence over claims: are the AI systems that read, match, grade and source people accurate, fair and reliable?
- Starts
- 4 Jan 2027
- Length
- 12 weeks
- Format
- Remote
- Pay
- Unpaid, educational
- Tracks
- 4
- Apply by
- 20 Dec 2026
HQ-W27 is the second cohort of the HyperQuark Research Fellowship, an unpaid, educational research programme: remote, twelve weeks, from Monday 4 January to Sunday 28 March 2027, with about 12 fellows. AI now grades people's work, ranks their applications, maps their skills and sources them for jobs. Each track tests one new claim about one of those systems, on public, open-licence or synthetic data with open models, to the standard of a publishable paper, and builds a release so anyone can check it.
Four tracks, one question each
- Track 1
Judges You Can Trust
Whether open AI judges that pass the standard human-agreement check still score the same answer lower when it is rewritten into Indian English with the content held fixed, against a UK English control.
OutputAims at a paper and a public variety-invariance benchmark for AI judges.
- Track 2
Fair Matching, India Edition
Whether bias in open résumé-screening pipelines runs through caste-coded names or through proxies such as college tier, city tier and English variety, measured against a placebo edit, and which mitigations close the gap at what cost in ranking quality.
OutputAims at a paper and a gated counterfactual benchmark.
- Track 3
India Skills and AI-Exposure Graph
A validated crosswalk from India's 8-digit occupation codes to ESCO skills and O*NET, and whether scoring AI exposure from India's own task text changes the picture that transferred scores give.
OutputAims at a paper and an open crosswalk with its gold set.
- Track 4
Reliable Recruiting Agents
Whether invented candidates and unsupported claims explain why open sourcing agents that succeed once fail on later runs, and whether a cheap deterministic check fixes it.
OutputAims at a paper and a public closed-pool sourcing benchmark.
Twelve weeks, five phases
Twelve weeks in five phases: problem comprehension, direction lock, a quality gate at week 5 that decides a track's scope and never who stays, defence and validation, then write-up and release. Each track names the strongest prior result it must beat, fixes its claim and analysis before seeing results, and reads the best work of 2025 and 2026. A research pack opens each phase. Fellows share a short report each Sunday, and track leads answer it in writing. It takes about six to eight hours a week at your own schedule, with one weekly cohort session, so it fits alongside study, and you may leave at any time.
Who may apply
Enrolled students and unaffiliated researchers (not employed to do research in the track's field), aged 18 or over, who live in and take part from an eligible country. No PhD is needed.
- India, Singapore and JapanEnrolled students and unaffiliated researchersEligible
- South KoreaEnrolled students aged 19 or overEligible
- United StatesEnrolled students only, not resident in California or New YorkEligible
- United Kingdom, Canada and AustraliaOnly where your course requires the placement and your institution confirms it in writingWith conditions
- Germany, France, the rest of the European Union and the European Economic Area, Israel, and every country not listed hereNot eligible for this cohortNot eligible
What fellows receive
Mentorship and written feedback, access to the lab's studio and guidance on free compute for the work, a certificate anyone can verify if you meet its requirements, and credit by name on what the lab publishes from your work. The fellowship is unpaid, now and later. It is not employment and not a route to a job.
What fellows produce
For each track: a paper in the format of a venue the track chooses (or an honest retrospective if the claim does not hold), a release of the benchmark, dataset or crosswalk it built with the code to rerun every number, and a recorded demonstration. The lab decides whether, what and when each is published, and credits the fellows who did the work by name on what it publishes.
Data and ethics
- Public, open-licence or synthetic data only, open-weight models and published code
- Fellows never receive insiderOne product or user data, in any form
- No data collected from people, no scraping of people and no experiments on people
- No dataset or result is ever linked to insiderOne member records
- Systems are audited with published methods, and no company is named as discriminatory
Fellowship agreement
Every fellow signs the HyperQuark Research Fellowship Agreement (HQ-FA 2.0) before taking part. Under it, everything made in the fellowship belongs to Aurist Private Limited, which runs insiderOne and the lab, and may be used in any product, tool or research without payment or approval. The company alone decides whether, what and when anything is published, and fellows are credited by name on what it publishes from their work. Taking part gives no claim, now or later, to payment, equity, a job or ownership. Fellows keep what they owned before they joined.
Selection
We look for curiosity, rigour and the ability to finish. Backgrounds in machine learning, natural language processing, statistics, economics and labour data, fairness and AI policy, or software engineering are all relevant. The application includes one short written question on checking a research claim. Applications close on 20 December 2026.
HQ-W27 · Applications open until 20 Dec 2026
Apply to be part of HQ-W27
Starts 4 Jan 2027 · 12 weeks · Remote · Unpaid
An unpaid, educational fellowship for students and researchers in eligible countries. Fellows sign the Fellowship Agreement before taking part. Read it before you apply.