Cubrim — Matrix regeneration from a descriptor

The central question

Suppose a matrix M has a short descriptor D (a reference, a seed, a feature set), |D| ≪ |M|. Is there a procedure with procedure(D) = M exactly (lossless) where |D| plus the amortized cost of regeneration is smaller than storing or transmitting the matrix itself? And does quantum search or a supercomputer change the answer? We take it apart piece by piece: the general case, the lexicographic index, the quantum angle, and the subset where regeneration is genuinely possible.

First — the honest wall

There are only 2^(d+1) descriptors of length ≤ d bits, but exactly 2^B matrices of length B bits. So any «descriptor → matrix» procedure addresses only a vanishing fraction of the space: a 32-byte descriptor covers about 10^−1156 of all 4096-bit matrices. This is an information bound, not a speed problem — it is not bypassed by a supercomputer or a quantum machine (Grover only doubles the reachable descriptor length, leaving the wall exponential; Holevo’s theorem forbids any «quantum bypass»). Regeneration is therefore possible only where a matrix has a short description — and such matrices are exponentially rare in the space. The real research question is their share not in the space but in the data actually encountered on the fleet. That is what the REGEN axis measures.

Research in progress

The regeneration hypotheses are run by a separate research axis (REGEN, CUBR-0055). The cards below are its live log: statuses come straight from the research database. While measured is empty a card honestly stays OPEN («under study»); measured numbers will appear here automatically once the REGEN axis obtains them. Nothing is invented in advance.

NO-GO · 4

structurally cannot win (boundary) · 3 near-match + delta · 1

Generated: 2026-07-16T21:01:03Z · db:addressor_hypotheses

Regeneration hypotheses · 4

AH-25 NO-GO REGEN · 2026-07-16

Matrix regeneration from a descriptor: the counting wall (2^B)

Can an arbitrary exact matrix M be reconstructed from a short descriptor D (|D| ≪ |M|) by search or generation — instead of storing and transmitting the matrix itself? The general case, with no assumed structure.

Data class (Z)

Arbitrary B = 4096-bit matrices (512-byte cube), no assumed structure

Address target

n/a

Predicted lever

None — an information-theoretic impossibility, not a speed problem

Ceiling category

structurally cannot win (boundary)

Mechanism

There are at most 2^(d+1) descriptors of length ≤ d bits, but exactly 2^B matrices of length B bits. Any deterministic procedure(D) → M covers at most a 2^(d+1)/2^B fraction of the space: a 32-byte descriptor addresses ≈ 10^−1156 of the B = 4096 space. Average-case incompressibility: the fraction of B-bit strings compressible by even k bits under any fixed lossless scheme is ≤ 2^(−k+1) (for k = 32 bits, < 5·10^−10). So regeneration is possible ONLY for matrices with a short description (low Kolmogorov complexity), and their share of the space is exponentially small. This is the same 2^B wall as the matrix-count card, seen from the decoder side.

Falsification test

GO only if there exists a non-empty class of matrices actually encountered on the fleet for which |D| + amortized_cost(regeneration) < min(CAS reference, Cubrim-1 block) losslessly. Phase-2 prediction (REGEN): that share is single-digit percent or less, and it is degenerate (const/zeros that CAS already dedupes to one copy).

Full cost (total_cost)

Descriptor + amortized cost of the procedure + hash-verification of each candidate. For an incompressible M a blind search is ≈ 2^B candidates; the Lloyd bound (~10^120 ≈ 2^399 ops for a universe-sized computer) rules out descriptors d ≥ 400 bits forever. An address into an exhaustive/random catalogue costs no less than the content itself.

Probe verdict

NO-GO — the counting wall is confirmed on the real fleet: 93–98% of unique matrices carry no generator structure at all; their empirical Kolmogorov estimate equals the entropy-coding price of 339–391 B/matrix (66–76% of raw), and no shorter descriptor is informationally possible. Blind search measured live: a 2^32 sweep costs 4.6 CPU-minutes PER MATRIX (the whole devs corpus = 6.8 years on 16 cores), 2^64 = 2,330 years; no break-even N exists. Regenerating an arbitrary matrix from a short descriptor fails not on speed but on counting: there are 2^|D| descriptors versus 2^B matrices.

