From e921b7a899643267c3a20cd3c0b42f996c734ea6 Mon Sep 17 00:00:00 2001 From: Claude Date: Thu, 16 Jul 2026 10:15:15 +0000 Subject: [PATCH] =?UTF-8?q?docs(v3):=20operator=20ruling=20=E2=80=94=20the?= =?UTF-8?q?=20analytic=20Fisher-z=20codec=20IS=20the=20canonical=20cosine/?= =?UTF-8?q?centroid=20replacement;=20helix=20:=20Fisher-2z=20::=20cosine-r?= =?UTF-8?q?eplacement=20:=20Fisher-z?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit EPIPHANIES E-FISHERZ-CANONICAL-COSINE-REPLACEMENT-1 prepended + le-contract §3 canonical-note subsection (append-only): - Ruling 1: the canonical cosine/centroid replacement = the analytic Fisher-z codec (FamilyGamma 8-byte codebook + normalized i8; certified rho>=0.999, E-PALETTE-NNUE-COSINE-GREEN-1). A materialized k x k table is a CACHE of the formula, never the canon: ranking = raw-i8 monotone compares; distance = |delta-i8| x (z_range/254); hydration = affine + tanh. i8 SIMD applies at 4x f32 density with F16/F32-equivalent RANK fidelity (measured; reconstruction-grade beyond certified corpora stays [H]). Boundary: replaces the distance/rank READ, not the semiring COMPOSE. - Ruling 2: helix is to Fisher-2z what the cosine-replacement is to Fisher-z — grounded in the tested identity hyperbolic_depth = 2 x fisher_z (helix fisher_z.rs:110-121) and the residue pipeline's batch_fisher_z; Signed360 = the hemisphere-doubled register. 1z = similarity scale, 2z = hyperbolic location/depth scale; same analytic-codec pattern, zero materialization at both rungs. - Queued follow-up: extract the FamilyGamma contract into lance-graph-contract (BBB-safe); i8 kernels stay ndarray-side. Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_01K3RyLEbuNSHxxB3NTTrGki --- .claude/board/EPIPHANIES.md | 8 ++++++++ .claude/v3/soa_layout/le-contract.md | 19 +++++++++++++++++++ 2 files changed, 27 insertions(+) diff --git a/.claude/board/EPIPHANIES.md b/.claude/board/EPIPHANIES.md index f7bb719b..92936a12 100644 --- a/.claude/board/EPIPHANIES.md +++ b/.claude/board/EPIPHANIES.md @@ -1,3 +1,11 @@ +## 2026-07-16 — E-FISHERZ-CANONICAL-COSINE-REPLACEMENT-1 — THE canonical cosine/centroid replacement = the analytic Fisher-z codec (FamilyGamma, no materialization); helix : Fisher-2z :: cosine-replacement : Fisher-z + +**Status:** OPERATOR RULING (2026-07-16) + FINDING for the grounding identity. No layout changed; the L4 semantics are sharpened, not altered. +**Ruling 1 — the canon:** the workspace's canonical cosine/centroid replacement is the **analytic Fisher-z codec** — `bgz-tensor::fisher_z::{FamilyGamma, FisherZTable}`: cosine → `atanh` → per-family affine (`FamilyGamma` = 8 bytes) → normalized i8; hydrate via `tanh`. Certified ρ≥0.999 (21 roles) + E-PALETTE-NNUE-COSINE-GREEN-1 (ρ_all 0.99971) with the standing rule "for RANKING you never materialize — read the Fisher-z distance directly." Consequences: (a) L4 "cosine-replacement palette256" reads through THIS canon — a materialized k×k table (bgz17's explicit 256², the FisherZTable buffer) is a **CACHE of the formula, an adapter — never the canon**; ranking = raw-i8 monotone compares, distance = |Δi8|·(z_range/254), i8 SIMD (VNNI/`matmul_i8_to_i32`/wasm i8x16) applies at 4× f32 density with F16/F32-equivalent RANK fidelity (measured; reconstruction-grade beyond the certified corpora stays [H]). (b) The "no residue" scoping is intrinsic: cosine-valued quantities are bounded and single-scale — the i4+i2 ladder and Passthrough exist for unbounded heavy-tailed rows, not for this lane. (c) HONEST BOUNDARY: Fisher-z replaces the distance/rank READ, not the semiring COMPOSE (tanh(z_a)·tanh(z_b) ≠ tanh(z_a+z_b)); multi-hop keeps the compose table. +**Ruling 2 — the ladder:** **helix is to Fisher-2z what the cosine-replacement is to Fisher-z.