SilverSight/docs/weird_machine_conservation_law.md

Weird Machine Conservation Law: Proven with Real Bytes

The Claim That Was Tested

"A Turing-complete weird machine can beat unpredictability by finding generating programs instead of predicting."

The Conservation Law (now measured)

compressed_size = program_size + residual_size ≥ entropy_floor × data_size

The weird machine moves bits between the program column and the residual column. It never reduces the sum below the entropy floor.

Measured Results (Claude Code demo, lossless round-trip PASS)

order k	tape B	model B	TOTAL B	amortized B
0	101,812	440	102,252	101,812
1	75,806	9,728	85,534	75,806
3	55,777	501,392	557,169	55,777
xz -9	35,492	~60KB	35,492	35,492

As k increases:

• Tape SHRINKS (better prediction, smaller residual)
• Model EXPLODES (every new context = bytes to ship)
• TOTAL bottoms out at k=1, then BLOWS UP at k=3

Why xz Wins

xz's decoder is ~60KB, amortized across all files by the standard. It never ships a fat per-file model. The model column is effectively zero per file. That's why total = tape = 35,492.

The One Real Win (not Hutter)

Frozen model + arithmetic coder: k=3 amortized = 55,777 bytes, sub-xz on tape alone. A real frozen LLM would drive this lower. But the model must be shared out-of-band (not scored). The instant you ship the model (Hutter Prize), the model column dominates and you lose.

What This Permanently Gates

• "Turing-complete weird machine beats unpredictability" = FALSE
• Conservation forbids it. The machine is never free; it's on the invoice.
• Generation = prediction. The generating program = the model. The residual = what can't be predicted/generated. Sum is conserved.
• The Braille/T9/hachimoji substrate is a different decomposition, not a different bound. It changes where bits go, not whether they exist.

The GW SNR Sweep (same law, different data)

SNR	program	residual	total	ratio
clean	9 coeff	0	tiny	583x (zero-noise artifact)
60 dB	9 coeff	small	small	2.3x
30 dB	9 coeff	noise	~floor	1.5x (ties LPC)
20 dB	9 coeff	more noise	~floor	1.5x (LPC wins)

Same conservation: bits move from program to residual as noise increases. Total converges to entropy floor. Nobody beats it.

The Honest Map

Approach	Text (enwik8)	Signals (GW)	Verdict
Order-2 PPM	3.088 b/B	—	Honest baseline
Braille/T9	4.167 b/B	—	Dead (worse than PPM)
16D braid	—	1.5x (ties LPC)	Dead (adds nothing)
Polynomial	Receipt	Receipt	Receipt, not compressor
xz	1.989 b/B	—	The floor
cmix	~1.2 b/B	—	SOTA (461 models)
LPC	—	~1.5x	The signal floor
Frozen LLM + AC	sub-xz (amortized)	—	Real, but model not scored

Semantic Mass Number: Base Conversion Proof (Final Branch)

The Test

"Encode data as a semantic mass number A(H)" = represent the message as one big number (the nuclide address / 10-adic residue reading).

Measured Results (lossless round-trip PASS)

Method	bits/char	bytes	ratio	lossless
M1 base-256 mass number	8.000	100,000	1.00	PASS
M2 mixed-radix (155 symbols)	7.276	91,107	1.10	PASS
M3 freq-weighted (=arithmetic)	5.401	67,823	1.47	needs model
xz -9 (order-N + matching)	2.551	31,892	3.14	PASS

Why It Fails

Base conversion is a bijection. A bijection moves information around, never destroys it — so it cannot compress below its radix. M1 IS the data (1.00x). M2 only beats 1.00 because the data uses 155 of 256 byte values (dropping unused-symbol slack, not compression). M3 = arithmetic coding wearing a nuclide costume — and the model must ship = conservation wall.

The Doctrine Already Knew

The de-anthropocentric revision explicitly flags "English-facing semantic compression" as the OLD ERROR and redefines: "MassNumber is the admissibility / recoverability RECEIPT projected from SemanticMass."

The measurement just put numbers behind the flag.

