Research library
Progress toward AI-managed advancement
Assessment date: November 7, 2026. This extends the mathematics rubric to every research area without changing the existing mathematics scores.
Endpoints
0%: the project starting point before ChatGPT's public release in 2022. This is a convention, not a claim that AI or automation began in 2022.
100%: AI manages the field's continuing advancement without human participation in choosing worthwhile problems, planning work, developing methods, executing research, validating results or improving the research process. Reliable execution of a fixed human-designed task alone cannot meet this endpoint.
The endpoint concerns the management and progression of the field. Patients, learners and other beneficiaries may still be human. Existing instruments, machine checkers and computing infrastructure may support the process; necessary ongoing human scientific direction, validation, troubleshooting or repair prevents a 100% rating.
Scoring rubric
Each outcome receives five component judgments, each worth at most 20 points. Their sum is its displayed percentage.
| Dimension | Requirement for all 20 points |
|---|---|
| Execution | Dependably performs the outcome's work to the required scientific or practical quality. |
| Breadth | Transfers across the field's relevant domains, conditions and genuinely unfamiliar problems. |
| Agenda-setting | Selects worthwhile questions, objectives and research plans without human direction. |
| Validation | Checks correctness, intended meaning, relevant novelty and practical validity without human repair or adjudication. |
| Sustained autonomous improvement | Maintains and improves the process over repeated research cycles without human management. |
Apply the same anchors used for mathematics:
| Points per dimension | Evidence interpretation |
|---|---|
| 0 | No demonstrated post-anchor autonomous progress in that dimension, or insufficient evidence to assign credit. |
| 5 | Isolated or heavily supported demonstrations. |
| 10 | Repeatable bounded performance with material human involvement. |
| 15 | Broad, mostly autonomous performance with remaining gaps. |
| 20 | Dependable autonomous performance throughout the relevant scope. |
Equal weights are a project judgment, not an empirically calibrated law. Scores use five-point increments to avoid spurious precision. Component values are also judgments, not independently measured quantities. A 30% rating is the sum of these assessments, not evidence that 30% of the work, discoveries or difficulty has been completed.
Evidence and uncertainty
Every score has low confidence. It is derived from the dated findings already saved in that outcome file; this scoring pass adds interpretation, not new experimental measurements or a fresh exhaustive literature search. Follow each file's baseline links for source dates, systems, budgets, denominators and limitations.
Missing evidence receives no demonstrated-progress credit in the affected component. Thus 0% can mean an evidence gap, a pre-2022 reference with no documented later increment, or no qualifying result. The accompanying explanation distinguishes these cases. None proves the absence of relevant AI capability.
A nearby enabling demonstration can justify limited partial credit where its relevance is explained, but is not proof that the specified outcome has been achieved. In particular, a project announcement, simulated result or improved control algorithm cannot be counted as measured grid reliability, physical confirmation or net electricity.
Independent human review remains useful evidence for this assessment. It does not imply autonomous validation by the AI. Good performance with substantial researcher assistance therefore earns less autonomy credit than the same quality achieved independently. Negative results, corrections and newly disclosed assistance can reduce estimates.
Interpretation and updates
These percentages are not benchmark accuracy, percent of jobs replaced, the probability of a singularity, a prediction of its date, or a literal measure of how much science remains. Equal numerical increments need not require equal time or effort.
Do not average outcomes into an overall area score: the indicators overlap, differ in scope and include evidence gaps. Do not treat the area tables as a precise ranking. Even all listed outcomes reaching 100% would require separate field-wide evidence that AI sustains independent advancement.
Preserve the November 7, 2026 estimates and append future dated assessments. For each update, record the old and new component values, the new evidence, changed human roles, and why each changed component crossed an anchor. Keep empirical benchmark series separate from these estimates. If the rubric changes, version it and rescore a common snapshot rather than silently treating the new scale as historical progress.