nishi code wiki / SOTA by generation · siblings: SOTA rung plan · method brief · the family tree

SOTA by generation

The ecosystem is a family tree: gen 0 = the god primitives; each generation builds on the ones above. This page scores every generation against what state-of-the-art means for that stratum, and clusters the work accordingly. Method per July-2026 research (core–periphery/DSM economics, CNCF maturity dimensions, topology-constrained clustering) — sourced in the companion research brief.

⚠ HAND-AUTHORED SNAPSHOT — THE FIGURES HERE ARE DATED, NOT LIVE
This sentence used to read “every number below was measured by a named instrument, not asserted”. It was true the day it was written and then became the most misleading claim on the page, because a dated snapshot that says it was measured is indistinguishable from a live surface to any reader who does not re-run the instrument. Figures are as of 2026-08-01 unless a run_epoch says otherwise; 2026-08-07 corrections are marked CORRECTED inline.

For a live answer, run the instrument rather than reading this page: nx_sota_status (PROVEN / EXCEED / frontier gaps) · nx_gensota (per-generation attach) · nx_ecosystem_maturity_rollup (maturity; it now stamps autonomy_age_s so a stale input reads as stale) · /sota, which nx_sota_page REGENERATES from live instruments — unlike this page, which occupies the more discoverable URL.

Drift measured 2026-08-07: PROVEN 0/41 here vs 41/41 live · capabilities 144 vs 160 · frontier gaps 280 vs 335 · gen-3 attach 235‰ vs 239‰. The attest_keys.conf blocker described below as open was resolved 2026-08-01 19:10 — hours after authoring.
A layered system is only as SOTA as its weakest generation.
Core–periphery research is unambiguous: defect cost concentrates in the tightly-coupled core — a small number of components generate most of the maintenance cost, and a root defect is a fleet event. We measured that here on 2026-08-01: one type defect in the compiler made ~10% of organs unbuildable at once. So the headline is the MIN across generations, never the mean — and when this was authored the binding constraints were gen 0 evidence (PROVEN 0/41, then believed structurally blocked on one operator act — CORRECTED 2026-08-07: PROVEN is 41/41 and has been since 2026-08-01, so gen 0 is no longer a constraint at all) and the gen 3 adoption floor — measured at 235‰ attach, the ecosystem MIN (the one-way import path that capped it was root-fixed 2026-08-01; the harvest of the capability it unblocked has not yet been collected).
Generations
8
16,222 organs · 42,088 import edges (atlas tree)
Independently proven
41/41
nx_sota_status run_epoch 1786127231 (2026-08-07) · CORRECTED: this box read 0/41 from 2026-08-01 until 2026-08-07
Gates a ruler can read
106/413
256‰ · the non-vacuity ceiling for every generation
Orphans (the noise class)
3,851
+ 2,227 tested-only — validated, unadopted

The generations, each scored on its own bar

Composition numbers from the atlas family tree (living-tree / total per generation). The SOTA bar differs by stratum — scoring a root like a product, or a product like a root, is a category error.

RED
gen 0 · the god primitives — 118 living / 295 total

What it is: organs that import nothing; all 16,222 descend from here. Exemplars by descendant reach: nx_tier (2,864), nx_syscalls_x86_64, nxasm_v2, nx_types, nx_tls13, nx_f32_hw, nx_axioms.

The SOTA bar here: independently proven correctness — external oracles (NIST/RFC test vectors), negative controls that can fail, mutation-proof gates. Nothing less, because a gen-0 defect propagates to everything.

Where we stand (CORRECTED 2026-08-07): PROVEN 41/41 — every domain, measured live by nx_sota_status at run_epoch 1786127231. knowledge/attest_keys.conf exists, and has since 2026-08-01 19:10 — hours AFTER this page was authored. This paragraph then advertised the blocker as open for six days.