Measured result (raw research log; descriptors partly in the RU original)

date: 2026-07-16
probe: REGEN-spectrum-v1 + REGEN-bruteforce-cost-v1 (CUBR-0055)

corpus

AH-23 кросс-девайс, 13.48 ГБ / 3 хоста; уникальные матрицы B=4096 бит: devs 12 451 314, prod 1 531 446, www 2 824 381 (payload не покидал хосты, только JSON-агрегаты)

verdict: NO-GO

verdict_note

NO-GO — the counting wall is confirmed on the real fleet: 93–98% of unique matrices carry no generator structure at all; their empirical Kolmogorov estimate equals the entropy-coding price of 339–391 B/matrix (66–76% of raw), and no shorter descriptor is informationally possible. Blind search measured live: a 2^32 sweep costs 4.6 CPU-minutes PER MATRIX (the whole devs corpus = 6.8 years on 16 cores), 2^64 = 2,330 years; no break-even N exists. Regenerating an arbitrary matrix from a short descriptor fails not on speed but on counting: there are 2^|D| descriptors versus 2^B matrices.

bruteforce_measured

d64_16cores_years: 2 330
candidates_per_sec_1proc: 15 667 239
exa_1e18cps_2pow96_years: 2 510
devs_corpus_16cores_years: 6.8
sweep_2pow32_cpu_min_per_matrix: 4.57

coverage_arithmetic

compressible_by_32bit_share: <2^-31
descriptor_32B_covers_share_of_2pow4096: 2^-3839 (~1e-1156)

store_bar_bytes_per_matrix

raw: 512
zlib9_per_block: 391.2
zstd19_discovery: 339.3
cubrim1_discovery: 352.1

no_generator_structure_share_pct

www: 92.68%
devs: 93.97%
prod: 97.9%

AH-26 NO-GO REGEN · 2026-07-16

Lexicographic index of a matrix-as-number

If all length-B matrices are ordered lexicographically, a matrix’s binary index is its content — regeneration is trivial and no search is needed. Does this save anything?

Data class (Z)

B = 4096-bit matrices; the full space versus the actually-encountered subset M ≪ 2^B

Address target

n/a

Predicted lever

Zero saving over the full space; over the encountered subset it reduces to CAS (log₂M ≈ 24 bits); the only new measurable is an order bonus (neighbour front-coding)

Ceiling category

structurally cannot win (boundary)

Mechanism

Lexicographic order on B-bit strings is the identity bijection «string ↔ number 0..2^B−1»; the index of an arbitrary matrix takes exactly log₂(2^B) = B bits — as many as the matrix itself (the AH-25 counting wall from the other side: any bijection preserves the count, no other ordering helps). Saving appears only when you index not the space but the list of M actually-encountered matrices: the index costs log₂(M) bits (union scan M ≈ 1.5·10^7 → 24 bits ≈ 3 B) — which is exactly the Addresser’s CAS/dedup (AH-02/05, ordinal references AH-08), and the catalogue itself must still be stored. The one new measurable claim is an order bonus: store the sorted list of encountered matrices as numbers + deltas to the neighbour (front-coding); real saving beyond log₂(M) appears ONLY if encountered matrices cluster lexicographically (long shared neighbour prefixes). Geometric honesty: lexicographic closeness ≠ near-match closeness — matrices differing by 1 bit in the first byte sit 2^4088 positions apart; the AH-15 axis works in chunk space, not in lex order.

Falsification test

Saving beyond the theoretical minimum log₂(M) ≈ 0.6% appears only under strong clustering of shared neighbour prefixes. REGEN prediction (P3): the bonus ≈ the theoretical minimum, prefix clustering is weak (real matrices do not fill lex neighbourhoods).

