docs(research): a withdrawn naming-consistency rule, reconstructed as evidence — first pass, negative result

[joyful-ii-V-I]

Follows Wang, Zhang, Jiang, Tang, Li & Liu, Deep Learning-Based Identification of Inconsistent Method Names: How Far Are We? (EMSE 2025, doi:10.1007/s10664-024-10592-z, arXiv:2501.12617): DL-based naming-inconsistency detectors collapse once scored against a realistic, unbalanced distribution instead of the balanced benchmarks earlier evaluations used. The row this traces through is our own withdrawn deterministic rule, already recorded at docs/LINEAGE.md L104 (MNire / the CIC name-versus-body proxy). Full write-up: docs/research/naming-consistency-withdrawal.md.

This is a first pass, and the honest result is a withdrawal, not a feature. No src/ changes — one document plus a one-off sampling script behind its base-rate estimate. (Base note: measured against main at 755f9026, before 0.6.2; the diff still merges cleanly onto today's main, 15a20855, tag 0.6.2.)

Headline finding, our own negative result. src/naminglens.h shipped a deterministic name-vs-body vocabulary-overlap rule (naming-body-mismatch) and pulled it the same day — measured on ripwire's own source it produced 159 of 217 (73%) of the naming lens's findings and flagged the best-named functions in the tree. This PR reconstructs that withdrawal as reproducible evidence (the old commit rebuilt and rerun, not just quoted), generalizes the lesson against the paper's own finding and against ripwire's existing rename-history calibration instrument, and measures an honest, capped base-rate estimate of genuine name/body inconsistency in real code: a seeded sample of 38 judged (name, body) pairs across four real OSS checkouts plus our own src/, 0/38 judged genuinely inconsistent, reported as a rule-of-three ~95% upper bound (≈7.9%) and explicitly not a point estimate. It states plainly what ripwire ships today that touches names versus what it does not — none of it is a semantic name-accuracy judgment.

This is a first pass with a lot further to push, and we would like critique on two open questions: whether any deterministic, non-learned axis could clear a realistic-population precision floor for this kind of check at all, and whether a better-labelled real-world dataset exists to tighten our base-rate estimate past a self-sampled, 38-function upper bound.

Our working premise across this line of investigations: algorithmic, deterministic checks applied while an AI writes are the practical way to keep code sound at the speed AI now writes it — no label, no vendor claim, a check that either fires or it doesn't. This withdrawal is exactly that premise holding itself to its own bar.

🤖 Generated with Claude Code

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

_{Comment @coderabbitai help to get the list of available commands.}