README.md

---
license: other
license_name: sofron-project-license
pipeline_tag: text-generation
tags:
  - code-generation
  - agent-architecture
model-index:
  - name: sofron-v7 (combination)
    results:
      - task:
          type: text-generation
        dataset:
          name: openai_humaneval
          type: openai_humaneval
        metrics:
          - name: pass@1
            type: pass@1
            value: 90.2
        source:
          name: Sofron HumanEval benchmark writeup (own harness, self-reported)
          url: https://github.com/s4h8a/sofron-benchmark-results
  - name: sofron-v7 (solo)
    results:
      - task:
          type: text-generation
        dataset:
          name: openai_humaneval
          type: openai_humaneval
        metrics:
          - name: pass@1
            type: pass@1
            value: 89.0
        source:
          name: Sofron HumanEval benchmark writeup (own harness, self-reported)
          url: https://github.com/s4h8a/sofron-benchmark-results
---

# Sofron v7

Sofron is a combination-architecture coding agent: its own model weights (v7)
paired with advisory teacher models (GLM-5.2 and DeepSeek) gathered in
parallel as candidates on each problem. Sofron synthesizes its own final
answer from their input, using its own judgment as the foundation rather
than deferring to any single advisor.

**This is a metadata-only repository.** No model weights are hosted here;
it exists to record a real, execution-graded, self-reported benchmark
result with a citable, verifiable source.

Contact / homepage: https://sofron.one

## HumanEval result (self-reported, not third-party verified)

| Condition | Pass@1 (164/164 problems) |
|---|---|
| combination (v7 + GLM-5.2 + DeepSeek synthesis) | **90.2%** (148/164) |
| solo (v7 alone) | 89.0% (146/164) |

Full methodology, raw results log, and harness scripts:
https://github.com/s4h8a/sofron-benchmark-results

Sofron combines advisory input from external teacher models with its own
model weights — it is not a frontier proprietary model. This result should
be read as an honest, real placement rather than a state-of-the-art claim.