wcamon/circus-0.3-agentic-repair

Elicit, Attribute, Repair

This repository is the archival release for “Elicit, Attribute, Repair: Agentic Behavior in a Memory-Architecture MoE at a 120B-Token Budget.” It contains the paper PDF, LaTeX source, figures, number ledger, and checksums.

The paper studies a 5.89B-total / 0.537B-active memory-architecture MoE. A paired oracle-name intervention and factorial analysis localize a tool-use binding failure; a targeted GSPO round then moves BFCL v4 non-live Simple AST from 20.25% to 31.25%, while HumanEval pass@8 falls from 25.00% to its frozen 15.62% floor.

Scope and limitations

• This is a single seed, single architecture case study, not a scaling law.
• No interactive evaluation was run; BFCL multi-turn and tau2-bench are absent.
• Truncation makes several generation scores lower bounds and can inflate abstention-style metrics.
• Capability-profile raw artifacts were lost; the paper clearly marks the affected section as ledger-derived and not independently reproducible.
• The release makes no emergence claim and does not claim broad knowledge gains.

Related artifacts

• Repaired checkpoint
• Pre-repair checkpoint
• Minimized evaluation evidence
• circus-0.2 binding diagnosis
• circus-0.2 STEM architecture
• circus-0.3 release collection

AI assistance

Claude (Anthropic) assisted with experiment execution, analysis, and drafting. Wei Ciao Wu selected the research questions, reviewed the evidence, and takes responsibility for all claims and the release.

Citation

Until an arXiv identifier is assigned, cite this archival release:

@misc{wu2026elicit,
  title  = {Elicit, Attribute, Repair: Agentic Behavior in a Memory-Architecture MoE at a 120B-Token Budget},
  author = {Wu, Wei Ciao},
  year   = {2026},
  url    = {https://huggingface.co/wcamon/circus-0.3-agentic-repair}
}