CSF at cosmological scale — the honest frontier map (observer slices, generative members, and the Library of Congress)

0. What stays dead — restated once, then built around

"Novel higher-dimensional tesseract compression" as a codec is permanently closed. The Shannon/Kolmogorov counting bound is scale-invariant: geometric re-arrangement of bytes buys zero entropy atMB and zero atPB, and the repo killed its own version of this claim with run experiments (E1/E2, 2026-06-28). Nothing below reopens it.

What the vision's words CAN honestly mean — and this is where the frontier really is:

1. "Contain the universe" = store the description (laws + seed), materialize *observer

slices* on demand, verify each materialization against a recorded hash. Description-length, not entropy-defeating: lawful data was never random.

1. "Tesseract observer slice" = the wavefront pattern (materialize only what is observed)

applied to a generatively-defined space — the one part of the old tesseract design that was always sound engineering.

1. "Every PDF in the Library of Congress" = corpus-scale tiering where dedup and

amortized shared-model coding become honest — the exact regime where the E1 kill's model-accounting objection mathematically dissolves.

1. Defining a frontier (the specifics)

A compression frontier is a Pareto surface, not a number. CSF's axes:

Enhancing the frontier = moving one measured point outward under a pre-stated kill criterion, without regressing another axis silently. CSF's measured frontier today:

2. The observer-slice universe, honestly (F1 — seed shipped)

Shipped today [implemented]: generative members (spec §2.2.2, format v0.9). An archive entry stores a tiny JSON generator spec instead of bytes; readers materialize the member deterministically and verify against its recorded sha256. Closed registry only — zeros, repeat, sha256-ctr (counter-mode SHA-256 DRBG: stdlib-pure, deterministic forever) — no eval, no user code,repeatGiB materialization guard. Measured: aMiB lawful stream + siblings in a 3.5 KB archive, tamper-flip of the seed detected by hash.

This is the entire honest core of "the archive contains an observer slice of a generated universe": observation = verified materialization of a slice of a lawful process. The archive holds the laws + seed; the bytes exist when observed; the hash pins the observation. ("Multiverse" reading, labeled metaphor: each seed is one universe of the family; the registry is the physics; none of this is physics.)

The ladder up [design — falsifiable]:

• F1b — slice-addressable generators. read_slice(member, offset, len) without full

materialization — sha256-ctr is already block-random-access; per-slice Merkle hashes make partial observations verifiable. This is the wavefront made real: observe a window of a universe-sized stream at O(window) cost. Kill: slice reads must not regress whole-read throughput or integrity coverage.

• F1c — registered scientific generators. The cosmology community already lives this way:

simulations emit terabytes per timestep, and the field's answer is seeds + parameters + error-bounded recompression (SZ3 ~300× on cosmology fields, newer methods to ~3,000× with bounded error — ACM survey · arXiv:2503.20031 · GPU cosmology compression, arXiv:2004.00224). A registered nbody-emulator generator kind (spec + seed + version-pinned code hash) would put "a universe run, beginning to end" in a CSF store as its recipe, with observer slices as verified snapshots. Kill: a registered generator whose regeneration drifts across versions is removed — determinism is the contract.

2.5 The organizational reading of the geometry (the steelman, adopted)

The tesseract/qutrit design's durable value — stated by the project owner and adopted here — is organizational, not entropic: restructure data so compression and dedup can act. This is the half of the old geometry that every measurement this session vindicated, and it splits cleanly:

• The dust principle is the real inheritance. The lattice's deep idea was never base-3 —

it is "absence is free: store a baseline, pay only for deviations." That principle, extracted from the geometry, is what Col v2 does across fields (+5.2–5.8%), what solid mode does across files (+18.4–30.8%), what generative members do in the limit (a lawful stream is one large duplicate of its recipe,MiB → 3.5 KB), and what F2's chunk store does at corpus scale (canonical chunk + references = baseline + dust). RKD (#1594) is the same body in delta clothing.

• **The trit types state, not storage.** {-1, 0, +1} is the natural type for what the

system tracks — refuted/unknown/confirmed; superseded/baseline/novel — and makes a clean lifecycle vocabulary for a chunk store (retired / dust-reference / live delta). Measured fact: a 3-symbol column costs the entropy stage ≤ log₂3 bits automatically, so this elegance is free — and worth exactly zero extra compressed bytes. Use the trit because it is the right type; never expect it to compress.