Full cost (total_cost)

The catalogue of encountered matrices must be stored and transmitted — its price is precisely the price of the matrix-as-storage (the canonical total_cost). An index without a catalogue is meaningless.

Probe verdict

NO-GO — an ordering bonus does exist (2.07–2.38% vs the raw list; clustering above the uniform bar of 61–77 bits/matrix; front-coding +4–13% vs shuffle), but as a catalog format delta-coding of sorted indices (504 B/matrix) is 48% WORSE than a plain zstd-19 store in natural discovery order (339 B/matrix), and lex-sorting makes real codecs WORSE than discovery order (zstd-19 −2.7%, Cubrim-1 −1.04%): file locality beats lexicographic proximity. Gap bits are approximated by the highest differing byte (big-int validation on 100k pairs: max error 49–160 bits on rare pairs, biased in favor of the lex scheme — the verdict stands).

Measured result (raw research log; descriptors partly in the RU original)

date: 2026-07-16
probe: REGEN-lex-order-v1 (CUBR-0055)
verdict: NO-GO
full_space: индекс = B бит тождественно (биекция), экономия 0 — подтверждено арифметикой, замер не нужен

verdict_note

NO-GO — an ordering bonus does exist (2.07–2.38% vs the raw list; clustering above the uniform bar of 61–77 bits/matrix; front-coding +4–13% vs shuffle), but as a catalog format delta-coding of sorted indices (504 B/matrix) is 48% WORSE than a plain zstd-19 store in natural discovery order (339 B/matrix), and lex-sorting makes real codecs WORSE than discovery order (zstd-19 −2.7%, Cubrim-1 −1.04%): file locality beats lexicographic proximity. Gap bits are approximated by the highest differing byte (big-int validation on 100k pairs: max error 49–160 bits on rare pairs, biased in favor of the lex scheme — the verdict stands).

order_bonus_vs_raw_pct

www: 2.1%
devs: 2.07%
prod: 2.38%

hamming_lt64_to_lex_neighbor_pct

www: 1.03%
devs: 0.85%
prod: 0.26%

encountered_subset_flat_index_bits

www: 21.43
devs: 23.57
prod: 20.55

clustering_above_uniform_bits_per_matrix

www: 64.5
devs: 61.4
prod: 76.8

front_coding_zlib9_sorted_vs_shuffled_pct

www: 6.5%
devs: 4.1%
prod: 12.6%

catalog_format_bytes_per_matrix_devs_sample

zstd19_lex_sorted: 348.4
cubrim1_lex_sorted: 355.7
zstd19_discovery_order: 339.3
cubrim1_discovery_order: 352.1
delta_coded_sorted_elias: 504

AH-27 NO-GO REGEN · 2026-07-16

Quantum search (Grover): √-speedup, not a wall bypass

Does quantum search (Grover) or a supercomputer break the information wall of regeneration — can a descriptor for an arbitrary matrix be found quantumly?

Data class (Z)

Arbitrary B = 4096-bit matrices; quantum hardware or a hypothetical exaflops brute-forcer

Address target

n/a

Predicted lever

The reachable descriptor length doubles (exponent d → d/2), the wall stays exponential; the information bound is not quantum-bypassable

Ceiling category

structurally cannot win (boundary)

Mechanism

Unstructured search over N candidates is Θ(√N) oracle calls (π/4·√N), and this is PROVABLY optimal (BBBV): there is no exponential quantum speedup of blind search. For a descriptor of length d: classically 2^d → quantumly ≈ 2^(d/2). The effect is that the reachable length doubles, no more: the stand’s classical limit (matrix-count threshold) d ≈ 51 → Grover-equivalent d ≈ 102; a generous planetary limit ~2^90 ops/year → Grover d ≈ 180 (ideal hardware); still ~3.9 thousand bits short of B = 4096. Quantum does not bypass pigeonhole — Holevo’s theorem: n qubits carry ≤ n classical bits, superposition does not turn a short descriptor into a long matrix. There is NO «quantum bypass» of the information bound — any claim otherwise is hype.

