|
3 | 3 | ## READ BY: truth-architect, savant-research, family-codec-smith, |
4 | 4 | ## palette-engineer, integration-lead, container-architect |
5 | 5 |
|
6 | | -## STATUS: 5 corrections absorbed, 0 probes run. All claims are CONJECTURES |
7 | | -## until the probe queue drains. |
| 6 | +## STATUS: 6 corrections absorbed, 0 probes run. All claims are CONJECTURES |
| 7 | +## until the probe queue drains. Correction 6 (2026-06-14) reclassifies |
| 8 | +## the whole float-reconstruction framing as a category error for |
| 9 | +## byte-encoded models — see the Correction Chain. |
8 | 10 |
|
9 | 11 | --- |
10 | 12 |
|
@@ -42,6 +44,28 @@ Iteration 5: Slot D = CLAM tree path (bucketing > resolution) |
42 | 44 | → Correct: Slot D is PRIMARY (bucket address), Slot V is REFINEMENT (exact value) |
43 | 45 | → Correct: 3-level 16-way CLAM = 12 bits → 16 → 256 → 4096 alignment |
44 | 46 | → UNVERIFIED: requires Probe M1 (CLAM tree fit on 256 Jina centroids) |
| 47 | +
|
| 48 | +Iteration 6 (operator, 2026-06-14): float codec on byte-encoded model = CATEGORY ERROR |
| 49 | + → Measured: bgz-hhtl-d certified on real Qwen3-TTS-1.7B FAILS reconstruction on |
| 50 | + all 16 roles (palette/hhtld cos ~0.05–0.32, Base17 cos ~0.01–0.37 at the first |
| 51 | + fold, matvec rel-err ~1.0, fisher-z ρ 0.03–0.68) — cert at |
| 52 | + .claude/knowledge/certification/hhtld_qwen3tts17b.json. The scalar Slot V (BF16 |
| 53 | + residual) adds ≈0 over palette-only (palette_cos ≈ hhtld_cos). |
| 54 | + → Root insight: this is NOT just a weak residual. It is a category error. Qwen's |
| 55 | + information lives in a BYTE-QUANTIZED REGISTER (GGUF / byte-BPE / bgz7), not a |
| 56 | + continuous float field. Scoring float-RECONSTRUCTION fidelity (Base17 fold + |
| 57 | + cosine/L2) on a discrete/byte representation measures the wrong category. Same |
| 58 | + failure mode as I-VSA-IDENTITIES, one layer down: decoding a quantized register |
| 59 | + back to f32 to score cosine IS the register-loss the iron rule forbids. |
| 60 | + → Correction: for byte-encoded models the success metric is discrete CODE-AGREEMENT |
| 61 | + / bucket-ROUTING (the rolling floor), NOT reconstruction. The codec should never |
| 62 | + decode to f32 to score cosine. Reconstruction is the wrong question. |
| 63 | + → Evidence the right-category path works (ndarray PR #218 probe family): |
| 64 | + rolling_floor_probe (coarse code routed via a self-calibrating μ+3σ floor, |
| 65 | + ρ=1.0 tracking, reject stable ~0.1% under SD drift), helix_bitdepth_probe |
| 66 | + (direction as a discrete golden index, 24-bit lossless vs ≤f16), and |
| 67 | + morton_perturbation_probe (discrete frame drives a field on-demand). None |
| 68 | + decode-to-f32-to-score-cosine. |
45 | 69 | ``` |
46 | 70 |
|
47 | 71 | ## The Five Hard Constraints |
|
0 commit comments