Final Sealed Map

Idea	As compressor	Honest home
char-poly	receipt, adds overhead	GCCL integrity receipt
Braille/T9	4.167 b/B, lose to xz	dead
16D / 583x GW	zero-noise artifact, LPC win	LPC in costume
weird-machine	conservation, k=3 worst total	bits relocate, never shrink
semantic mass number	base conversion, 1.00–1.10x	recoverability receipt

One rule: you can move bits between columns, never beat K(data). Everything that "compresses" is either base conversion (bijection, no gain) or arithmetic coding (needs model, ship cost). The clever geometry buys nothing over boring xz/LPC.

LLM Recoverable Drop: Same Law, Different Substrate

How LLMs Actually "Drop Data Recoverably"

1. Residual stream = accumulate, don't overwrite. No bytes saved — the data was never dropped. The stream is a workspace, not a compressor.

2. Superposition = pack N features into d < N dimensions via near-orthogonal directions (Anthropic's Toy Models of Superposition). Recovery is exact only when k-sparse: d ≥ k·log(N/k) (RIP bound). Past that → interference → lossy.

3. Attention = soft retrieval, keeps everything. KV-cache eviction (StreamingLLM/beacons) is explicitly lossy — evicted tokens gone, reconstructed only approximately.

4. Quantization = drop precision bits, recover approximately. Lossy, bounded by weight tolerance.

The Law Under All of It

Recoverable ⟺ sparse/redundant. Every mechanism recovers the structured part and loses the noise part. Feed it dense/random data and recovery fails — same wall as every compression branch measured.

Connection to Pipeline Measurements

• GW ringdown at 30dB: signal is k=5 sparse in N=10,000 samples. RIP bound: d ≥ 5·log(2000) ≈ 37.5. Our d=9 polynomial is BELOW the bound → lossy recovery, residual = noise → 1.5x (ties LPC).

• Order-2 PPM on enwik8: 256×256 co-occurrence packs N=65,536 transitions into d=256 contexts. Sparse text = good recovery. Dense Wikipedia = interference = 3.088 b/B residual.

• QR decomposition (O-AMMR) IS compressed sensing. Eigenvalue spectrum IS the sparse feature set. Golden spiral IS RIP-compliant recovery. But it's LOSSY — the residual is lost.

The Honest Split

• Lossless: dead. K(data) is conserved across every substrate.
• Lossy recoverability: real, useful, must be judged vs JPEG/Opus at equal distortion. The QR projection recovers the dominant structure (sparse part), loses the residual (noise part). The mass number IS the honesty tag for what was kept vs lost.

Final Consistency Check

The doctrine says: "MassNumber is the admissibility / recoverability RECEIPT projected from SemanticMass." The measurement confirms:

• As lossless compressor: dead (base conversion, 1.00x)
• As lossy projection: real (recovers sparse structure, loses noise)
• As receipt: exactly what the spec says (records what was kept)
• LLM superposition: same mechanism, same RIP bound, same lossy floor

Coda: π as Tape LUT (Cleanest Proof)

The Test

"Encode data as an offset into π — every string appears in π's digits, so store just the offset."

Measured on Real π (1,000,000 digits)

k (data digits)	avg 1st-occurrence pos	log10(pos) = offset digits
1	9	0.96
2	118	2.07
3	983	2.99
4	10,007	4.00
5	109,256	5.04
6	403,165*	5.61*

*only 198/300 found in 10^6 digits (survivorship-biased small; extend to 10^7 and avg → ~10^6, offset → 6.00)

The Result

Slope = exactly 1. Offset digits ≈ data digits. The pointer-into-π is the same size as the data it points to. Base conversion — the offset is the data's rank in another base. Zero gain.

The BBP Nuance (Real but Doesn't Beat Conservation)

π has the BBP formula: compute the n-th hex digit without previous ones, O(n log n) time, tiny space. So π is a random-access tape you never store. The "LUT is astronomically big" objection doesn't apply.

But: random access is free; the address is not. You still must store a ~k-digit offset to name k digits of data. BBP makes the shelf free; the call number is still as long as the book.

Also: π-normality is only conjectured. "Every string appears" isn't proven. Losslessness not guaranteed for all inputs.