The registry's own header carries the reasoning and is worth reading rather than paraphrasing: role=machine IS registered, because for EV_CLASS_ORACLE independence comes from the REFERENCE, not the signer — a machine signature over “this gate ran and matched RFC 9807 byte-exact” certifies nothing about truth; RFC 9807 certifies that, and any reader can re-run the gate. The signature buys tamper-evidence and attribution only. role=human remains deliberately absent: an agent registering a human key would forge the very leg that exists so an automated actor cannot certify its own work.

What the staleness cost, measured: on 2026-08-07 an agent read this page, took “PROVEN 0/41, blocked on one operator signature” at face value, and reported it as the estate's single highest-leverage act — ranking a six-day-resolved item first. A HAND-AUTHORED SNAPSHOT THAT SAYS ITS NUMBERS WERE MEASURED IS INDISTINGUISHABLE FROM A LIVE ONE TO ANY READER WHO DOES NOT RE-RUN THE INSTRUMENT. A date line is not protection: the reader has no way to know WHICH numbers moved.
Evidence for the propagation claim, measured here: the 2026-08-01 compiler regression (a scalar static typed as its own slot address) made ~10% of organs unbuildable from one root defect. The promoted compiler could not compile itself — the promote had skipped the self-host stage.

The work: NONE — C1 is closed. This read “one operator signature, then the already-derived oracle rows convert” for six days after the signature had been made and the rows had converted. A WORK ITEM THAT OUTLIVES ITS COMPLETION SPENDS SOMEONE ELSE'S ATTENTION EVERY TIME IT IS READ.

AMBER
gen 1 · first composition — 1,275 living / 4,130 total

What it is: the first layer built on the roots. Contains the single most load-bearing organ in the ecosystem: nx_syscalls with 13,268 transitive importers; plus the f32 plane, nx_gguf, nx_image, nx_tensor.

The SOTA bar here: non-vacuous, ruler-readable gate coverage on the hubs, with mutation testing as the exit criterion — a surviving mutant is a gate that cannot fail. A silent threshold shift in this generation is a fleet event.

Where we stand: gates a rollup can read = 106 of 413 (256‰); the first mutation sweep scored 16 gates and found 7 survivors (3 fixed); the integer-truncation class (a bar silently divided down before compare) was found at 9 sites across 7 live gates, latent. The hubs are correct-by-run but under-proven-by-instrument.

The work: cluster C2.

AMBER
gen 2 · the bulk of the body — 1,809 living / 10,053 total (62% of the tree)

What it is: where the ecosystem actually lives — stores (nx_seg_store), trust (nx_x509_trust_store), runtime, publishing, capability registry. The widest generation by far.

The SOTA bar here: the golden path — extending a plane costs one data row, never bespoke code (the CNCF/Gartner maturity dimensions land hardest on this stratum). Duplication is measured, then collapsed.

Where we stand: the deploy plane proved the pattern — 27 near-duplicate deploy/rollback implementations collapsed into one row-driven path, coverage 333‰→666‰ doubled by rows, not code. But that was ONE plane; the same duplication class is unmeasured across the rest of the widest generation, and the debt-cluster board's top families (nx_mgmt_api, nx_hostctl, nx_debt…) all live here.

The work: cluster C3.

RED
gen 3 · domain libraries — 294 living / 1,468 total

What it is: the reusable domain layer — nx_str, nx_strconv, nx_log, nx_search_inverted, the sketch plane.

The SOTA bar here: adoption. A library's value is its importer count; discoverability is the capability, and import cost is measured before blaming discoverability.

Where we stand: the structural ceiling is FIXED, 2026-08-01: the resolver's subroot list lacked _hdl_build; it is now tried LAST at every walk-up level (additive — only previously-failing imports can change), shipped inside a compiler promoted through the canary gate to a byte-identical self-host fixpoint (sha256 e98e0f55… on both generations). Proof battery: the debt's own acceptance case (import nx_ppmi_lib.nx from runtime/) BUILDS; a nowhere-existent import still FAILS (not fail-open); sev-8 1785608999 EATEN. What remains is the harvest: the measured attach floor here is 235‰ — the ecosystem MIN — and the 2,227 tested-only organs await adopt-or-retire triage now that nothing structural blocks them.

