Progress

System record

AIDE² (September 2026 study)

Organization: Weco AI · base model: Claude Opus 4.7 (outer loop), Gemini 3 Flash (selection inner loop)

System and setup

Reported system information
Model versionnot reported
Software used to run the modelAIDE² nested harness search
Software versionarXiv:2609.26457v1
Tools available to the modelCode editing and task evaluators; fixed selection rule
Release dateNot reported

Autonomy assessments

tracker evidence synthesisbounded

ai rd frontier model development

Framework levels: L5 · assessed 2026-09-26

Assessor: RSI Tracker (Codex evidence synthesis)

Revised research harness is retained and runs later; the separate outer-improver test invokes an evolved harness. Stronger effective recursion remains inconclusive.

Decisions or controls outside the AI system: Private evaluator; Task families; Budget; Outer selection rule

Evidence against this assessment: rec-aide2

Extraction: agent checked. This checks whether the record matches its source; it does not mean the experiment was repeated independently.

tracker evidence synthesisbounded

software engineering

Framework levels: L5 · assessed 2026-09-26

Assessor: RSI Tracker (Codex evidence synthesis)

Same harness mechanism viewed at the software-engineering boundary; not a domain-wide stage assignment.

Decisions or controls outside the AI system: Task families; Protected grading

Evidence against this assessment: None cited in this assessment

Extraction: agent checked. This checks whether the record matches its source; it does not mean the experiment was repeated independently.

Recursive-improvement evidence

tracker evidence synthesis

Assessment by RSI Tracker (Codex evidence synthesis) · 2026-09-26 · extraction review agent checked.

Resource budget:known · See study resource accounting; evaluation constraints are distinct from cumulative search cost.·Independence:known · Source-author report; no independent replication recorded.

Retention

not reported

Not recorded in this dataset.

Evidence history

Published capability results

No capability result currently links to this system.

Related research

AIDE²: inherited harness and outer-improver testpreprint · 2026-09-22Primary source status belongs to the linked paper; the tracker assessment is separate.
Additional system details
System ID
aide2
Organization
Weco AI
Base model
Claude Opus 4.7 (outer loop), Gemini 3 Flash (selection inner loop)
Release date
Not reported
Assessment records
2
Recursive evidence records
1