nx_debtcluster.nx

nx_debtcluster.nx -- THE LEDGER STOPS BEING A LIST AND BECOMES A RECOMMENDATION. license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0 ─── THE DEFECT THIS EXISTS TO KILL ──────────────────────────────────────────────────────────────────── MEASURED 2026-07-31: the debt plane holds 2068 rows, 905 open, 281 open at sev>=8, and it is FLAT. Every seat reads it top-down by severity and picks a row. But the rows are not independent -- in the first FORTY sev>=8 rows alone there are at least five families, each one incident wearing many row numbers: mgmt_api availability seq698 crash-loop · seq710 hang · the live sev-9 :18098 outage resource exhaustion seq883 23.6GiB alloc-without-free · seq774 20,440 respawns · seq783 · seq727 · seq975 dup-source seq207 · seq701 · seq945 · seq965 · seq974 (965/974 report GREEN divergent=0 WHILE THE PARENTS STAY OPEN) law-warden magic nums seq274 L001 (_hdl_build, 2348) · seq315 L006 (core, 3476) -- same rule, same detector, two corpora, filed as two unrelated problems gx-body seq804·805·808·819·820·824·826·828·900 = NINE rows, ONE skull-architecture arc A seat that fixes the top row by severity fixes ONE MEMBER of a family and leaves the family. The ledger is sorted by how loud a row is, never by how much work a single act would close. ★LAW: A LEDGER SORTED BY SEVERITY ANSWERS "WHAT SCREAMS LOUDEST", NEVER "WHAT SHOULD I BUILD". ─── WHAT SOTA SAYS (researched 2026-07-31, and what we deliberately DID NOT copy) ───────────────────── 1. NORMALIZE BEFORE YOU EMBED. LLM-augmented ticket clustering (FLAIRS 2026, ABCD) beats raw-text embedding on silhouette AND human-rated coherence purely by extracting a canonical statement first. GPTrace (ICSE'26, arXiv 2512.01609) does the same for crash dedup: normalize -> embed -> HDBSCAN, scored on silhouette + DBCV + P/R/F1, beating CrashWalk and Igor. 2. DENSITY-BASED, NOT PARTITIONAL. HDBSCAN is chosen in that literature precisely because it DECLINES to cluster outliers. A debt ledger is mostly singletons; a partitional method (k-means, plain agglomerative) has nowhere to put them and pollutes every cluster with noise. We implement the principle, not the library: a core-density test with an EXPLICIT noise class that is always reported. 3. TOPOLOGY BEATS TEXT SIMILARITY FOR CAUSALITY. TopoEvo (arXiv 2605.15611) is unambiguous -- correlation is constrained by the ACTUAL dependency graph, and topology-agnostic baselines manufacture spurious links by treating all relations as equivalent. TWO ROWS BOTH SAYING "LEAK" IN UNCONNECTED ORGANS ARE TWO PROBLEMS. This is the ic_ident_gate below, and it is the single most important line in the organ. 4. RANK BY RISK-REDUCTION OVER EFFORT, and cap the recommendation list at 5-8 (postmortem practice). 5. ★DO NOT BUILD THE AGENTIC RCA TIER. ITBench-AA (Artificial Analysis + IBM Research, launched 2026-05-27, 59 live Kubernetes SRE tasks): EVERY frontier model scores below 50% -- Claude Opus 4.7 47%, GPT-5.5 46%; IBM's own ITBench reports best-agent 13.8% SRE / 25.2% CISO / 0% FinOps. Meanwhile the plain correlate-and-rank layer delivers a reliable 80-85% noise reduction in production. So we build the deterministic layer that demonstrably works and refuse the one that demonstrably does not. 6. GRADE ORDINALLY. Arvo's AI Investigation Capability Ladder (May 2026, upd. July 2026): L0 manual · L1 correlation · L2 summarized timeline · L3 single-shot diagnosis · L4 agentic · L5 closed-loop. THIS ORGAN IS L2, EVIDENCE-CAPPED, AND SAYS SO IN ITS OWN OUTPUT. Claiming L4 would be the banked