Version the evidence, not just the model name.

Claim records, not floating facts

An AI claim is stored or written with enough context to remain interpretable: exact model or system name, version when available, organization, claim type, direct source, publisher, publication date, access date, and a caveat when the evidence is incomplete or self-reported. “Model X can do Y” is narrowed to what was demonstrated, under which conditions, and by whom.

First-party technical reports and system cards are primary evidence for what their publisher disclosed. They are not independent verification. Research papers can provide method and results but may not generalize beyond their dataset. Product pages can document available features but can change without preserving a historical record.

Benchmark results need a comparison class

Flintglade does not combine unrelated evaluations into a universal intelligence score. A published benchmark result should retain:

• benchmark name and version or dated release;
• exact evaluated model, system, or agent;
• metric, unit, and whether higher or lower is better;
• split, subset, language, and evaluation date where applicable;
• prompting, tools, scaffold, pass count, cost limit, or other material setting;
• publisher, primary result source, and whether the result is first-party or independently reproduced;
• comparability class and a specific caveat.

A result published by an evaluation team may be displayed as a published third-party result. It is not relabeled as a Flintglade run. Results from different benchmark versions, contaminated test sets, different tool access, rolling question pools, or retired evaluations are separated rather than sorted into one leaderboard.

Benchmark definitions and benchmark scores are separate

A definition explains what an evaluation was designed to measure, its dataset scale and task format, known constraints, and the responsible primary source. A result snapshot reports what a particular source published under a particular setting. Keeping those objects separate allows the Compendium to describe HumanEval, MMLU-Pro, ARC-AGI-2, OSWorld, SWE-bench, or another evaluation without implying that every reported score can be compared directly.

Research, theory, and mathematics

Mathematical entries identify notation, assumptions, and the level of abstraction. An equation is accompanied by a plain-language interpretation and its role in a model or learning process. Theoretical discussions are labeled as theory, conjecture, scenario, or open question; they are not presented as a released product or demonstrated capability.

Physics, quantum computing, and AI futures

Work at the boundary of AI and the physical sciences distinguishes mathematical analogy from physical mechanism, classical simulation from quantum execution, and proposed advantage from empirical evidence. Research into scientific machine learning, physical modeling, quantum information, or quantum algorithms records the computational setting and validation route needed to interpret the result. A future-facing scenario is labeled as a scenario rather than presented as an established trajectory.

Trends and future systems

Trends are based on multiple dated signals, such as peer-reviewed or preprint research, official releases, standards work, documented adoption, and reproducible practitioner evidence. A company announcement supports the fact that something was announced, not a prediction that it will succeed. Future-product pages distinguish confirmed roadmaps, reported development, inference, and speculation.

Human review and AI assistance

Models may assist with classification, drafts, cross-references, or candidate-source discovery. They are not accepted as a source. Material factual entries require a human-checkable record and deterministic validation for required fields, URL schemes, dates, section membership, and internal relationships. A generated citation is rejected unless the linked source exists and directly supports the claim.

Corrections, deprecation, and uncertainty

Entries expose publication and revision dates. A benchmark can be marked retired, superseded, saturated, contaminated, disputed, or rolling without erasing its historical value. Model names are not silently reused for newer versions. Material corrections change the affected claim and revision date; source updates become new evidence rather than pretending the original always said something else.

Reference standards and examples

• National Institute of Standards and Technology, AI Risk Management Framework.
• National Institute of Standards and Technology, Generative AI Profile.
• Stanford Center for Research on Foundation Models, Holistic Evaluation of Language Models.
• National Institute of Standards and Technology, AI RMF Playbook.

Published 13 July 2026 · Last materially reviewed 13 July 2026 · Research corrections: support@flintglade.com