** Grounded in shipped, tested code, not analogy: helix pins `hyperbolic_depth = 2·fisher_z` as an identity at 1e-15 (`crates/helix/src/fisher_z.rs:110-121` — 2·arctanh(s) = ln((1+s)/(1−s)), the hyperbolic/Poincaré depth), and the residue pipeline names the Fisher-Z arctanh as its ONE transcendental (`crates/helix/src/simd.rs:3-4`, `batch_fisher_z`). So: **1z = the similarity scale** (FamilyGamma-normalized i8, the cosine/centroid replacement); **2z = the hyperbolic location/depth scale** (the helix residue register; `Signed360` doubles the hemisphere — bidirectional full-sphere over the same analytic pipeline). Same pattern at both rungs: variance-stabilized transform → normalized small-int register → formula hydration → zero materialization. +**Queued follow-up:** extract the `FamilyGamma` contract (2×f32 + encode/decode formulas, BBB-safe zero-dep) from bgz-tensor into `lance-graph-contract` so consumers (L4 ClassViews, ndarray SpoDistanceMatrices analytic planes, thinking-engine lenses, stockfish-rs) speak the canon without the bgz-tensor dep; i8 SIMD kernels stay ndarray-side per the hardware/thinking split. [H] until shipped. +**Cross-ref:** E-PALETTE-NNUE-COSINE-GREEN-1 (the certification), E-PALETTE-RESIDUAL-LADDER-1 (the ladder this lane does NOT need), E-H268-PROBE-WAVE-1-RESULTS (survivors context), le-contract §3 L4 row + the new canonical-note subsection, `encoding-ecosystem.md` (codec map), TD-BGZ-TENSOR-PRE-LANE-REVIEW (the extraction follow-up feeds its outcome). ndarray §8 mirror folds into the next ndarray doc pass. + ## 2026-07-16 — E-H268-PROBE-WAVE-1-RESULTS — three H.268 conjectures measured: WH-tile NEUTRAL/not-transferring, signature-checksum PASS (depth-2 blind spot), walk-spectrum KILL of §10(g)'s "decorrelated" half **Status:** FINDING (three probes run against shipped code, deterministic seeds, Opus-adjudicated; grades E-WH-TWO-SIDES-SIG-CHECKSUM-1 legs 1+2 and §10(g); plan `h268-probe-wave-v1.md`). No canon changed; one claim-leg CLOSED NOT-TRANSFERRING, one KILLED, one PASSes with a bounded caveat. diff --git a/.claude/v3/soa_layout/le-contract.md b/.claude/v3/soa_layout/le-contract.md index daff6aa5..53a114d5 100644 --- a/.claude/v3/soa_layout/le-contract.md +++ b/.claude/v3/soa_layout/le-contract.md @@ -154,6 +154,25 @@ facet payloads, lens-selected per class. Consequences: `edges[16B]` block in the CANON node is untouched by this (it is one-byte SLOT refs, not packed mantissas). +### The canonical cosine/centroid replacement is ANALYTIC (operator, 2026-07-16) + +L4's "cosine-replacement palette256" reads through the **analytic Fisher-z +codec** (`bgz-tensor::fisher_z::{FamilyGamma, FisherZTable}` — cosine → +`atanh` → 8-byte per-family affine → normalized i8; hydrate via `tanh`; +certified ρ≥0.999, E-PALETTE-NNUE-COSINE-GREEN-1). **A materialized k×k +table is a CACHE of the formula, never the canon**: ranking compares raw +i8 (monotone), distance is `|Δi8|·(z_range/254)`, and hydration is one +affine + `tanh` — so an L4 ClassView may declare an analytic codebook +(the 8-byte gamma) with zero table materialization. Boundary: this +replaces the distance/rank READ; the semiring COMPOSE keeps its table +(z-addition does not compose cosines). The location-side sibling, one +rung up: **helix is to Fisher-2z what the cosine-replacement is to +Fisher-z** — helix pins `hyperbolic_depth = 2·fisher_z` as a tested +identity (`crates/helix/src/fisher_z.rs:110-121`) and runs its residue +pipeline on that doubled scale (`Signed360` = the hemisphere-doubled +register). Canonical ruling text: EPIPHANIES +`E-FISHERZ-CANONICAL-COSINE-REPLACEMENT-1`. + ### CAM_PQ grounding (digital vs analog) CAM_PQ's codebook is the **DeepNSM 4096-word English-native-speaker