1# Plan/protocol vs code — discrepancies reported by the Part IV writer (2026-11-03) and their disposition23| # | Discrepancy | Disposition (Agent A) |4|---|---|---|5| 1 | ontology.yaml note says "max over matched topics"; classify.py sums weights then caps at 1 per signal | code is the rule; YAML note to be corrected (taxonomy/src) |6| 2 | VALIDATION.md promises Krippendorff α (MASI); validate_sample.py computes mean Jaccard | add α (MASI) to validate_sample.py --metrics |7| 3 | Protocol E7 = penumbra; code e7 = generic-deposit exclusion | rename in protocol: E7 generic deposit (A-002), penumbra = separate flag, amendment A-005 |8| 4 | H5 says Gini on 24-month citations; code uses c36; c12 = pub year + next calendar year | report both c12 and c36 Gini; define windows as calendar-year windows explicitly (Part IV already does) |9| 5 | Placebo breaks run over baseline AND post months | add `--placebo-baseline-only` option and report both |10| 6 | Right truncation by last-complete-month cut + sensitivity, not lag correction; no donut windows / binary segmentation / sup-F | document; add donut-window option to ITS (exclude ±3 months); sup-F not planned (DP + BIC/LWZ kept) |11| 7 | ITS event = first full month after each date | document (already in text) |12| 8 | Semantic centroids by calendar-year periods; split-half ARI on one UMAP fit | document as implemented; optional bootstrap later |13| 9 | Emerging detector: topics only, volume-only label, fewer features, cutoffs to 2021-12, unstandardised logistic | add feature standardisation + novelty/cross-area merge when available; label with citation-share variant as sensitivity; cutoffs extended to 2022-09 (labels observable by 2024-09 for h=24) |14| 10 | No igraph author/institution networks; CAGR, bootstrap CIs, diversity indices not scripted | add bibliometrics/04_diversity_cagr.py (CAGR, Rao–Stirling, Leinster–Cobbold, bootstrap CIs) and institution/country co-authorship graph metrics |15| 11 | DBLP arXiv-step linked 0 records (DBLP now uses 10.48550 DOIs) | documented in Part III; DOI step already captures them |16