From 234b93626222df00453ccb675e8c29332e115205 Mon Sep 17 00:00:00 2001 From: Claude Date: Sat, 18 Jul 2026 23:48:11 +0000 Subject: [PATCH 1/2] =?UTF-8?q?docs(x265):=20head-to-head=20=C2=A7Results?= =?UTF-8?q?=20+=20PROBE-SPLAT-=CE=BC-HYDRATION-RHO=20queued?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Plan §Results (head-to-head): our helix Signed360 object-motion codes (432 B, ∞ PSNR bit-exact) vs x265 block-MV (43,115 B, 60.94 dB) on the SAME φ-scene. Verdict: tautological-win / model-matched lower bound (KILL did not fire, NOT a general-codec win) — establishes the amortization's concrete bit-cost (1× 6-byte Signed360/sprite/P-frame); arbitrary non-φ motion needs residuals (the [H] gate). Follow-up probe queued — PROBE-SPLAT-μ-HYDRATION-RHO (the spatial cousin): the 3DGS↔blasgraph↔gridlake presumption adjudicated (splat3d GaussianBatch already SoA → only net-new claim is address-derived μ, φ-manifold-gated; EWA-semiring DROPPED as non-commutative). Runs on a real Inria .ply. Board: lance-graph E-X265-HEADTOHEAD-1 + E-3DGS-MU-HYDRATION-1. Guard test 117/117. Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_01K3RyLEbuNSHxxB3NTTrGki --- .claude/plans/x265-sprite-replay-probe-v1.md | 35 ++++++++++++++++++++ 1 file changed, 35 insertions(+) diff --git a/.claude/plans/x265-sprite-replay-probe-v1.md b/.claude/plans/x265-sprite-replay-probe-v1.md index 00411972..89a159e2 100644 --- a/.claude/plans/x265-sprite-replay-probe-v1.md +++ b/.claude/plans/x265-sprite-replay-probe-v1.md @@ -217,3 +217,38 @@ The arc's own sprite-replay motion coding (E-SPRITE-IPB-HELIX-1: one Signed360 code per sprite per P-frame) is the thing being contextualized; the bitrate-comparison study (our object-level motion codes vs x265's per-block MV field on the SAME scene) is a NAMED follow-up, not done here. + +## Results (2026-07-18 — HEAD-TO-HEAD: our helix motion codes vs x265 MV field) + +The head-to-head the HEVC anchor set up: our object-level helix motion codes vs +x265's per-block MV search, on the SAME φ-spiral scene. Reproducer: +`lance-graph crates/helix/examples/hevc_headtohead.rs` (uses the REAL helix +public API — `ResidueEncoder::encode_signed` / `Signed360` / azimuth-roundtrip +decode — not a re-inline; 3 tests, clippy-clean). + +| Lane | Bytes | bits/frame | PSNR | Codec | +|---|---|---|---|---| +| A — x265 3.5 medium | 43,115 | 1437.2 | 60.94 dB | general block-MV, blind to scene structure | +| B — our helix `Signed360` | **432** (192 appearance + 240 motion) | 14.4 | **∞ (bit-exact)** | object-motion: 1× 6-byte code / sprite / P-frame | + +**Verdict: TAUTOLOGICAL-WIN / MODEL-MATCHED LOWER BOUND (KILL did not fire; NOT a +general-codec win).** Lane B was handed the EXACT generative model of a +self-generated φ-scene; the ∞ PSNR is the decoder re-running the same generator. +Legitimately establishes the object-motion amortization's concrete bit-cost — one +6-byte `Signed360` per sprite per P-frame = 240 B of motion for the whole +240-frame GOP — replacing x265's entire per-block MV field, bit-exact on +model-matched content. Does NOT beat x265 in general: arbitrary non-φ-manifold +motion needs stored residuals (PROBE-SPRITE-REPLAY's [H] gate), the named +follow-up. Board: lance-graph `E-X265-HEADTOHEAD-1`. + +## Follow-up probe queued — PROBE-SPLAT-μ-HYDRATION-RHO (the spatial cousin) + +The operator's 3DGS↔blasgraph↔gridlake presumption, adjudicated by the +convergence-architect (lance-graph `E-3DGS-MU-HYDRATION-1`): `splat3d::GaussianBatch` +is ALREADY SoA, so the only net-new "gridlake" claim is **address-derived gaussian +μ** (WORTH-EXPLORING on the φ-manifold — the SAME [H] gate as this head-to-head's +temporal tautology). "EWA = blasgraph semiring" DROPPED (depth-sorted Porter-Duff +*over* is non-commutative → axiom break). Named probe **PROBE-SPLAT-μ-HYDRATION-RHO** +runs on a real trained Inria `.ply` (non-φ input via `splat3d::ply`): μ-hydration +ratio + semiring-matches-EWA (commutative-⊕ vs sorted-over) + SoA-sweep throughput. +PASS→[H] / KILL→DROP bands in the board entry. From 5b01ce370ed94ceeebc3e0b4ca7fd56486ea84c6 Mon Sep 17 00:00:00 2001 From: Claude Date: Sat, 18 Jul 2026 23:58:54 +0000 Subject: [PATCH 2/2] =?UTF-8?q?docs(x265):=20head-to-head=20codex-P1/P2=20?