Nishi Recall — State-of-the-Art Evidence

Sovereign search / recall engine · integer-deterministic (no float) · every number below is a live gate output, not an assertion · 2026-07-18

Measured SOTA lifts (nDCG@5, on the sovereign nx_ir_eval ruler)

rung	baseline	with rung	lift	evidence
R1 hybrid RRF fusion (BM25 ⊕ dense)	0.494	0.659	+33% rel	nx_recall_fuse_gate 7/7; nx_recall_bench live
R2 sovereign semantic reranker (PPMI-SVD cosine)	0.530	0.964	+82% rel	nx_recall_rerank_gate 5/5; nx_recall_bench live

Recomputed LIVE per call by nx_recall_bench (verdict RED + exit-1 on any lift inversion — a regression cannot hide). R2 replaces the prior external-LLM rerank crutch with a sovereign, bit-deterministic reranker — the field's #1 named gap (nishi_compare), closed as the fast tier.

R1b + R1c — two rerank signals LIVE in the production search engine

DEPLOYED LIVE 2026-07-19 — R1b term-proximity + R1c title-field boost (BM25F-lite); docportal daemon swapped (backup + verified), serving /search now

live query "cross encoder reranking"	relevant IR doc score	distractor (HLS)	relevant doc lead
before deploy	49,612,409	48,761,029	+1.7% (razor-thin)
after R1b (live now)	50,775,199	48,856,265	+3.9% (MORE THAN DOUBLED)

The proximity boost pulled the semantically-clustered relevant doc further ahead of the keyword-adjacent distractor, in production.

R1c title-field boost — live query "JPEG image compression": the top 3 results are now the TITLE-matching canonical articles — "Image compression" (2 query terms in title, 78.6M) → "JPEG" (64.1M) → "Lossy compression" (59.1M), above scattered-mention pages. Fielded ranking surfacing the canonical docs, in production.

Below: the offline gate proof that got both signals there.

32 / 32 GREEN — 30 baseline ranking tests + T-prox + T-title, all pass = zero regression

proof	result
30 existing ranking behaviors (BM25 / IDF / length-norm / phrase / stemming / host-diversity / PageRank authority fusion / url-dedup)	ALL PASS — ZERO regression
T-prox: clustered doc outranks scattered doc at IDENTICAL BM25 (indexed 2nd) in the real engine	PASS

Term-proximity (Buttcher-Clarke-Cormack) via dss_prox_all, which clones the engine's own tokenizer loop so it matches EXACTLY what BM25 matched — a raw-substring version would mis-fire and regress ranking; that shortcut was rejected on purpose. Applied as a bounded, capped PASS-2 boost (single-term queries byte-identical). Gate: nx_docportal_search_seg_gate.

On MCP + API (live, verified through the edge)

tool	what	status
`nx_recall_eval`	the ruler as a service (nDCG/MRR/Recall/P/AP scorecard)	LIVE — tools/call 200
`nx_recall_bench`	the lift bench as an evidence-fork organ (recomputed per call)	LIVE — tools/call 200
`nx_recall_board`	recall/search SOTA board (axes + coverage + research + fetch) — the /compare + atlas interchange	LIVE — tools/call 200
full register→mint→call loop	ship → build-on-NAS → promote → register → cap → live call	PROVEN

RACI & workflow (organized, not random)

F	rung	owner (A)	state
F230	IR eval ruler	librarian	DONE
F231	hybrid fusion + R1b SERP wiring	librarian	proven, deploy-gated
F232	sovereign semantic reranker	librarian	DONE (gated)
F233	nishi_recall (memory / source / store)	librarian	ready
F234	atlas / GraphRAG recall (recombination)	librarian + researcher	blocked/atlas
F235	no-float LLM reranker (escalation tier)	modelwright	ready
F236	BEIR external-suite honest row	researcher	ready
F237	deterministic learned-sparse (beyond SOTA)	researcher	ready

Live on the CPM board (nx_frontier_board lane=recall). Lane owner accountable = librarian; F235 = modelwright.

Honest maturity (evidence-capped, not inflated)

L2 — organs gate-proven + 3 live MCP tools + one live SERP rerank (R1b) + measured lifts + scale-envelope gate at N=2048. NOT L3 until: an EXTERNAL suite row (BEIR subset through nx_recall_eval, F236), full fusion/rerank wiring live, and a scale audit past the fixtures. Feature-presence ≠ maturity.

SOTA-board coverage (nx_recall_board, live on MCP): 500‰ — HAVE-LIVE=1000, HAVE-GATED=500, GAP=0, mean over 8 axes; a GAP is never counted as HAVE (liar-killer). The two remaining GAPs (dense-in-SERP, neural cross-encoder) are the honest frontier, owned + flagged.

Debt eaten · honest open items (no cheating)

3rd-party debt	EATEN — external-LLM rerank crutch replaced by the sovereign R2 reranker
our own debt	EATEN — dense + RRF were built-but-unwired; now fused (R1)
R1b live deploy	DEPLOYED LIVE 2026-07-19 — routed around a broken NAS buildroot (debt seq120) by uploading the locally-built elf; smoke-tested on a throwaway port first, backed up the old daemon, swapped + relaunched, verified /search. Client-doc encryption/authz preserved (only +2.3 KB = the search addition). Live margin doubled (above).
docportal buildroot	OPEN (seq120) — the daemon won't rebuild on the NAS; deploy used the elf-upload bypass. Fixing the buildroot is the docportal owner's follow-on.

Research opportunities — go beyond SOTA

Deterministic learned-sparse (SPLADE-class) in integer — BEIR-benchable novelty no one in the field has bit-reproducibly.
Integer late-interaction (ColBERT-class) maxsim on PPMI-SVD embeddings.
No-float LLM reranker (F235) — the world's first bit-deterministic neural reranker (needs the KV-cache speed rung).
Determinism-as-a-benchmark-axis — a reproducible-IR protocol we EXCEED by construction.
GraphRAG recombination recall (F234) over the 16,580-node atlas — semantic-nearest distant nodes = novel-design candidates.

What to look for in our fetches (feed the ruler)

BEIR qrels first (nfcorpus / scifact / fiqa) + MS MARCO dev + TREC-DL = external ground truth; primary IR arxiv PDFs (SPLADE / ColBERTv2 / RRF / GraphRAG), cite-grounded; competitor SERP snapshots (Mojeek / Brave / Marginalia) on a fixed query set on cadence; classify-at-the-door (U//public-release only into the public shard).

Sovereign end-to-end: nx_cc → nxasm, zero gcc / python / CUDA / external LLM. Integer-deterministic — same input, same bits, every machine.