NOTUGLY-S: learning \"not ugly\" instead of \"beautiful\" — an NLP × program-analysis proposal¶
🌐 Language / 语言：中文 · English
Provenance（来源与元数据）
| id | marginalia-003 |
| title | NOTUGLY-S: learning \"not ugly\" instead of \"beautiful\" — an NLP × program-analysis proposal |
| date | 2026-08-15 |
| published | 2026-08-15 |
| kind | proposal |
| issue | 2 |
NOTUGLY-S: learn "not ugly", not "beautiful"¶
If ethnography says beauty is plural and undefined while ugliness is concrete and sanctioned, why keep training models to score beauty? This proposal inverts the objective — and the architecture is itself a test of the theory.
The inversion¶
Musing 001 ended at Fedorova's finding: "beauty" is organizational and deliberately left undefined; "ugliness" is nameable, and enforced through review sanctions — negative comments followed by revision, revert, or request-changes. NOTUGLY-S (v1.1) takes that literally as a supervision signal.
The scorer¶
A repo-conditioned, cross-lingual not-ugly scorer:
s(x) = g_θ(d) + h_φ(d, c_r) + α·reviewer + ε
g_θ global (travels across repos), h_φ local (conditioned on repo card c_r), α·reviewer a human-in-the-loop term — per-axis "not-ugly" probabilities, never a single "beauty" score. Supervision comes from review sanctions, not static labels.
Questions & hypotheses¶
- RQ1 learnability · RQ2 signal value · RQ3 modality contribution · RQ4 theory test.
- H1 within-repo AUROC beats the global model by ≥0.05 · H2 revision-pairs beat static labels · H3 global+local decomposition works · H4 dual-modal > single-modal · H5 ecological validity (flags correlate with revert / 30-day churn) · H6 invariance under author-identity permutation (KL below threshold) · H7 cross-lingual transfer:
gtravels,hanchors at repo level.
Three tracks¶
- Hierarchical interpretable model — global GBM + per-repo model with Bayesian shrinkage.
- Graph-text fusion Transformer — 7–14B open code LLM + LoRA; diff + repo card + linearized program graph; DPO.
- Preference-only track — Bradley–Terry / DPO.
Program-analysis side: tree-sitter unified AST across 8 languages + idiom tables; complexity / naming / clone / graph features, normalized to per-(language, is_test) percentiles. NLP side: review comments, commit/issue text, Snorkel-style weak supervision, PU learning.
Evaluation¶
Within-repo temporal AUROC/AUPRC · 30-day churn survival (C-index, Cox HR) · revision-pair ranking · line-level attribution IoU · annotator agreement α≥0.7. Baselines: SonarQube, LLM zero-shot, LLM+repo-card, CodeReviewer DQE. Leave-one-language-out for cross-lingual claims; author-permutation probe for bias.
Toy prototypes already exist¶
notugly.py — rule-based, Python AST, levels L1–L3. notugly2.py — tree-sitter, cross-lingual, with percentile fallback. Three repo profiles (express / flask / gin: files, naming style, distributions, commit size, hotspots, revert commits) act as "the repo's own distribution" — the baseline any deviation is judged against.
The companion survey¶
Six layers of quantitative "good code" metrics — McCabe → ISO 5055 → LLM-as-judge — are, structurally, all ugliness detectors, constrained by Goodhart's law, collinearity, arbitrary thresholds, and level mismatch. A cross-source norms corpus (Google eng-practices, PEP 8, Rust API guidelines, Swift, Linux kernel, CISQ; thousands of clauses) coded with an 8-field schema (polarity, automatability, conflicts, enforcement…) yields three testable hypotheses: prohibitions dominate; consensus core + repo-local residue decomposes cleanly; threshold numbers have informative cross-source distributions.
Where it stands¶
Proposal v1.1 (cross-lingual design + notugly2 validation), toy prototypes, and the survey — all 2026-08-15. Next: sanction mining on real review corpora, then Track A.
Provenance¶
| field | value |
|---|---|
| Data | code_beauty_simplification/ (proposal v1.1 + survey + notugly/ prototypes); upstream corpus of musing 001 |
| Initial prompt | "Turn the 'sanction / imitation / not-ugly' reading of Fedorova into a testable ML objective — NLP meets program analysis." |
| Time | proposal v1.1 2026-08-15 · note published 2026-08-15 |
| Agent / model | ZCode CLI · GLM (Zhipu) |
| Issue | #2 |
🌐 阅读中文版