docs(adr): ADR 0113 representative density graph (semisupervised-v3) + validation plan #1254
Loading…
Reference in a new issue
No description provided.
Delete branch "docs/adr-0113-representative-graph"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Summary
Draft ADR 0113 (
proposed) + validation plan for the representative density graph optimization of face clustering.Decision (V1–V7): new single-pass method
semisupervised-v3: threshold graph over U ∪ R (unlabeled pool + ADR 0108 representatives as labeled identity-evidence nodes), pairwise computation restricted to U×U + U×R — no L-blocks. Steady state: (L+U)² → U² + U·R. Group formation on the U×U subgraph only; gates generalized (not forked) to read rep adjacency; contract version 2, fail-closed matrix on worker + Core; v1/v2 byte-identical and operator-selectable (rollback = stay on v2). Sub-decision: fingerprint-keyed representative cache (method-independent, benefits v2 immediately). Honest limits recorded: no bulk-import win (ANN = follow-up), no feed-level incrementality (future work), accepted rep-induced edge flips near threshold.Validation plan G3 filled with measured prototype numbers: steady state 111× faster (3.35s→0.03s), computed pairs 228× fewer (78.1M→342k), peak RSS 733→8.5 MiB, max HAC component 102→10; mid 21.7×; bulk 5.7× faster (no regression). Identity sanity: 100% own-rep adjacency, 0% foreign-rep adjacency.
Status: ADR stays
proposeduntil G1 (owner-library CV) and G4 (deploy-host run) pass.Test plan
python3 -m json.toolfeat/representative-density-graph-prototype(additive module + benchmark, all suites green)