Claims audit¶
Files audited: README.md, report\ch00_front_matter.md, report\ch01_introduction.md, report\ch02_landscape.md, report\ch03_analysis.md, report\ch04_latest_techniques.md, report\ch05_related_work.md, report\ch06_method.md, report\ch07_experiments.md, report\ch08_discussion.md, report\ch09_conclusion.md. Paragraphs and tables with numbers: 109; with a resolving claim id: 105; marked [verify]: 4; uncovered: 0. Claim ids cited: 111. Unresolved ids: 0. Cited rows lacking a source file or commit: 0. ‘first’ used as a contribution claim: 0. Result: PASS
Uncovered paragraphs (a number with no resolving claim id)¶
none
Unresolved claim ids¶
none
Cited rows without source file or commit¶
none
‘first’ as a contribution claim¶
none
Pending (marked [verify]; to be checked by the researcher)¶
report\ch01_introduction.md:3: > Status: draft written by Claude Code for the researcher to rewrite in their own words and defend; updated 22 Oct 2026 for the finished Tie…
report\ch03_analysis.md:41: The two stages count within-window variation twice, so the interval is wider than a plain standard error would give: about the square root o…
report\ch06_method.md:3: > Status: draft written by Claude Code for the researcher to rewrite and defend; completed for Tiers 1 and 2 on 22 Oct 2026. Every section d…
report\ch07_experiments.md:3: > Status: draft written by Claude Code in Phase 6 for the researcher to rewrite, verify and defend. Every number carries a claim id from `do…
What the results do not show¶
That the gap between random-split F1 and time-aware AUT is caused by temporal leakage: the per-slot comparator reaches a lower malware ratio than the time-aware series on LAMDA, and the E0 and per-slot comparators disagree about the 0.05 band on LAMDA (P-14, E-19).
A delay curve on BODMAS: with nE = 6 and retraining every third slot there is one useful refit per window, and periodic equals never at a delay of three by construction (P-12).
Independent replication on LAMDA: its 16 rolling windows are about four independent 12-month blocks, and logistic regression is deterministic given its pool, so seeds are draws of the C3 subsample and of the selection, not replicates of the model (P-12).
That more labels do not help when a flat result appears at the smallest budget: retraining changed 0.2% to 2.1% of decisions in the earlier diagnostic (P-13).
That conformal credibility adds anything over probability-based uncertainty (E3 does not separate them), or that a drift monitor beats periodic retraining: in E4 none did at a budget of 1% (E-59, E-60).
Anything about other platforms or EMBER2024: E5 and E6 (Tier 3) have not been run.
Anything about evasion or adversarial examples: the study describes decay, not defeat (docs/ETHICS.md).
Results on another machine: every cell was run on one laptop; numeric drift elsewhere is reported, not corrected (docs/GRID.md).