A one-way import path is a silent adoption ceiling — it breaks nothing that already works; it only prevents anything new. Fixed by construction: the new subroot is searched last, so no import that resolved yesterday can resolve differently today.

The work: cluster C4 — and it is now a RANKED QUEUE, not a number: nx_gensota worklist 3 emits this generation's unattached organs ordered by measured importers. Its first run found the highest-leverage item in the ecosystem: nx_gate_gn, imported by 437 gates and itself ungated — the integer printer under the entire verification fleet. Closed 2026-08-01 (nx_gate_gn_gate 7/7 GREEN, proven by redirecting fd 1 and reading the bytes back), together with nx_str (109 importers, 11/11). Both then dropped off the worklist on the next graph rebuild and gen-3 attach moved 235→236‰, ratcheting the bar.

A self-test inside a module's own main() is documentation that happens to compile. nx_str had one — and expand_imports strips main from every non-root file, so it ran only if someone built that module as an entry point, and no ruler could ever read its result.
AMBER
gen 4 · served composition — 37 living / 225 total

What it is: the daemons and served surfaces composed from the libraries: nx_docportal_search_serve, nx_search_inverted_persist, nx_frame_codec, nx_robots.

The SOTA bar here: readiness, not liveness — a port that accepts and never serves must read RED; remediation targets the actual holder of the resource and escalates on no-op.

Where we stand: self-healing was proven end-to-end on a real fault with no human in the path (beat fired → gate named 3 supervisor roots → sentinel reconciled 3→1). But the same week, a runaway daemon took 16.5GB RSS in 10s and starved the whole host before quarantine — admission control exists (memfloor) and is not yet universal on this stratum.

The work: folded into C2 (instrument coverage) + C5 hygiene.

AMBER
gen 5 · 5 living / 45 total   —   gen 6 · 1 / 4   —   gen 7 · 1 / 2   (the products spike)

What it is: the top of the tree thins to a single lineage — gen 5 (nx_survey_sentiment, nx_golive, nx_tpl_site…) → gen 6 (nx_survey_serve) → gen 7 (nx_survey_daemon). One product line carries the ecosystem's entire height.

The SOTA bar here: the field — product benchmarks against external competitors (the 41 compare domains), not internal gates.

Where we stand: the thinness IS the finding. The ecosystem holds 160 unique EXCEED capabilities and 335 declared frontier gaps (nx_sota_status, run_epoch 1786127231, 2026-08-07 — this read 144 / 280 when authored 2026-08-01), yet almost every product flattens at gen 2–3 instead of composing upward. The survey spike proves tall composition works; nothing else uses it.

The work: clusters C6 + C7.

The three quarantine classes (not generations, and never mixed in)

classorgansreading
validation6,604tests/gates — paperwork attached to its subjects. Healthy; its defect class is unreadability (see C2).
tested-only2,227validated, never adopted — the adoption trap, quantified. Worklist for C4.
orphans3,851one-offs nothing builds on — the honest noise class. July-2026 clustering practice is explicit: decline to cluster outliers rather than force them. Janitor worklist (C5), never force-fit into a family.

The work order — 7 clusters, generation-ordered

Capped at 7 per postmortem-practice guidance (5–8 recommendations). Ordered by the weakest-path rule: evidence on a leaf buys nothing until its whole chain lifts, so lower generations come first. Blast radius is the measured descendant reach; effort tiers are assessed, and marked so.

