HyperQuark Intelligence Labs · San Francisco Bay Area · Global research
Counterfactual, reproducible audits of the large language models and learned ranking systems that advise, score and screen people for work: calibration, bias under perturbation, and the rules that govern them. Coordinated from the San Francisco Bay Area, with fellows and research around the world.
Method
How the work moves
Eight steps every piece of HyperQuark research goes through. Watch the core take the shape of each.
- 01QUESTION
Name the decision AI is making about a person, and who lives with the outcome.
Every track starts from one question about a real decision, stated in a sentence.
- 02EXPLORE
Read what is already known: the rules, the systems in use, the prior studies.
Fellows read the rules, the systems in use and the prior work before proposing anything.
- 03MODEL
State the claim precisely enough that it could turn out to be wrong.
By week 3 each track locks one claim: “We propose … and demonstrate …”.
- 04EXPERIMENT
Test it on a fixed set of cases, chosen before the results are seen.
A fixed set of test cases and a baseline, chosen before any result is seen.
- 05EVIDENCE
Keep every result, including the ones that do not help the claim.
Results go into a frozen evidence table, including the failures.
- 06CHALLENGE
Put the work in front of reviewers whose job is to break it.
Another track reviews the work; then the HyperQuark Review Board scores it.
- 07REVISE
Change the claim to fit the evidence, and date the change.
Claims narrow to what the evidence holds. Every change is dated.
- 08CONTRIBUTE
Publish the method, the data where the law allows, and the limits.
A paper or an honest retrospective per track, with its data and code, under the fellows' names, citable by its permanent link.
HyperQuark Intelligence Labs is an applied AI research lab. It studies how AI systems screen, rank and advise people looking for work, around the world, and publishes what it finds, including what did not work.
HQ-W27 · Applications open until 20 Dec 2026
Apply to be part of HQ-W27
Evidence over claims: are the AI systems that read, match, grade and source people accurate, fair and reliable? · 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.
Take part
Three ways into the lab
- HQ-W27 · applications open
Research, credited by name
Twelve remote, unpaid weeks in a track, a report every Sunday, written feedback every week, and credit by name on what the lab publishes. Applications close 20 Dec 2026.
Apply to HQ-W27 - Open to anyone
Write for the lab
Essays, research notes and preprints on AI and work. Draft in the studio, run the checks, and the editor reads every piece before it is published under your name.
How writing works - Past fellows
Your record and certificate
Everything your cohort was given, your track's work and your own, with the certificate you were sent. Sign in with the address you used, or ask for access.
Sign in Ask for your record
Inside the lab
Questions in, open evidence out
Papers, evaluations, Review Board decisions and the fellowship, all moving through one lab and published in the open.
Research
What we study
One lane and four programmes around it. Each asks a single question about AI and people, and links to the work that answers it so far.
- Programme A · The lab's lane
AI in hiring and careers
When software screens, ranks or advises a job seeker, is the decision fair, lawful and explainable?
The rules that govern automated decisions about job seekers, open audits of AI career advice, and what a fair process looks like, across jurisdictions worldwide.
- Hiring and profiling law
- Role-fit and ranking
- AI career advice
- Programme B
Reasoning and evaluation
Does a model's reasoning actually support its answer, and can an AI judge tell?
How to tell whether a model's reasoning holds up, with benchmarks built to fail loudly rather than flatter.
- Reasoning faithfulness
- AI judges
- Benchmarks
- Programme C
Agentic systems
When software acts on a person's behalf, where does it help, where does it fail, and where must a human decide?
Software that takes steps on a person's behalf: where it helps, where it fails, and where a human must decide.
- Tool use
- Retrieval
- Recovery from failure
- Programme D
Governance and ethics
Can an outsider check whether a deployed AI system is accountable?
Documentation, accountability and trust for systems that affect people's working lives.
- Accountability
- Documentation
- Public scoring
- Programme E
Skills and the labour market
What does a role really ask of a person, and which of those skills is AI changing?
Open maps from occupations to skills, built on public labour data, with AI-exposure evidence from several sources and a plain account of where they disagree.
- Skills taxonomies
- Occupation data
- AI exposure
The chain
From human capability to impact
Six links the lab works along. Choose one to see what we can show for it today.
Link 1 of 6
Human capability
What people can actually do, shown rather than asserted.
What we can show today
- Skill receipts on insiderOne, checkable by anyone
Latest
Recent writing
- Research
Within-role, proficiency-aware discrimination on profile-derived skill graphs
Two candidates can hold the same skills at very different depths. Skill-counting cannot tell them apart. An HQ-S26 track built role graphs from real profiles and tested a score that can.
- Research
An accountability rubric for deployed AI systems: design, and a first scoring
Frameworks tell organisations how to govern AI; few let an outsider score whether a deployed system is accountable. An HQ-S26 track built a five-criterion rubric from public documents and tried it on two frontier chat assistants.
- Research
Recovering from weak retrieval: design notes for a research query assistant
A single search-then-answer pass assumes the question finds the right documents. An HQ-S26 track designed an assistant that notices when it did not, and tries once more. This note records the design and what the track observed.
- Research
Cross-channel disagreement as an operational diagnostic for LLM reasoning faithfulness
When an AI judge rates another model's reasoning, it can reward fluency over correctness. An HQ-S26 track found cases where frontier models reasoned their way to a wrong answer, and both judges rated the reasoning fully faithful.
Fellowship
Twelve weeks, one question per track
A remote, unpaid research fellowship. Fellows work in tracks, report every Sunday, defend their work, and are credited by name on what the lab publishes.
Connect
Work with us
Prospective fellows
Applications for HQ-W27 are open until 20 Dec 2026.
Researchers and institutions
Bring a question, a dataset or a method you want tested in the open.
Press and funders
Ask us anything, including how the lab relates to insiderOne. Write to contactus@insiderone.in.