Roadmap¶
What may come after 0.1.0. Research items come from chapter 9 section 9.5 of the report. Each needs its own pre-registration (an addendum and a prereg-vN tag) before any run, because a change chosen after seeing results is no longer a test. Nothing here is promised. Order and timing are the researcher’s.
Releases¶
Version |
Content |
Precondition |
|---|---|---|
0.1.x |
Fixes only: documentation, packaging, dependency refreshes that leave every stored number unchanged |
quarterly dependency refresh (docs/MAINTENANCE.md) |
0.2.0 |
The 8-page workshop paper ( |
after the capstone hand-in (10 Dec 2026) |
0.3.0 |
Any of the research items below that is registered and run |
its addendum and tag |
Research items (chapter 9 section 9.5)¶
Tier 3: E5, a cross-platform comparison, with the rule for disagreement written in advance; E6, EMBER2024 (
reports/G5-tier3-decision.md).Drift monitors at other settings: a larger sample per slot, a stricter significance level, or another reference rule, reported with the alarm rate [E-63].
A longer BODMAS evaluation window, so that the delay axis has more than one refit (needs an ADR).
Label noise and measured delay: noise injected at a stated rate, and an observed delay distribution if one can be obtained.
A trap arm whose released labels also train the model [E-45].
Controls for class prior: a random baseline matched to the malware share, and a threshold-free score (PR-AUC) from stored predictions [E-66, E-47].
A registered lifetime measure: threshold and start slot fixed in advance, per-window curves, an interval [E-68].
A label-using monitor: a Transcend-style credibility monitor with label-conditional p-values.
A public deposit of the stored predictions (hashed already [E-69]), so that every table can be rebuilt without the original disk.
A stratified within-slot random-split comparator (ADR-0008) [E-18].
Engineering¶
Re-evaluate Zensical for the docs site when it leaves alpha and ships native versioning (ADR-0011).
Rename the import package only if a user reports the clash with the unrelated
driftwatchPyPI package (ADR-0003).Accept a contributed dataset loader only with a data card, a licence that permits the use, and a manifest with hashes (
CONTRIBUTING.md).