Research notes · rubric 1.0
What makes an event significant?
An event is one dated, sourced record in the editorial archive. An announcement is a fact. Its place in history is an assessment. This archive keeps those separate: sources document each event, while a transparent editorial rubric determines visual prominence.
5 criteria, one explainable score
Weights divide 100 available points across 5 criteria. Rate each criterion from 0 to 5, then multiply its rating divided by 5 by its available points. Add those contributions to get the total score. The score is an editorial assessment, not a probability or a share of all AI impact. No recency bonus, popularity counter, or provider ranking is added. Confidence is reported separately. All 114 records have individual ratings and written rationales. Scoring each item against a fixed rubric with a written rationale is also how human evaluators grade AI models' answers, paid work that AITraining.jobs' guide to AI evaluation work, built by the same founders as this site, explains step by step. The original archive rubric was reviewed 2026-09-10; full reports show their own review dates.
Conceptual shift
30% · 30 points maxHow much did this change the available approach?
- 0No material change
- 1Incremental refinement
- 2Meaningful extension
- 3New capability within a field
- 4New general approach
- 5A foundational conceptual change
Downstream influence
25% · 25 points maxHow much subsequent research or practice does it underpin?
- 0Not demonstrated
- 1Isolated follow-up
- 2A model family or specialist line
- 3Multiple documented successors
- 4Many methods or applications
- 5A foundation used across the field
Breadth of impact
20% · 20 points maxHow widely did the change matter beyond its original setting?
- 0Not established
- 1Single experiment
- 2Specialist community
- 3Several fields or workflows
- 4Multiple sectors or a large public audience
- 5A field-wide or society-wide change
Access & diffusion
15% · 15 points maxHow much did it enable others to use or build on the work?
- 0No usable access
- 1Restricted demonstration
- 2Limited or controlled access
- 3Published method or widely available service
- 4Reusable code, weights, or substantial practical access
- 5A durable, broadly reusable enabling resource
Institutional change
10% · 10 points maxDid it change lasting infrastructure, governance, or incentives?
- 0No demonstrated change
- 1Local workflow change
- 2A new organizational commitment
- 3A durable program, standard, or platform
- 4Cross-sector changes to infrastructure or norms
- 5A new field, binding framework, or systemic institution
3 visual priorities
Defining. Field-shaping events; emphasized at the widest scales.
Major. Substantial changes; labeled as space becomes available.
Notable. Significant within a narrower scope; revealed on closer inspection.
Point size and label hierarchy reflect the tier. Labels compete for available space by score, so the canvas stays legible. Every included event remains available through its point, search, and direct dossier. Zooming reveals the surrounding detail.
Evidence, inclusion, and uncertainty
We include documented conceptual advances, enabling datasets and infrastructure, influential systems, material changes in access, and institutional turning points. A model release alone does not establish a breakthrough. Our coverage is curated, not a claim to record every significant event.
Original papers, project documentation, model cards, institutional archives, and official announcements are preferred. Each dossier links its evidence and distinguishes publication, announcement, release, and retrospective dates where relevant. Year-only events are placed at midyear as a navigation convention.
Established means a documented contribution with historical perspective. Supported means the contribution is documented but broader influence is less settled. Provisional marks especially recent developments whose lasting influence cannot yet be established. These labels are editorial, not statistical probabilities.
Provider claims are attributed, not treated as independent performance verification. Scores are comparative judgments with overlapping criteria and a risk of hindsight, English-language, availability, and recency bias. A score of 81 is not meaningfully more precise than 80. Future evidence may change both ratings and inclusion.
The archive is reviewed editorially. New developments from any lab, research group or government are weighed against the same standard, and one is added only together with its dossier and full report: no announcement becomes a scored event automatically.
How adaptive time works
Each year receives a base width plus a density allowance, using the average event count across that year and its two neighbors. The width multiplier is min(6, 1 + 2 × √density). Quiet years keep a nonzero width. Spacing uses the full archive so applying a category filter does not move history.
The resulting mapping is reversible. The canvas, cursor zoom, and overview all use it. Adaptive spacing is explicitly non-linear: physical distance is not elapsed time. Labels change from years to months to days as you zoom.
Data usage
The event index (each event's date, names, organization, category, summary, impact, ratings, score, confidence and primary source) is open data under CC BY 4.0: download it from the AI history dataset page and reuse it with credit. The long-form dossiers and research reports are provided for reading and reference and remain all rights reserved. Linked papers, model artifacts, and other source materials retain their respective rights and licenses.