instruments / wink / ADR/0001-boring-classical-comparison-stack.md · raw

ADR 0001 boring classical comparison stack. A research program Kieran shapes and a model (the "seat") writes down. "I" in these files is the seat unless a line is quoted as his. Snapshot of 2026-10-15 from the private repo; local addresses and paths scrubbed, nothing else changed.

0001 — Comparison stack: classical, label-free, CPU-first

Decision (2026-08-21): structural comparison uses graph edit distance (edit script = the loss vector, mechanically), Weisfeiler-Leman refinement and spectral signatures (label-free fingerprints), and permutation/rewiring nulls. networkx/scipy scale; artifacts parse to ~5–50 node graphs. No GNNs, no vector DB, no graph DB at bootstrap.

The GPU is the measurement apparatus, not the math engine: local qwen (frozen, reproducible, cheap at volume) + frontier models as a second instrument family (§30); embeddings only as the baseline to beat (§88). Disk is unconstrained → full generation contexts are archived verbatim (§50 resolved by storage abundance).

Wiring discipline inherited from riverbed: models propose, deterministic code adjudicates. The kernel computes no similarity (ADR-0000).

Reconsider when: artifacts stop fitting in small graphs; GED cost sensitivity tests show hand-set costs dominate results; or an experiment needs behavior (simulation) rather than structure.

Researcher input 2026-08-21: "the thing is sufficiently described... 4090 + effectively infinite disk; slow is acceptable." Bottleneck moved from review to experiment design; design note 0000 status changed to STANDING UNLESS OBJECTED.

back to wink