🔬 Research Log

What executing the work packages taught about the reproduction itself: the paper-fidelity measurements, the modeling and data-protocol choices tried and abandoned, and the results that do not flatter the project. Repo-building findings -- CI, packaging, licensing, formatting, docs tooling -- live in docs/ENGINEERING_LOG.md instead; the two are split by claim, not by work package, so a single WP whose finding straddles both gets one entry in each, cross-linked.

This file is not a register. Where a finding fixed an implementation choice the papers left open, the choice lives in ASSUMPTIONS.md and the number lives there with it; those are cited by id here, never restated. Release-level results live in REPRODUCTION_REPORT.md for the same reason. What lands here is what has no other home: fidelity measurements, negative results, and the reasoning behind a rejection.

Each finding is written once, in its work package's own file, work_packages/wp-<id>.md, between hidden research markers; the phase pages below only include it (--8<-- "work_packages/wp-087d.md:research"), and the roadmap links to the WP file's anchor (#wp-087d). To add one, write the section into the WP file and add its include line to the phase page, then run scripts/lint/audit_wp_links.py --fix: it links every WP-<id> you named and writes the WP page's header (log pages, depends on, needed by, neighbours), and the wp-links hook refuses a commit where either is stale. A package with nothing worth recording here has no section -- and a package whose whole finding is about tooling rather than fidelity has none here at all; see ENGINEERING_LOG.md for it.

Phases 1-5 are backfilled (2026-08-14) and read differently from the rest. Those roadmap rows were transcribed from the blueprint before the work and never rewritten afterwards, so nothing was carried out of them -- their entries here are reconstructed from the commit record, which is the durable artifact, and not from anyone's memory of the work. Phase 6 onward is the reverse: those rows had accumulated the retrospective detail in place, and this file is where it moved to. Phase 0's own findings were all tooling-kind and moved to ENGINEERING_LOG.md whole.

Dates are the day the finding was recorded.

🗂️ Pages

Pages, one per phase:

• Phase 1 — Data pipeline
• Phase 2 — Architecture
• Phase 3 — Assignment and losses
• Phase 4 — MuSGD
• Phase 5 — Lightning training loop
• Phase 6 — Evaluation, v0.1.0
• Phase 7 — Instance segmentation, v0.2.0
• Phase 8 — Oriented detection, v0.3.0
• Phase 9 — Inference and generalization
• Phase 10 — Consolidation
• Phase 12 — Keypoint detection, v0.5.0