Falsification test

NO-GO by construction for arbitrary matrices (the information bound is not quantum-bypassable). OPEN note without numbers (not a basis for design): if the generator-candidate space is structured (a hierarchy of families with priors), amplitude amplification / QAOA might speed up a structured fit — but today there is no hardware, no metric, and no demonstrated case on data of our kind.

Full cost (total_cost)

Hardware calibration: a Grover attack on AES-128 (d = 128, the same arithmetic) is ≈ 2^83 logical gate-operations, millions of error-corrected physical qubits, and time on the order of decades-to-a-century per SINGLE target. The physically conceivable Grover zone today is d ≤ 80–100 bits — covering only matrices with K(M) ≤ 16 B, where an algebraic fit is already cheaper than any search.

Probe verdict

NO-GO — quantum gives the sqrt of the same exponential and never touches the information wall: reachable |D| moves from ~50–64 bits (classical; measured stand at 15.7M candidates/s) to ~100–128 bits (idealized Grover), while real matrices need ≈2,700+ bits. No quantum bypass of the counting bound exists (Holevo/BBBV) — recorded explicitly as hype protection.

Measured result (raw research log; descriptors partly in the RU original)

date: 2026-07-16

probe

REGEN-theory §4 (арифметика/литература) + классический замер REGEN-bruteforce-cost-v1 как базис экстраполяции (CUBR-0055)

grover

speedup: sqrt (оптимум по BBBV — экспоненциального квантового ускорения слепого перебора не существует)

d128_cost

2^64 оракул-вызовов ≈ масштаб Grover-атаки на AES-128: ~2^83 гейт-операций с QEC, миллионы кубитов, десятилетия на одну цель

grover_equiv_reach_bits: 102
classical_stand_reach_bits: 51

verdict: NO-GO

open_tail

структурированный квантовый фит (амплитудная амплификация/QAOA по иерархии генераторов) — спекулятивно, железа/метрики/кейса нет; не основание для дизайна

verdict_note

NO-GO — quantum gives the sqrt of the same exponential and never touches the information wall: reachable |D| moves from ~50–64 bits (classical; measured stand at 15.7M candidates/s) to ~100–128 bits (idealized Grover), while real matrices need ≈2,700+ bits. No quantum bypass of the counting bound exists (Holevo/BBBV) — recorded explicitly as hype protection.

information_bound

pigeonhole неквантуем: теорема Холево — n кубитов передают ≤ n классических бит; 2^|D| дескрипторов не адресуют 2^B матриц ни на каком железе

required_descriptor_bits_for_real_matrices: ≈2700+ (эмпирическая K-оценка 339–391 Б/матрицу из AH-25/AH-28 замеров)

AH-28 NO-GO REGEN · 2026-07-16

The genuinely regenerable subset (generators and recipes)

Is there a non-empty class of matrices actually encountered on the fleet where a short descriptor + amortized regeneration losslessly beats CAS storage and the Cubrim-1 block?

Data class (Z)

Encountered matrices of low Kolmogorov complexity: const/zeros, RLE, arithmetic/affine, periodic, PRNG output with a recoverable seed, version-chain recipes

Address target

n/a

Predicted lever

Descriptor = generator-family id + parameters (ones-to-tens of bytes); a categorical win only on raw PRNG output (8-B seed versus ≈1.0× codec compression)

Ceiling category

near-match + delta

Mechanism

Three classes. (a) Procedurally generated (const/zeros, RLE, arithmetic/affine, periodic, PRNG with recoverable seed): descriptor = family + parameters; but an entropy codec already compresses exactly these structures close to their K, so a win exists only if the generator-descriptor is shorter than the Cubrim-1 block. The exception where regeneration categorically beats the codec is high-quality PRNG output: statistically indistinguishable from noise (Cubrim-1/zstd compress it ≈1.0×), yet an algebraic seed-fit yields an 8-byte descriptor. (b) Known-generator «recipe instead of result» (build from source, scene render, deploy from git-SHA) — already absorbed by the version-chain axis AH-15 (GO: delta 4.2–8.5× vs the ultra bar; router-MVP factor 64.37 on 339 real git pairs) as base+delta; hash-exact toolchain rebuild is a separate engineering axis (reproducible builds), not an Addresser one. (c) Low-entropy blocks: «regeneration» = decompression, «descriptor» = the compressed block — precisely the Cubrim-1 local backend, nothing beyond it.

