Skip to content

STATS.md

Latest commit

 

History

History
37 lines (26 loc) · 1.58 KB

STATS.md

File metadata and controls

37 lines (26 loc) · 1.58 KB
Raw

DeepInit — computed stats (STATS.md)

Generated by tools/build_stats.py from the committed validation records + the harness summary. Do not hand-edit — every figure self-derives. Re-run python tools/build_stats.py.

Harness: 473/473 checks PASS across 106 oracle sections. External metamorphic oracle (§26): recall 14/22 = 64% (Wilson95 LB 43%); metamorphic-FP 0 (the hard gate). Recall is INDICATIVE — §18 stays the headline.

Mirror Test (core-output coverage vs the project's own docs)

HELD-OUT (8 repos: commercetools-sync-java, eabe, helix, kagent, kotlinx-schema, saluki, scrapy, uniffi-rs) — the REPORTED number:

  • Coverage 126/192 = 65.6% (Wilson95 LB 59%)
  • Faithfulness 144/147 = 98.0%
  • The one hard gate deepinit_wrong_high Σ = 0; beyond-doc 482

Per-kind coverage (pooled held-out):

  • component-exists: 36/44 = 82%
  • component-role: 36/43 = 84%
  • dependency-edge: 11/15 = 73%
  • data-store: 3/4 = 75%
  • boundary-rule: 9/28 = 32%
  • key-invariant: 8/28 = 29%
  • entry-point: 7/8 = 88%
  • technology-choice: 16/22 = 73%

TUNE (not reported): excalidraw, otelcol — coverage 37/43 = 86.1%.

Precision (naive-vs-guarded, real repos)

  • 90 naive false positives avoided across 4 repos, 0 false defects.

Cross-language field sweeps

  • 15 stacks swept (C, C#/.NET, C++, Crystal, Elixir, Go, Java 21, Kotlin, OCaml, PHP, Python, Ruby on Rails, Rust, Swift, TypeScript / React).

Cost

  • 1 ledger(s); measured range ~$1.28–$8.25 (tier-S, INDICATIVE). No published $ until a clean api_usage run.