From e784032d9f0319d18a0dfc62755c255c13ea848c Mon Sep 17 00:00:00 2001 From: Claude Date: Thu, 16 Jul 2026 09:10:10 +0000 Subject: [PATCH 1/2] =?UTF-8?q?docs:=20probe-wave-1=20verdicts=20=E2=80=94?= =?UTF-8?q?=20=C2=A710(g)=20corrected=20(KILL=20of=20the=20decorrelation?= =?UTF-8?q?=20half),=20=C2=A710=20RESULT=20lines,=20=C2=A75=20results=20su?= =?UTF-8?q?b-table?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Companion to lance-graph E-H268-PROBE-WAVE-1-RESULTS (probes live there; Opus-adjudicated per h268-probe-wave-v1.md): - PROBE-WH-MAG NEUTRAL, bare-tile leg CLOSED not-transferring (B/A 0.929/1.411/1.869); WHT16 spreads outlier energy tile-wide; the row-level win needs passthrough+centroid, which the probe isolates away; shipped row codec untouched; PROBE-WH-MAG-2 deferred. - PROBE-SIG-CHECKSUM PASS with the depth-2 parallel-chord blind-spot bound (null space exceeds tree-like equivalence). - PROBE-WALK-SPECTRUM KILL of §10(g)'s "decorrelated by construction" (lattice |R| 0.875 vs PRNG 0.0205 = 42.7x; C(13)=-15 sidelobe; coprimality != decorrelation); known-structure half CONFIRMED (R(17m)=1-8m/N); D-QUANTGATE / bijectivity unaffected. Knowledge-doc suite green (117/117). Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_01K3RyLEbuNSHxxB3NTTrGki --- .claude/blackboard.md | 22 +++++++ .../pr-x12-h268-morton-wgpu-synergies.md | 66 ++++++++++++++++--- 2 files changed, 79 insertions(+), 9 deletions(-) diff --git a/.claude/blackboard.md b/.claude/blackboard.md index 044eb289..9bc475a4 100644 --- a/.claude/blackboard.md +++ b/.claude/blackboard.md @@ -784,3 +784,25 @@ default is x86-64-v3 (avx2) so ndarray_simd activates on avx512 builds only. digest's null space, the formal "which detours leave no comma"; PROBE-SIG-CHECKSUM on the Pillar-11 harnesses). Neither adds stored tile fields. + +--- + +## 2026-07-16 (4) — h268-probe-wave-v1 RESULTS (adjudicated) + +- **PROBE-WH-MAG → NEUTRAL; bare-tile leg CLOSED NOT-TRANSFERRING.** + B/A 0.929/1.411/1.869 — WHT₁₆ spreads outlier energy tile-wide, + inflating the per-cell quantization floor; the row-level win needs the + passthrough escape + centroid residual the probe deliberately omits. + Shipped row codec untouched. PROBE-WH-MAG-2 named, deferred. +- **PROBE-SIG-CHECKSUM → PASS** with the depth-2 bound: parallel-chord + interior displacement is EXACTLY signature-invisible — null space + exceeds tree-like equivalence; mitigate via depth 3 or paired digest. +- **PROBE-WALK-SPECTRUM → KILL** of §10(g)'s "decorrelated by + construction" (walk lattice |R| 0.875 vs PRNG 0.0205 = 42.7×; + C(13)=−15 sidelobe, coprimality ≠ decorrelation); "known period-17 + structure" half CONFIRMED (R(17m)=1−8m/N). D-QUANTGATE unaffected. +- Doc updates: §10(g) corrected, §10 forward-synthesis RESULT lines, + §5 results sub-table. Canonical verdicts: lance-graph + E-H268-PROBE-WAVE-1-RESULTS + plan h268-probe-wave-v1.md Results. + Probes: bgz-tensor probe_wh_mag / jc sig_checksum / helix + walk_spectrum (all suites green). diff --git a/.claude/knowledge/pr-x12-h268-morton-wgpu-synergies.md b/.claude/knowledge/pr-x12-h268-morton-wgpu-synergies.md index 8131f690..823ee567 100644 --- a/.claude/knowledge/pr-x12-h268-morton-wgpu-synergies.md +++ b/.claude/knowledge/pr-x12-h268-morton-wgpu-synergies.md @@ -115,6 +115,15 @@ any float-path GPU bit-exactness