Verdict

π-as-tape-LUT = base conversion + a free-to-read shelf. The address carries all the bits. Same wall as semantic mass number, now wearing π. Cleanest single proof of the base-conversion law: real π, slope-1, measured in half a second.

Implication for Dense Computation Targeting

Even with a free-to-read tape (π via BBP), the address IS the information. The conservation law is substrate-independent — it doesn't matter whether the tape is stored, computed, or given by physics. The sparse structure you target must be genuinely sparse (low-rank, k-sparse, RIP-compliant), not just "looked up from a big table." Look-up = base conversion = no gain.

SLOS Direction: What the Pipeline Actually Needs

The Two Computation Paths

Path	What it computes	Cost	Gives
SLOS (Perceval, K=2)	Full output distribution, all M_n states	O(n × M_n)	Exact Omega
QR/spectral (numpy, K=1)	Eigenvalue spectrum → first column of U	O(n³)	Approximate Omega

K=2 (SLOS) gives the correct Sidon vs non-Sidon signal (4/4). K=1 (spectral) gave an INVERTED signal initially.

What Each Pipeline Query Actually Needs

Query	Needs SLOS?	Why
Omega (exhaust mode probabilities)	YES (K=2)	K=1 approximation was wrong
Sidon property check	NO	is_sidon() = pure integer arithmetic
GCCL coherence gate	NO	IS the eigenvalue spectrum
QR rank/spectrum	NO	O(n³) eigendecomposition
Collision count	NO	Integer pairwise-sum check

The pipeline runs SLOS to compute ONE number (Omega) from the full M_n-state distribution. Everything else is computed without SLOS.

The Conservation Law Applied to SLOS

The K=2 interference pattern IS the "residual" — the part the K=1 spectrum can't predict. This is the same conservation law:

K=1 spectrum (program/model) + K=2 interference (residual) = full distribution

The spectrum captures the sparse structure (block-diagonal modes, dominant eigenvalues). The K=2 interference captures the dense correlations (photon-photon interactions, Hong-Ouandel interference).

You can't predict Omega from the spectrum alone because Omega depends on the K=2 interference — which IS the incompressible residual.

The Honest Application

The shortcut for the SLOS direction is NOT "skip SLOS using the spectrum." The shortcut is: don't run SLOS for queries that don't need it.

• Sidon check: run is_sidon() (O(n²) integer arithmetic, no SLOS)
• GCCL gate: compute eigenvalue spectrum (O(n³), no SLOS)
• QR rank: compute eigenvalue spectrum (O(n³), no SLOS)
• Collision count: integer pairwise-sum check (O(n²), no SLOS)
• Omega: MUST run SLOS (K=2 interference is irreducible)

The pipeline was running SLOS for ALL queries when only Omega needed it. The shortcut: skip SLOS for 4/5 queries, run it only for Omega.

What This Saves

For an n=8, m=8 system:

• Full SLOS: O(8 × 6435) = O(51,480) operations
• Skip SLOS for 4/5 queries: save 4 × 51,480 = 205,920 operations
• Only run SLOS for Omega: 51,480 operations (1/5 of before)

This is a real 5x speedup — not from compression, but from not running expensive quantum simulations for queries that only need integer arithmetic or eigenvalue decomposition.

What This Does NOT Save

The Omega computation itself still needs full SLOS (K=2). The K=2 interference pattern is the irreducible residual. The conservation law says: you can't predict it from the K=1 spectrum. The spectrum is the model; the K=2 interference is the tape. Both must be computed.

The Quandela SLOS as "Weird Machine"

The SLOS IS a weird machine: a photonic quantum simulator repurposed as a Sidon set classifier. The "weirdness" is that a quantum optics experiment (designed for photon counting) is being used to detect combinatorial properties (Sidon sets) of integer sets.

But the conservation law applies: the quantum computation (K=2 SLOS) gives information that the classical computation (K=1 spectrum) can't recover. This IS the quantum advantage — not in computation speed, but in information content. The K=2 interference pattern contains genuinely new information (the residual) that the K=1 spectrum lacks.

This is the honest quantum advantage: SLOS computes something the spectrum can't predict. Not because of quantum speedup, but because the K=2 correlations are fundamentally denser than the K=1 spectrum.