Proposed
A construct, requirement, instrument, or explanation offered for criticism and testing. It has not yet been empirically established.
Methods and versions
The status of a claim should remain visible.
A conceptual proposal, a preliminary observation, and a replicated result can all be useful. They cannot carry the same evidentiary weight. This page records the distinctions used across the research program.
Evidence status
Four states used on this site
Labels describe what kind of support a claim presently has. They are not quality scores, and they should change when the underlying evidence changes.
Preliminary evidence
An observation supported by limited evidence whose scope does not yet justify broad generalization.
Published analysis
An argument available in a public repository with identifiable authorship, date, citation, and stated limitations.
Unresolved
A question, conflict, or boundary that remains open because the available evidence or reasoning does not settle it.
Evaluation practice
Minimum publication record
Any released evaluation should allow a reader to reconstruct what was tested and what the resulting score leaves out. Aggregate performance is insufficient when failures, refusals, incomplete runs, or scoring assumptions materially change the interpretation.
Task record
Publish the task purpose, source materials, required output, staged changes, and criteria used to determine success. When the full corpus cannot be released, provide a representative inspectable subset and explain the restriction.
Configuration record
Identify the model, version, settings, tools, prompt, run date, and number of repetitions. Report completion and failure separately from accuracy on successful runs.
Ground truth and review
Record how the answer key was created, who reviewed it, how disagreements were reconciled, and where model assistance entered the process. Human review improves an answer key; it does not remove the need to disclose how the key was produced.
Limits and conflicts
State corpus limits, unresolved ambiguities, known sources of bias, sponsorship, and relationships that could affect task selection, scoring, or interpretation.
Versioning
Revision is part of the record
A correction should not erase the existence of an earlier claim or instrument. Material changes receive a new version with a date, a description of what changed, and a reason. Superseded versions remain identifiable unless legal or ethical constraints require removal.
- Version 0.xDesign-stage instrument. Criteria and tasks may still change substantially.
- Version 1.0First released instrument with documented tasks, scoring, review, and limitations.
- Version 1.xCompatible correction or extension that preserves the instrument’s central construct.
- Version 2.0+Material change to construct, corpus, scoring, or interpretation that limits direct comparison.
Independence and assistance
Independence requires a record
Perfinitive is presented as an independent research program. That description does not eliminate possible conflicts. Any released evaluation should identify its funding, model or data access, compute support, provider participation, advance review, and other relationships that could affect task selection or interpretation.
OpenAI’s Codex assisted with recovery of the prior website, implementation of this redesign, source discovery, and editorial drafting. AI assistance does not determine which claims are retained or what evidence is sufficient. Micheal Preble remains responsible for the published language and research decisions.
Current boundary: no model evaluation results are published on this site. A result-specific disclosure must accompany any future release.
Current record
Site and evaluation status
The four papers listed in the publication archive are publicly available through SSRN. The evaluation and case-registry materials described on this site remain proposals in design. No benchmark scores, released datasets, or replicated findings are presently claimed.
Last reviewed: September 2026. Material changes to this policy should be dated and summarized here.