claim. | Plan E bench | bits/Gaussian on Mip-NeRF 360 | ≤4 bits | R-10 re-derived; web-streaming claim withdrawn | | a2ui N2 | wgpu `webgl` feature + texture upload, wasm32-tested | render parity headless vs browser | GPU raster tier deferred; CPU raster only | +**Run this wave — h268-probe-wave-v1 (2026-07-16, reviewer-adjudicated; +probes live in lance-graph `bgz-tensor`/`jc`/`helix`):** + +| Probe | Verdict | Numbers | Consequence | +|---|---|---|---| +| PROBE-WH-MAG | NEUTRAL / bare-tile leg NOT-TRANSFERRING | B/A 0.929 / 1.411 / 1.869 (grad+spike / heavy-tail / noise) | §10 [H] leg: WH-on-bare-tiles does not transfer; shipped row codec (lance-graph E-PALETTE-RESIDUAL-LADDER-1) untouched; PROBE-WH-MAG-2 deferred | +| PROBE-SIG-CHECKSUM | PASS (depth-2 blind-spot bound) | replay 0.0; tree-like 0.0<1e-9; non-tree 11.31>0.05 | §10 [S] leg confirmed; digest ≠ catch-all — parallel-chord displacement is signature-invisible at depth 2 | +| PROBE-WALK-SPECTRUM | KILL of §10(g) "decorrelated" half | period-17 \|R\| 0.998/0.996/0.994; lattice max 0.875 vs PRNG 0.0205 = 42.7× | §10(g) corrected: known structure ≠ low correlation; D-QUANTGATE / bijectivity unaffected | + ## 7. Comma closure — the replayable irrational (constants correction folded in) The Pythagorean comma is the residue of a stack of pure fifths that never @@ -351,15 +360,30 @@ bytes actually stored are magnitudes. Same object, two consumers. (the integer-only GPU caveat) unless a deterministic-backend probe is added. Within that scope, counterfactual replay stores **zero** exploration state. -- **(g) Anti-confabulation = anti-moiré in concept space [H, probe-gated]** - — a coprime probe schedule is decorrelated from the palette lattice by - construction, and a *known* period-17 dependence structure is plausibly - friendlier to `I-NOISE-FLOOR-JIRAK`'s weak-dependence analysis than an - *unknown* PRNG correlation structure. The "friendlier" half is an - **unverified inference**: the period-17 permutation self-test proves - bijectivity, NOT decorrelation. Promotion to FINDING requires a measured - dependence probe — the walk's correlation spectrum against the palette - lattice vs a PRNG baseline, judged under Jirak weak-dependence rates. +- **(g) Anti-confabulation — the "decorrelated by construction" half + KILLED 2026-07-16 (PROBE-WALK-SPECTRUM); the "known structure" half + CONFIRMED.** The coprime walk's period-permutation self-test proves + bijectivity and gives a KNOWN, replayable, closed-form dependence + structure (period-17 harmonics measured |R| = 0.998/0.996/0.994, + matching R(17m)=1−8m/N) — real value for auditability and conformance. + But it is **NOT decorrelated from the palette lattice**: the parity of + a full-period-17 permutation is an exact period-17 ±1 waveform, so its + autocorrelation at lag τ is governed by C(τ mod 17) at EVERY lag, and + the lattice periods land on large sidelobes — measured walk |R| max + 0.875 at τ=64 (64 mod 17 = 13, C(13)=C(4)=−15, −15/17 ≈ −0.882) vs a + PRNG baseline's 0.0205, i.e. 42.7× (≫ the 3× kill bar). + `gcd(17,τ)=1` guarantees a full-period permutation, NOT low + autocorrelation. So the "anti-moiré in concept space / friendlier to + `I-NOISE-FLOOR-JIRAK`" inference is **struck** — strong periodic + autocorrelation is if anything worse for weak-dependence rates than a + PRNG's ≈−1/N floor. Genuine decorrelation, if ever needed, requires a + DESIGNED low-autocorrelation sign sequence (Legendre/quadratic-residue + Paley over a prime p≫17, or a maximal-length LFSR m-sequence), NOT the + coprime walk. **Unaffected:** the walk's bijectivity/quantization role + and OGAR D-QUANTGATE stay canon — the coprime walk QUANTIZES by + construction (bijective closure + WGSL-portable integer stride), which + this probe does not challenge; only the anti-moiré / lattice- + decorrelation consequence dies. - **(h) Exact phase-side unbinding** — sign is recomputable per address with no cleanup codebook needed; a cleanup codebook is only needed for magnitudes. The two-algebra rule (sign = XOR, magnitude = `vsa_bundle`, @@ -402,6 +426,20 @@ fourth-mode/anchor discussion (§8): **PROBE-WH-MAG:** reconstruction error of WHT₁₆+i4/i2 vs direct i4/i2 on real tile magnitudes (bgz-tensor's row-level win does not automatically transfer to 16-cell tiles). + **RESULT (2026-07-16, PROBE-WH-MAG, plan h268-probe-wave-v1):** + NEUTRAL — bare-tile leg NOT-TRANSFERRING. B/A = 0.929 (gradient+spike) + / 1.411 (heavy-tailed) / 1.869 (uniform-noise) over 256 tiles/class at + dim 16: misses the <0.9 PASS bar everywhere, regresses past 1.1 on two + classes. WHT₁₆ spreads a tile's outlier energy across all 16 cells, + inflating the per-cell quantization floor where the direct cascade + quantized 13-14 near-zero cells for free; WH helps only where energy + is pre-spread (the 0.929 smooth-gradient class) or where an outlier + escapes to a passthrough tier first. The bare-tile probe has neither + the passthrough escape nor the centroid residual the shipped row codec + (lance-graph E-PALETTE-RESIDUAL-LADDER-1) pairs WH with — so the + row-level win does not transfer to per-tile granularity, and the + shipped row codec is untouched. Follow-up PROBE-WH-MAG-2 (WH + escape + tier + centroid residual) deferred to a per-tile-codec consumer. - **[S] Signature as the replayable-trajectory checksum.** The x264 contrast in §7 asked for a replayable, checksummable stream; the Hambly–Lyons signature is the canonical path digest with a uniqueness @@ -414,6 +452,16 @@ fourth-mode/anchor discussion (§8): the Pillar-11 harness) of a temporal-stream trajectory as its digest — verify replay-identity, tree-like edits invisible (correctly), non-tree perturbations caught. + **RESULT (2026-07-16, PROBE-SIG-CHECKSUM, plan h268-probe-wave-v1):** + PASS, with a depth-2 blind-spot bound. Replay bit-exact; tree-like + mid-path A→B→A excursion invisible (0.0 < 1e-9); non-tree interior + displacement caught (11.31 > 0.05). Bound (exact identity): at DEPTH=2 + an interior displacement PARALLEL to the neighbor chord + `path[i+1]−path[i−1]` is signature-invisible (the level-2 signed-area + term vanishes; level-1 is endpoint-invariant), so the digest's null + space strictly EXCEEDS tree-like equivalence — the depth-2 digest does + not catch ALL non-tree edits. Mitigable by depth 3 or a paired + independently-projected digest. Both stay conditional on the same ledger above; neither adds a stored field to the tile (the WH rotation is derivable, the signature is a derived digest). From 95aec8e63568749ff933650a813f62a8e5f8159c Mon Sep 17 00:00:00 2001 From: Claude Date: Thu, 16 Jul 2026 09:23:37 +0000 Subject: [PATCH 2/2] docs: update heavy-tailed ratio to 1.317 (post review fix on the probe side) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The lance-graph probe wave's gen_heavy_tailed collision fix (distinct outlier indices) shifted only the heavy-tailed class: B/A 1.4113 -> 1.3169. Verdict unchanged (NEUTRAL, bare-tile leg not-transferring). Updated in the §10 RESULT line, the §5 results sub-table, and the blackboard entry. Knowledge-doc suite green (117/117). Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_01K3RyLEbuNSHxxB3NTTrGki --- .claude/blackboard.md | 2 +- .claude/knowledge/pr-x12-h268-morton-wgpu-synergies.md | 4 ++-- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/.claude/blackboard.md b/.claude/blackboard.md index 9bc475a4..63590010 100644 --- a/.claude/blackboard.md +++ b/.claude/blackboard.md @@ -790,7 +790,7 @@ default is x86-64-v3 (avx2) so ndarray_simd activates on avx512 builds only. ## 2026-07-16 (4) — h268-probe-wave-v1 RESULTS (adjudicated) - **PROBE-WH-MAG → NEUTRAL; bare-tile leg CLOSED NOT-TRANSFERRING.** - B/A 0.929/1.411/1.869 — WHT₁₆ spreads outlier energy tile-wide, + B/A 0.929/1.317/1.869 — WHT₁₆ spreads outlier energy tile-wide, inflating the per-cell quantization floor; the row-level win needs the passthrough escape + centroid residual the probe deliberately omits. Shipped row codec untouched. PROBE-WH-MAG-2 named, deferred. diff --git a/.claude/knowledge/pr-x12-h268-morton-wgpu-synergies.md b/.claude/knowledge/pr-x12-h268-morton-wgpu-synergies.md index 823ee567..fe203c1a 100644 --- a/.claude/knowledge/pr-x12-h268-morton-wgpu-synergies.md +++ b/.claude/knowledge/pr-x12-h268-morton-wgpu-synergies.md @@ -120,7 +120,7 @@ probes live in lance-graph `bgz-tensor`/`jc`/`helix`):** | Probe | Verdict | Numbers | Consequence | |---|---|---|---| -| PROBE-WH-MAG | NEUTRAL / bare-tile leg NOT-TRANSFERRING | B/A 0.929 / 1.411 / 1.869 (grad+spike / heavy-tail / noise) | §10 [H] leg: WH-on-bare-tiles does not transfer; shipped row codec (lance-graph E-PALETTE-RESIDUAL-LADDER-1) untouched; PROBE-WH-MAG-2 deferred | +| PROBE-WH-MAG | NEUTRAL / bare-tile leg NOT-TRANSFERRING | B/A 0.929 / 1.317 / 1.869 (grad+spike / heavy-tail / noise) | §10 [H] leg: WH-on-bare-tiles does not transfer; shipped row codec (lance-graph E-PALETTE-RESIDUAL-LADDER-1) untouched; PROBE-WH-MAG-2 deferred | | PROBE-SIG-CHECKSUM | PASS (depth-2 blind-spot bound) | replay 0.0; tree-like 0.0<1e-9; non-tree 11.31>0.05 | §10 [S] leg confirmed; digest ≠ catch-all — parallel-chord displacement is signature-invisible at depth 2 | | PROBE-WALK-SPECTRUM | KILL of §10(g) "decorrelated" half | period-17 \|R\| 0.998/0.996/0.994; lattice max 0.875 vs PRNG 0.0205 = 42.7× | §10(g) corrected: known structure ≠ low correlation; D-QUANTGATE / bijectivity unaffected | @@ -428,7 +428,7 @@ fourth-mode/anchor discussion (§8): automatically transfer to 16-cell tiles). **RESULT (2026-07-16, PROBE-WH-MAG, plan h268-probe-wave-v1):** NEUTRAL — bare-tile leg NOT-TRANSFERRING. B/A = 0.929 (gradient+spike) - / 1.411 (heavy-tailed) / 1.869 (uniform-noise) over 256 tiles/class at + / 1.317 (heavy-tailed) / 1.869 (uniform-noise) over 256 tiles/class at dim 16: misses the <0.9 PASS bar everywhere, regresses past 1.1 on two classes. WHT₁₆ spreads a tile's outlier energy across all 16 cells, inflating the per-cell quantization floor where the direct cascade