#clusterstratumthe one actblast radius (measured)status
1C1 · Root attestationgen 0DONE 2026-08-01. attest_keys.conf exists; role=machine is registered (for oracle rows independence comes from the REFERENCE, not the signer), role=human stays deliberately absent. The oracle rows route.all 16,222 organs inherit a proven floorRESOLVED — PROVEN 41/41 (this row read “operator-gated, not agent-runnable” for six days after it was done)
2C2 · Hub non-vacuitygen 1–2 + validationMake the remaining 307 gates ruler-readable; mutation-test the hubs; sweep the 9 truncation sites (cross-multiply, never divide).413 gates; every rollup that reads themagent-runnable now
3C3 · Golden-path the bulkgen 2Repeat the deploy-plane collapse (27 copies → 1 row-driven path) on the next measured duplication family from the debt-cluster board.62% of the tree lives in this stratumagent-runnable; one plane at a time
4C4 · Lift the adoption ceilinggen 3DONE 2026-08-01 (first half): resolver fixed, compiler promoted to a byte-identical self-host fixpoint, probes prove both directions, sev-8 1785608999 EATEN. Remaining: adopt-or-retire triage of the 2,227 tested-only organs, tracked by the nx_gensota attach floor (235‰ ratchet).every future consumer of every gen-3 libraryceiling fixed; harvest agent-runnable now
5C5 · Janitor the orphansnoise class3,851 one-offs: adopt, attach as validation, or retire — each an explicit decision, never a forced cluster.24% of the tree exits limboagent-runnable, long-tail beat
6C6 · Convert unblocked evidencedomainsWork the measured queue where evidence raises delivered capability TODAY: automation · containers · doctor · genealogy · graphviz · supervisor · videoops · workermesh; fix the 3 RED (typography, librarian, office — build or retract, never unwire).8 unblocked domains + 3 RED (nx_capgraph_derive roots, 2026-08-01)agent-runnable now
7C7 · Build heightgen 5–7Recombine existing gen-2/3 capability into taller product lineages — the survey spike is the existence proof; the 160 EXCEED axes are the raw material.the whole moat: 160 unique capabilities (144 when authored)agent-runnable after C4 (adoption ceiling caps it)

The instrument — nx_gensota, live since 2026-08-01

The rollup organ this page declared as future work now EXISTS: nx_gensota (promoted sha256 c5346740…, gate 5/5 GREEN incl. a neg-control that forces RED and an absent-store INSTRUMENT-BLIND tooth; bar in knowledge/gensota_bars.conf, ratcheted at today's measured MIN; hourly beat in the self-applying cron registry). Live artifact: /code/gensota.txt. First measurement, whole graph (17,966 modules / 48,002 edges, partition SUM-OK):

gentotalvalidationorgansattachedattach‰
034019321144448
13,8938273,0662,441796
211,3095,4215,8881,838312
32,0347341,300306235 — the MIN
239 live 2026-08-07
430114715461396
576453113419
612932666
71100na
The instrument independently confirmed the hand assessment: the measured weakest generation is gen 3 (attach 235‰) — the same stratum this page had assessed RED for adoption. Verdicts cross-multiply (att×1000 vs bar×organs), never divide-then-compare.

Method honesty

Measured: every composition count (atlas family tree), evidence classes and EXCEED/gap counts (nx_sota_status run_epoch 1785613017), the unblocked queue (nx_capgraph_derive roots), gate readability 106/413, deploy coverage 10/15, the graph itself (nx_eco_graph_build: 17,966 nodes / 48,002 edges on the fuller module population), and now the per-generation validation-attachment rollup above (nx_gensota, gate-proven). Still assessed, and declared so: the RED/AMBER band colors synthesize the measured rollup with named incident evidence (one axis is now derived; the multi-axis color is not yet), and the effort tiers in the work order.

Two instruments, two populations, on purpose: the atlas tree counts 16,222 organs (living-tree snapshot); nx_eco_graph_build counts 17,966 nodes (full module scan, same day). Neither is wrong; a headline that mixed them would be. Each number above names its instrument.