Completed example — AI-Assisted Research Charter

Project: Beyond PISA rankings: moving international benchmarks and UK trajectories
Researcher / team: Simon Rudkin and collaborators
Discipline / field: Quantitative social science / social statistics / education comparison
Status: Retrospective completion from the mature project record

Template history: the blank charter now includes a distinct pre-use permission gate (introduced in the v1.1.0 resource revision). This older retrospective example does not document whether permission assessments occurred at a particular earlier point. No missing contemporaneous permission record is invented here.

1. Substantive research problem

International rankings can change even when a system’s own performance deteriorates, because the comparison field is also moving.

The research problem is therefore to represent education-system performance in a way that distinguishes:

own-score movement;
rank movement;
movement relative to a changing international comparison field;
overlapping multidomain similarity.

2. Existing researcher contribution

Before AI assistance, the project already had:

an established researcher interest in rankings and multidimensional social outcomes;
prior work using Ball Mapper / topological data analysis;
substantive knowledge of the UK education-system context;
a preference for treating England, Scotland, Wales and Northern Ireland as analytically distinct systems.

3. Analytical object

Published PISA education-system mean performance in reading, mathematics and science for 2018, 2022 and 2025.

Key design choices:

education system is the unit of analysis;
England, Scotland, Wales and Northern Ireland are treated separately;
the aggregate United Kingdom observation is excluded to avoid double counting;
the main balanced panel contains 73 education systems;
Ball Mapper is used to represent local overlapping neighbourhood structure;
PCA and hard clustering provide comparator analyses.

4. Non-negotiable constraints

AI-assisted exploration must not silently:

aggregate the four UK administrations into a single UK observation;
include the aggregate UK observation alongside the four administrations;
treat radius 50 as fixed without re-testing it;
suppress the PISA 2022 UK sampling-standard caveat;
describe generated code as empirical evidence before local execution and inspection;
infer direct same-system flows from overlapping endpoint memberships when that coupling is not identified;
present Ball Mapper as replacing simpler global summaries when comparator evidence does not support that claim.

5. Permitted AI roles

6. Verification boundary

Generated code becomes evidence only after:

versioned delivery;
local execution;
production of logs / handbacks;
inspection of the resulting objects;
reconciliation with the research design.

Generated interpretations remain proposals until checked against the identified research object and inferential target.

7. Consequential decisions to record

Record when the project:

adopts or rejects a consequential AI-originated question;
fixes the unit of analysis or inclusion/exclusion rules;
changes the preferred methodological claim after a comparator;
withdraws an interpretation because the design does not identify it;
revises language after robustness analysis;
authorises the manuscript claim boundary;
discovers a problem through direct inspection of the finished scholarly object.

8. Escalation rule

Any AI-generated claim that would strengthen the contribution, identify movement between states, alter the analytical object or remove a qualification must return to empirical or substantive review before entering the manuscript.

9. Release rule

Only evidence and wording that can be supported from the frozen project record should appear as factual claims about the research process.