Claims have boundaries
Technical conformance, documentary reconstruction, causal effectiveness, legal conformity, and institutional adoption require different evidence.
A claim should remain attached to the evidence, method, and conditions that support it.
These commitments guide the Registry and the research it describes. The private preview keeps unnormalized research states and draft case records visible as such. Source review is a documentary check; it does not constitute independent replication or a completed incident investigation.
Technical conformance, documentary reconstruction, causal effectiveness, legal conformity, and institutional adoption require different evidence.
An incomplete record may leave a question indeterminate. Absence from a search result does not establish absence from the underlying record.
Record the source, date, scope, relationship to the claim, and known limitations. Identify derived analysis separately from source observations.
Preserve the method version, evidence packet, and decision rule that produced each result.
A material correction identifies the earlier state, what changed, why it changed, and which dependent claims were affected.
A passing schema or test suite demonstrates only the properties it examines. Other claims require separate testing.
Name the actual role: external scorer, reviewer, coauthor, replicator, funder, or adopting institution.
Material AI assistance should be disclosed. Human authors remain accountable for evidentiary judgment and publication.
Commissioning an evaluation should not transfer authority over its evidentiary conclusion. Factual review, confidentiality review, and approval of a finding are separate powers.
| Field | What the record should disclose |
|---|---|
| Question authority | Who selected the question and success criteria |
| Funding provenance | Who paid and what material interests exist |
| Method authority | Who selected or changed the method |
| Evidence access | What could be inspected and what remained unavailable |
| Review rights | Who could correct facts, request redactions, or challenge interpretations |
| Publication rights | Who could restrict publication and on what grounds |
| Conclusion authority | Who retained responsibility for the evidentiary judgment |
An assurance conclusion is bounded by the investigator’s access as well as by the evidence discovered.
Research integrity →METR’s investigation guidance discusses access to models, transcripts, environments, employees, and training evidence, together with disclosure of engagement terms and redaction effects. Read METR’s methodology ↗
External methodological context. This source does not validate TAE, HIT, GDI, CDFI, or Node & Norm’s case classifications.
| Investigation condition | Required disclosure |
|---|---|
| System and environment access | Full, partial, reproduced, documentary only, or unavailable |
| Transcripts and logs | Access scope, sampling, missingness, and known integrity limits |
| Model and personnel access | What could be queried or independently examined |
| Search process | Queries, tools, sampling, and limits of AI-assisted evidence discovery |
| Redactions and withheld evidence | Known restrictions and their effects on supportable conclusions |
| Resources and authority | Time, personnel, scope, publication rights, and conclusion control |
| State dimension | Question |
|---|---|
| Publication | What version has been frozen or released? |
| Technical | What specification or implementation exists? |
| Research maturity | What empirical progress has occurred? |
| Claim | What does the evidence support? |
| Review | What independent examination has occurred? |
| Adoption | What external institutional use is documented? |
Corrections identify changes to evidence, method, interpretation, and claim boundaries separately. Source history and Node & Norm record history must also remain distinct.
Inspect a changing source record →The case drafts were prepared with AI-assisted source discovery, page review, and drafting. No private system access, interviews, complete log inspection, or original model experiments were conducted. Human editorial admission remains pending.