Methodology

What is scored

An event is a documented beneficial or harmful outcome with a plausible link to one or more AI systems. A press release that says a model “accelerates science” is not an event. A specific result that later shows up in a paper, clinic, court, operational system, or widely deployed product can be.

Foundational architecture papers are out of scope. Attention Is All You Need made later systems possible; this index starts after that architecture existed and scores what named models actually did with it.

Significant product or feature launches are in scope when they change professional practice in an industry—GitHub Copilot for software engineering, Claude Code for agentic coding, the November 2022 ChatGPT deployment. Download counts, token volume, and “X million users” alone are not scores. The launch still needs evidence of adoption or a measured productivity effect, but that evidence belongs in the Evidence multiplier, not in a lowered base tier.

Benefit and harm stay on separate ledgers. Net Good is benefit minus harm. A model with large benefit and large harm should not look like a quiet success.

Contest performance, historical olympiad sets, and other benchmark exams are not Net Good. They are recorded on a separate Capability Ledger. Capability events stay on the public record; they do not enter benefit, harm, or leaderboard totals. Solving a known IMO problem is a capability milestone. Resolving an open Erdős problem is newly created knowledge and belongs on the Benefit Ledger.

Score formula

Points = Base Impact × Attribution × Evidence × Realization × Durability × Credit Share

Every multiplier is between 0 and 1. Credit shares across AI contributors to the same result cannot exceed 1.0.

• Tier I Useful / Limited harm: 1
• Tier II Notable / Significant: 3
• Tier III Major: 10
• Tier IV Historic: 30
• Tier V Civilizational / Catastrophic: 100

The four questions are separate. Do not answer the same uncertainty twice.

• Base impact is how important the accomplishment is if the claim is true.
• Evidence is confidence that the claim is true.
• Realization is how far the result has actually progressed from claim to outcome.
• Attribution is how much of that outcome the named AI system caused.

A claimed Millennium Problem solution is Historic with discounted evidence, not Major until someone reconstructs it. A claimed Alzheimer's cure would be Civilizational with very low evidence, not Major because it has not been replicated. A field-transforming product launch can be Historic without a measured GDP series, just as a protein-design method can be Historic without an approved drug. The score should rise as confirmation arrives rather than jumping both the category and the multipliers.

Domain rules

Mathematics. New theorems, constructions, and disproofs are realized knowledge. A public manuscript sits around 0.6 realization; a complete manuscript plus machine-verifiable formalization sits around 0.8–0.9; community acceptance is 1.0. A correct theorem is effectively permanent (durability ~1.0). If the proof is wrong, that is an evidence or retraction problem, not decay. Independent reconstruction raises evidence, not the base tier. A cluster of independent theorems published on one day is still a Historic body of work if the theorems would have been Historic or Major separately; publication format is not a 10-point ceiling.

Science. A new method that becomes standard practice can be Historic before downstream products exist. Replication and experimental validation live in evidence and realization.

Clinical. Phase 2a is Notable until Phase 3, approval, or a documented patient-outcome effect justifies Major. Peer review of a small trial raises evidence, not the base tier. An authorized clinical workflow can outrank an unapproved molecule.

Productivity and deployment. Changing how a profession works is in scope. Surveys, field experiments, and years of workplace use raise evidence. Missing GDP accounting does not cap the event at Major.

Civic harm. This is not a risk index. Score the realized incident. Deceptive calls with no demonstrated electoral effect are Significant harm, not Major, until turnout or outcome evidence appears.

Credit units

Organizations, families, and models are three different primitives. The scoring unit is a named model release—GPT-6 Astra, GPT-5.6 Sol, Gemini Deep Think—not a lab and not a product brand.

• Organization is the lab or company (OpenAI, Google DeepMind). Org scores are rollups.
• Family is a named generation or product line (GPT-6, GPT-5.6, Claude, Gemini). Family scores sum the releases inside it.
• Named release is the default model record. Pre-release and internally deployed snapshots of the same named line count as that release—GPT-6 Astra includes the in-development Astra that posted the August 2026 ten-proofs cluster. A lab still keeps a separate internal record when it documents a different unreleased system (for example, OpenAI’s September 2026 search model described as more capable than Astra).
• Internal / unreleased is its own model only while the lab has not identified it with a named release. Date it. When the identity is later learned, remap the contributions onto that release rather than leaving a permanent shadow model.
• If authors used “Claude” or “Codex” without a version, record a version-unspecified model in that family until a named release is identified.

Credit shares across AI contributors cannot exceed 1.0 when they contributed to the same result—a Lean formalization by Astra of a proof found by an internal system is a split. Independent systems that separately accomplish the same milestone are separate events. Do not force 50/50 credit on two non-collaborating results.

Rumors that Hodge, Birch–Swinnerton-Dyer, Riemann, Yang–Mills, or P vs NP have been solved are not events. Poincaré remains the only Clay Millennium problem with an accepted human solution. A lab announcement of a Millennium-scale result is still recorded, at Historic if that is the right magnitude, with evidence and realization held down until independent reconstruction or Clay evaluation.

Evidence and review

Developer blogs score as announcements, not proof. Peer review, independent replication, court records, and measured outcomes raise evidence. High-tier events require human review. Every published score change writes an immutable revision.

Provisional means the event is recorded and the claim may well be true, but independent reconstruction, acceptance, or downstream confirmation is still incomplete. That is not the same as disputed. Navier–Stokes-scale announcements can be provisional Historic results. Claims the mathematical community has not accepted, with substantive objections, remain disputed.

Disputed means there is a serious live argument that the claimed outcome is false or did not happen. Priority fights, credit fights, and disagreements about which official formulation matters should be flagged on the record. They do not by themselves make the event disputed.

Hourly ingest publishes matching items immediately as provisional events. They do not wait in the review queue. Routine items keep a conservative unspecified-model placeholder. Realized third-party harm and other significant claims get a magnitude-aware preliminary score on ingest — Major or Significant when the claim warrants it, with evidence and named-release identity held down until review. If the source already names a release, ingest maps that alias instead of leaving the placeholder. TypeSafe System One judges whether an ingest item is realized third-party harm, a notable beneficial development, or speculative-only. Those judgments choose the ledger and the preliminary score gate; the factor formula, source-type evidence discounts, and a cap at Major still live in code. Keywords remain the fallback if TypeSafe is unavailable. Editorial events are not rewritten by ingest classification.

The same ingest, plus a daily monitor pass, watches lab alignment blogs and misalignment-report pages — including OpenAI Alignment, Anthropic's Alignment Science blog, UK AISI, and METR — for realized third-party harm. Simulated eval writeups and speculative risk notes stay out of the ledger unless a documented real-world incident is attached. Follow-up reporting can update an existing event: an unspecified system later named, or a magnitude later revealed to be larger or smaller than first recorded. Those updates write a new immutable revision; they do not rewrite history. Automatic updates never promote an event to Historic or Civilizational tiers.

Retracted events remain on the record with a score of zero unless residual impact is documented. Speculative catastrophe, chatbot quality, token volume, and popularity are out of scope.

Corrections

Anyone can submit a factual correction, missing source, attribution dispute, or retraction notice. Submissions enter an editorial queue. Leaderboard numbers are never typed by hand.