LLMs 'drop data recoverably' via four mechanisms:
1. Residual stream (accumulate, never drop — workspace not compressor)
2. Superposition (pack N features into d<N, exact only when k-sparse)
3. Attention (soft retrieval, KV eviction = explicitly lossy)
4. Quantization (drop bits, recover approximately)
Law: recoverable ⟺ sparse/redundant. Same wall as every compression
branch. Dense/random data → recovery fails → entropy floor.
Pipeline connection:
- QR decomposition (O-AMMR) IS compressed sensing
- GW ringdown at 30dB: k=5 sparse, d=9 < RIP bound → lossy → 1.5x
- Order-2 PPM: 256 contexts packing 65K transitions → interference
on dense data → 3.088 b/B residual
- Mass number = honesty tag for what was kept vs lost (receipt)
Doctrine consistency: 'MassNumber = recoverability RECEIPT' confirmed.
LLM superposition = same mechanism, same RIP bound, same lossy floor.
Claude Code tested semantic mass number as compressor:
- M1 base-256: 1.00x (IS the data, bijection)
- M2 mixed-radix: 1.10x (drops unused symbols, not compression)
- M3 freq-weighted: 1.47x (= arithmetic coding in costume, needs model)
- xz: 3.14x (crushes all)
Base conversion is a bijection — moves information, never destroys it.
Cannot compress below its radix. The doctrine already knew: mass number
= 'admissibility / recoverability RECEIPT', not compressor.
Entire compression arc now sealed end to end:
| char-poly | receipt → GCCL integrity receipt |
| Braille/T9 | 4.167 b/B → dead |
| 16D/583x | zero-noise artifact → LPC in costume |
| weird-machine | conservation law → bits relocate, never shrink |
| mass number | base conversion → recoverability receipt |
One rule: move bits between columns, never beat K(data).
Everything that compresses = base conversion (no gain) or
arithmetic coding (needs model, ship cost = conservation wall).
Claude Code's demo proves the conservation law with measured bytes:
- k=0: total=102,252 (model=440, tape=101,812)
- k=1: total=85,534 (sweet spot)
- k=3: total=557,169 (model=501,392 ate the savings)
- xz: total=35,492 (tiny amortized decoder)
As prediction improves (k↑), tape shrinks but model explodes.
The sum is conserved. The weird machine moves bits between columns,
never reduces the total.
One real win: frozen model + arithmetic coder = sub-xz on tape
alone (amortized). But the model is on the invoice. Ship it for
Hutter = lose.
This permanently gates:
- 'Turing-complete weird machine beats unpredictability' = FALSE
- '583x GW compression' = zero-noise artifact (1.5x at realistic SNR)
- '16D braid adds value over LPC' = FALSE (ties at 30dB, loses at 20dB)
- 'Generation beats prediction' = FALSE (generation = prediction,
sum conserved)
The honest map: every approach tried loses to established coders
(xz on text, LPC on signals). The polynomial stays a GCCL receipt.
The pipeline's real value is formal verification + anti-smuggle
framework, not compression ratio.