Falsification test

The measured quantity is the share of unique matrices in a real scan whose generator-descriptor (class a) is shorter than the Cubrim-1 block. REGEN predictions: P1 the share is single-digit percent or less, its mass degenerate (const/zeros that CAS dedupes anyway); P2 the raw-PRNG class has ≈ 0 mass on the real fleet; P4 brute-forcing even a 32-bit seed space (~seconds of CPU × M matrices) does not amortize at any realistic N against storing a compressed block (~hundreds of bytes).

Full cost (total_cost)

Family descriptor + parameters + amortized fitting cost (an algebraic fit is polynomial; a blind search is exponential, 2^L). A win exists only if the total cost < the Cubrim-1 block on a real class with non-zero mass; otherwise class (c) reduces regeneration to ordinary decompression.

Probe verdict

NO-GO as a new mechanism — the regenerable subset is mathematically non-empty but economically empty: a generator descriptor beats the compressed block for 0.025% of matrices, and the exact classes save 900 bytes across the ENTIRE fleet (upper bound for all winners: 0.016% of the store); raw PRNG output — the only class where regeneration categorically beats entropy coding — never occurred (0 of 16.8M, consistent with AH-03). The living part of "store the recipe, not the result" has already shipped as version-chain AH-15.

Measured result (raw research log; descriptors partly in the RU original)

date: 2026-07-16
probe: REGEN-spectrum-v1 (CUBR-0055)
verdict: NO-GO

verdict_note

NO-GO as a new mechanism — the regenerable subset is mathematically non-empty but economically empty: a generator descriptor beats the compressed block for 0.025% of matrices, and the exact classes save 900 bytes across the ENTIRE fleet (upper bound for all winners: 0.016% of the store); raw PRNG output — the only class where regeneration categorically beats entropy coding — never occurred (0 of 16.8M, consistent with AH-03). The living part of "store the recipe, not the result" has already shipped as version-chain AH-15.

recipes_class

покрыт version-chain AH-15 (GO, фактор 64.37 в router-MVP CUBR-0054) — нового механизма не требует; hash-exact пересборка из тулчейна = отдельная инженерная проблема reproducible-builds, не адресаторная

prng_fits_total: 0

descriptor_beats_zlib9

www: 871/2824381 (0.031%)
devs: 3003/12451314 (0.024%)
prod: 262/1531446 (0.017%)
total: 4136/16807141 (0.025%)

falsification_P1_P2_P4: все подтверждены: доля <<1% и вырождена; PRNG-масса 0; брутфорс не амортизируется ни при каком N
generator_families_tested: constarithxor-deltaperiodic<=64xorshift32xorshift64LCG-NR32RLE
algebraic_fit_cost_us_per_matrix: 62–99 (store-side, one-time)
exact_class_savings_bytes_fleetwide: 900
winners_upper_bound_share_of_store_pct: 0.016%

What this means

Universal regeneration — «compress any file into a short descriptor and rebuild it by search» — is impossible twice over: information-theoretically (there are fewer descriptors than matrices) and computationally (for an incompressible matrix the search is an exponential that neither a quantum machine nor a universe-sized computer can take). What remains is a narrow subset: procedurally generated matrices (const, RLE, arithmetic, periodic, a recoverable PRNG seed) and data with a known recipe (build from source, deploy from a git-SHA), where a short descriptor really is shorter than the compressed block. The practical part of that subset the Addresser already covers with deduplication (CAS), ordinal references, and the version-chain axis (AH-15). How large the remaining share is on real data is precisely the measurable question; the numbers will appear in the cards above as they are obtained.