{
  "title": "Real Coding Decisions Local Pilot Receipts",
  "publication_date": "2026-10-07",
  "source_repository": "TrentCarter/llm-tuning-lab",
  "source_commit": "9fb9a99",
  "runner_checkpoint": "64bda7d",
  "model": {
    "identifier": "Mapika/decider-4b v2.1",
    "architecture": "Qwen3.5-4B-Base derivative",
    "format": "MLX affine Q4 (group size 64)",
    "weights_sha256": "d0ba3ca5953871030be574146a6eef2d8073d556041756647753921112afaa7f"
  },
  "dataset": {
    "identifier": "HumanEval 120-task mutant split (30 calibration, 90 evaluation)",
    "cases_sha256": "2f7994d4be5d3b80a040dde3422d1dd1dda2422536d3dd79409598769dbe87f9"
  },
  "files": {
    "summary.json": {
      "sha256": "6e6597ac7c8674e9fe90986b000f3ea148e24b898ccbaffc064b9ff724819e51",
      "description": "Aggregated benchmark metrics, confidence intervals, latencies, and calibration numbers."
    },
    "methodology.md": {
      "sha256": "4000024b123fb061be1ee135bbc1e55049064a6d79b5fa52992a2c18c4d1c19c",
      "description": "Full study methodology, experimental setup, and analytical disclosures."
    }
  }
}