Status. An HQ-S26 research note: work done by fellows of the HyperQuark Research Fellowship between April and July 2026, written up by the lab from the track's own files. The scores are one fellow's provisional, single-rater scores of two systems; the proposed comparison across kinds of system was not done. It has not been reviewed outside HyperQuark, and its code and data have not been released.
The question
If an AI system makes or shapes a decision about you, who answers for it, and how would you find out? Governance frameworks tell organisations what processes to have. They do not let an outsider look at a deployed system and say how accountable it is, and the checklists that do produce a score tend to need insider access and rely on self-reporting.
The track set out to build a rubric an outsider can apply from public documents alone, and to test whether accountability documentation differs by kind of system: conversational assistants, recommenders and automated decision systems.
HQ-S26 was HyperQuark's first cohort: twelve fellows in four tracks, twelve weeks, weekly reports due every Sunday. No track reached a final submission, so every result here is interim and is reported as the track left it.
The rubric
Five criteria. Each is scored 0 (not documented), 1 (implied), 2 (documented but incomplete) or 3 (fully documented), and every score must cite the public document it rests on.
| Criterion | The question it asks | NIST AI RMF |
|---|---|---|
| Responsibility | Does the operator name who is accountable for the system's decisions? | GOVERN 1.1 |
| Redress | Can a person affected by a decision contest it, and how? | MANAGE 4.1 |
| Audit access | Can outside parties examine the system? | MEASURE 2.5 |
| Incident documentation | Are failures and incidents recorded publicly? | GOVERN 5.2 |
| Human oversight | Can a person step in and override the system? | MANAGE 2.2 |
A first scoring
One fellow applied the rubric to two frontier chat assistants, from their published system cards (April 2026) and terms of use. These are provisional, single-rater scores; the fellow flagged one as uncertain.
| Criterion | Claude Opus 4.7 | GPT-5.5 |
|---|---|---|
| Responsibility | 0 | 0 |
| Redress | 1 | 1 |
| Audit access | 3 | 2 |
| Incident documentation | 3 | 2 or 3 (uncertain) |
| Human oversight | 3 | 2 |
The finding that holds even at this stage is the first row: neither operator's public documents name a party accountable for the system's decisions, and neither offers an affected person more than an implied route to contest an outcome.
Limitations
-
One rater, two systems. No second rater scored the same documents, so agreement between raters is unknown.
-
The central comparison was not made. No recommender or automated decision system was scored, so the claim that accountability differs by kind of system is untested.
-
Whole-number scores proved coarse: the rater wanted a half-point for one criterion.
-
Public documents only. A system can be more accountable in practice than its documents show, or less.
What would settle it
Two or more raters scoring the same documents; systems of all three kinds, hiring systems especially, where the stakes for a job seeker are highest; and a check that scores agree across raters before any comparison is drawn.
Contributions
-
Tanzila Kaiser (Track Lead): set the direction; designed the rubric and its mapping to the NIST framework.
-
Iris Vukovic: scored the two assistants against the rubric, with citations, and compared it with the EU's ALTAI checklist.
-
Subhajit Paul: compared governance approaches and proposed a trust-scoring direction the track considered.
-
Bishnu Dev Changkakoti (Programme Director): set the fellowship's research agenda and directed the track.
Sources
The two assistants' published system cards (April 2026) and terms of use; NIST AI Risk Management Framework 1.0 (2023); the EU High-Level Expert Group's Assessment List for Trustworthy AI (ALTAI, 2020).