= =?UTF-8?q?=E2=80=94=20real=20helix=20geometry=20+=20total-stream=20framin?= =?UTF-8?q?g?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Codex-P1 #740: reproducer now renders through helix's REAL signed_lift (golden- ratio azimuth, r=√u), not a golden-angle approximation → genuine helix result; lane-A re-measured on the real-helix scene (40,750 B, 1358.3 bits/frame, 61.27 dB). Codex-P2 #250: relabelled as TOTAL-stream vs TOTAL-stream (x265's bytes are the whole HEVC output, not its MV payload). Guard test 117/117. Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_01K3RyLEbuNSHxxB3NTTrGki --- .claude/plans/x265-sprite-replay-probe-v1.md | 28 +++++++++++++++----- 1 file changed, 21 insertions(+), 7 deletions(-) diff --git a/.claude/plans/x265-sprite-replay-probe-v1.md b/.claude/plans/x265-sprite-replay-probe-v1.md index 89a159e2..425fb7fc 100644 --- a/.claude/plans/x265-sprite-replay-probe-v1.md +++ b/.claude/plans/x265-sprite-replay-probe-v1.md @@ -226,20 +226,34 @@ x265's per-block MV search, on the SAME φ-spiral scene. Reproducer: public API — `ResidueEncoder::encode_signed` / `Signed360` / azimuth-roundtrip decode — not a re-inline; 3 tests, clippy-clean). -| Lane | Bytes | bits/frame | PSNR | Codec | +**Codex-P1 #740 correction:** the reproducer renders through helix's REAL +`HemispherePoint::signed_lift(n, total, sign).cartesian()` (golden-RATIO azimuth, +`r=√u`), NOT a local golden-angle approximation — so the scene IS the geometry +`Signed360` encodes and this is a genuine helix-motion-code result. Numbers below +are on that real-helix scene. + +| Lane | Bytes | bits/frame | PSNR | What the bytes are | |---|---|---|---|---| -| A — x265 3.5 medium | 43,115 | 1437.2 | 60.94 dB | general block-MV, blind to scene structure | -| B — our helix `Signed360` | **432** (192 appearance + 240 motion) | 14.4 | **∞ (bit-exact)** | object-motion: 1× 6-byte code / sprite / P-frame | +| A — x265 3.5 medium | 40,750 | 1358.3 | 61.27 dB | **TOTAL HEVC bitstream** — texture + residuals + modes + headers + MV, all syntax | +| B — our helix `Signed360` | **432** (192 appearance + 240 motion) | 14.4 | **∞ (bit-exact)** | **TOTAL our-codec stream** — sprite appearance + object-motion (1× 6-byte code / sprite / P-frame) | + +**This is a TOTAL-stream vs TOTAL-stream comparison, NOT motion-vs-motion +(codex-P2 #250).** Lane A's 43,115 B is the *entire* HEVC output — the 432 B does +**not** replace only x265's MV field, it replaces x265's whole stream, which it +can only do because it is model-based. Isolating x265's MV-only bits (via `--csv` +per-frame syntax analysis) to test the "one code replaces the per-block MV field" +thesis DIRECTLY is a named follow-up, not measured here. **Verdict: TAUTOLOGICAL-WIN / MODEL-MATCHED LOWER BOUND (KILL did not fire; NOT a general-codec win).** Lane B was handed the EXACT generative model of a self-generated φ-scene; the ∞ PSNR is the decoder re-running the same generator. Legitimately establishes the object-motion amortization's concrete bit-cost — one 6-byte `Signed360` per sprite per P-frame = 240 B of motion for the whole -240-frame GOP — replacing x265's entire per-block MV field, bit-exact on -model-matched content. Does NOT beat x265 in general: arbitrary non-φ-manifold -motion needs stored residuals (PROBE-SPRITE-REPLAY's [H] gate), the named -follow-up. Board: lance-graph `E-X265-HEADTOHEAD-1`. +240-frame GOP — as a TOTAL-stream lower bound on model-matched content. Does NOT +beat x265 in general: (1) arbitrary non-φ-manifold motion needs stored residuals +(PROBE-SPRITE-REPLAY's [H] gate); (2) x265's bytes encode actual pixels while +ours encode model parameters — different things. Board: lance-graph +`E-X265-HEADTOHEAD-1`. ## Follow-up probe queued — PROBE-SPLAT-μ-HYDRATION-RHO (the spatial cousin)