metadata
language:
- en
license: mit
library_name: transformers
tags:
- code
- swe-bench
- reasoning
- moe
- recursive-intelligence
metrics:
- code_eval
model-index:
- name: RI-Meta-Core-Grounded-DeepSeek-V4.1
results:
- task:
type: coding
name: Software Engineering Code Resolution
dataset:
name: SWE-bench Verified
type: SWE-bench/SWE-bench_Verified
split: test
metrics:
- name: swe_bench_%_resolved
type: swe_bench_%_resolved
value: 0
⚡ RI-Meta-Core-Grounded-DeepSeek-V4.1 (Pure Autonomous Zero-Shot)
Recursive Intelligence (RI) Meta-Core evaluated strictly on raw problem statements across 500 tasks of SWE-bench Verified with 0% Gold Contamination.
📊 Empirical Autonomous Metrics (No Gold Leakage)
| Metric | Pure Autonomous Result | Technical Detail |
|---|---|---|
| Gold Contamination Rate | 0.00% | Zero oracle leaks, 100% blind generation |
| Valid Git Diff Syntax | 83.60% (418/500) | Correct diff headers, file pointers, and hunks |
| Target File Identification | 36.20% (181/500) | Zero-shot file discovery from issue text alone |
| Target Patch Perplexity | 4.12 (Best: 2.24) | High token-level confidence and code structure |
| Testbed Harness Execution | Docker Containers | Official SWE-bench evaluation harness verified |
All predictions are available in predictions/swebench_verified_pure_500_honest.json.