e4cb242619
New folder: docs/living/ — for goals that change daily, not stable findings (those stay in docs/research/). Documents: - README.md: rules for living docs (overwrite freely, one sentence per target, link to findings, honesty tag) - TARGETS.md: 5 active targets (invariant geometry, reaction primes, merged O(1), formal cleanup, encoder fidelity) + 3 dead + 3 candidate - PROJECT_MAP.md: what SilverSight IS right now (not what it was) - MILESTONES.md: 5 near-term milestones with success criteria (3-SAT spectral test, cmix SVD, SLOS eigenvalue products, CRT hachimoji pairing, reaction prime proof) - OPEN_QUESTIONS.md: 8 unanswered questions - DIRECTION_LOG.md: 5 direction changes from this session, with the reason for each Living docs rules: - No measurement required (goals, not findings) - Overwrite freely (git history preserves old versions) - One sentence per target - Link to docs/research/ for backing evidence - Honesty tag: MEASURED / OPEN / SPECULATIVE
3 KiB
Direction Log — Why the Direction Changed
Purpose: Record when and why the project's direction shifted. Each entry is a decision point, not a finding.
2026-07-03: Compression → Invariant Geometry
Was: Build a spectral compressor (polynomial/Braille/T9/16D) that beats xz on enwik8 and LPC on signals.
Changed to: Stop compression. The conservation law (measured 8 times) forbids it. Pivot to computation shortcuts via the octagon principle (embed nonlinear in linear, detect via spectrum).
Why: Claude Code's conservation-law demo proved k=3 total (557,169B) > xz (35,492B). The model column eats the savings. No decomposition beats K(data). The polynomial is a receipt, not a compressor. Every branch measured the same wall.
What survived: The conservation law itself (as a pruning criterion), the octagon embedding (Sidon 4/4), and the formal verification framework (anti-smuggle scanner, GCCL, 20 bugs fixed).
2026-07-03: 16D Braid → LPC (承认 defeat)
Was: The 16D braid / golden spiral compresses GW ringdown 583x.
Changed to: 16D = LPC in a costume. The 583x was a zero-noise artifact. At 30dB SNR (realistic), the ratio is 1.5x — tying or losing to standard LPC.
Why: Claude Code's SNR sweep measured the parametric model vs LPC across noise levels. Clean signal: 111x. Realistic: 1.5x. The residual IS the noise, and noise is incompressible.
2026-07-03: Universal Shortcut → Problem-Specific
Was: ManifoldShortcut finds the Kolmogorov-optimal equation for any problem.
Changed to: ManifoldShortcut has ONE universal component (Shannon-entropy pruning). Everything else is problem-specific. K(data) is uncomputable; can't claim K-optimal.
Why: 5-way attack: K uncomputable (attack 1), alpha/beta free params (attack 2), RIP wrong for combinatorial (attack 3), AngrySphinx is timeout not accelerator (attack 4), coherence is linear only (attack 5).
2026-07-03: Weird Machine → Conservation Law
Was: A Turing-complete weird machine beats unpredictability by finding generating programs instead of predicting.
Changed to: Generation = prediction. The generating program = the model. The residual = what can't be predicted/generated. Sum is conserved. No machine beats K(data).
Why: Claude Code's demo: k=0 total=102K, k=1 total=85K (sweet spot), k=3 total=557K (model ate savings). Bits relocate between program and tape columns, never shrink. The machine is never free; it's on the invoice.
2026-07-03: DNA as Compressor → DNA as Invariant Carrier
Was: DNA (hachimoji) encodes data compactly for compression.
Changed to: DNA is the octagon carrier — structurally linear (compatible with the pipeline) but carrying nonlinear meaning (compatible with the problem). It's not a compressor; it's the embedding that makes nonlinear properties spectrally detectable.
Why: Braille/T9 on text: 4.167 b/B (dead). But the p-adic valuations (prime factorization) ARE the prime decomposition — the invariant signature. DNA carries invariants, not compressed bytes.