AI research: completion history

All area pages | Area definition | Scoring rubric

Created November 7, 2026. The three earlier snapshots are retrospective judgment-based estimates, not scores recorded at the time. The current assessment is unchanged. Historical confidence is very low; current confidence is low. Zero can indicate an evidence gap rather than absent capability.

Chart layout updated November 8, 2026: one graph with four dated lines. Scores and assessment dates are unchanged. Overlapping lines indicate equal values; point labels show November 7, 2026. The table below contains every date's exact values.

Scalable graph | Combined 14-page report

Percentage estimates

Reconstruction notes and component scores

Components below are execution / breadth / agenda-setting / validation / sustained autonomous improvement, each out of 20. Their sum gives the percentage. The component values are judgments, not published statistics. Source links document evidence available by the cutoff, not the numerical estimate itself. The source section lists availability dates and version caveats. No area-wide average is calculated.

Acceleration

Narrow feedback supports partial credit at all past cutoffs. The organizational acceleration assessment was published November 5, so its added execution/agenda evidence appears only in the current score.

Algorithmic efficiency

The optimization evidence already existed in May 2025. Carry the same bounded judgment rather than invent gains from the later December restatement.

Experiment efficiency

Earlier bounded search lacks full campaign efficiency accounting. November's organizational workflow assessment supports only a small additional planning judgment.

Research autonomy

Bounded agent research is documented before every historical cutoff. Human-defined tasks constrain agenda credit; the November assessment supports the current limited agenda component.

Research task performance

The same dated bounded-task evidence supports all snapshots. No comparable new task series is established here, so the score is held flat.

Successive improvement

Narrow feedback into model training predates 2025's cutoff, but does not establish a causal multi-generation series. November's assessment modestly strengthens planning evidence.

Valid discoveries

Accepted optimizations predate every snapshot. Missing independent multi-seed replication prevents an inferred upward trajectory.

Dated sources

All sources accessed November 7, 2026. Dates below are publication or conservative availability dates, as noted.

evolve25 | 2025-05-14 - AlphaEvolve announcement. Evaluated code optimization and narrow feedback into AI training; human-defined evaluators.

rebench | 2025-05-27 - RE-Bench research paper. Research-engineering evaluation; bounded tasks and fixed budgets, not a self-directed laboratory.

Future updates

Keep these reconstructions separate from subsequently recorded observations. Append new dated assessments with changed component scores and source evidence; do not silently replace historical reasoning. The shared historical-method notes describe how these snapshots were made.