• Where base-3 pays in bytes it is already registered elsewhere: trit-packing for the

ternary serving artifact (application map Q1) and PAM-3 signaling — regimes with no entropy coder in the path. Inside an archive, hand-rolled ternary arrangement measurably adds nothing (base3_cyclic lost to delta+zstd; re-verified this cycle).

One line: the geometry earns its keep as a discipline — baseline-plus-deviation, observe-only-the-slice, three-state lifecycle — and the discipline now has measured hardware under it; base-3 is the notation, not the mechanism.

Companion question, graded (owner, 2026-07-21): "use the infinite values between 0–2 as space to shrink bytes." The pure form dies on distinguishability: storage holds states, not reals — n bits of precision distinguish exactly 2ⁿ points of any interval (precision is bytes), and analog level-packing is taxed by noise (Shannon–Hartley; why PAM-3 stops atlevels and flash at ~16). But the rigorous form of the intuition is arithmetic coding — the whole message as one point in an interval, narrowed per-symbol by probability, code length = −log₂P = the Shannon optimum — which is exactly the FSE/ANS/range-coder stage inside every CSF backend codec. The continuum is load-bearing three ways here already: as probability space (shipped in every codec), as precision-priced floats (F1c error-bounded tier: bits ∝ log(range/ε)), and as the home of the surviving field/budget instruments (lapse, dilation, water-filling). Verdict: not a new lever — the mechanism of the existing floor, which is precisely why nothing hand-rolled beats that floor.

2.6 The optimum number of levels — why the wire stopped atand the cell at(researched)

One law governs both, and it is the physical twin of the archive's entropy floor: distinguishable levels are linear in the physical budget, information is logarithmic in the levels, and the cost of one more bit is exponential. With window W and lifetime noise σ (jitter over a symbol on a wire; ten years of charge leakage in a cell), usable levels are M ≈+ W/(kσ), bits = log₂M — and holding one more bit means doubling W or halving σ.

The wire (GDDR7 = PAM-3, verified): eye height shrinks as 1/(M−1) ⇒ required SNR grows ~6 dB per halving. NRZ baseline; PAM-3: −6.0 dB for 1.5 bits/symbol (2 symbols =states ≥⇒Micronbits/2 clocks, 94% packing); PAM-4: −9.5 dB forI-PEXbits (ProLabs · Samtec · I-PEX); PAM-300 would need ≈ −49.5 dB — categorically beyond aGbaud single-ended memory channel's capacity (Shannon–Hartley: usable bits/symbol = log₂(1+SNR) ≈ 1.5 here). Best margin-per-bit at speed pickedMicron(Micron · Rambus). Under linear cost-per-level the optimal radix is e ≈ 2.718 (radix economy, in canon) — the fastest bus on earth landed on its nearest integer.

The cell (NAND ≤16 levels, verified): fixed ~5–6 V window; σ_lifetime = years of leakage + P/E oxide damage + disturb + temperature (TechTarget · arXiv:1706.08642). Measured ladder: SLCFlexxonlevels ~100k P/E → MLCIEEE~10k → TLC~3–10k (3D) → QLC~150–1,000 (Flexxon · Lexar). Each added bit buys a shrinking capacity gain (+100% → +50% → +33% → +25%) at ~10× endurance cost + heavier ECC — QLC is the economic knee. PLC (32 levels) has been in pilot since(IEEE · Kioxia/WD) and theflagship push is more layers, not more levels (Tom's Hardware) — the market's revealed optimum: buying more cells beats buying more distinguishability per cell.

Unified: both maximize value(log₂M) − cost(M) with convex-to-exponential cost ⇒ optima are tiny (2–4 on wires, 8–32 in cells). The levers that escape the law are exactly this doc's ladder: organization (states you never spend — dedup, restructuring, dust) and amortization (one generator/model shared across a corpus).

3. The Library of Congress program (F2–F3)

Scale, verified: the Library manages 21 PB of digital collections acrossmillion files (2022; LoC digital collections FAQ), within a catalog of ~170M items.

• F2 — corpus-scale dedup tier [design]. Content-defined chunking before the coder:

FastCDC-class chunkers run 3–10× faster than classic CDC at equal-or-better dedup (USENIX ATC'16). Digitized-scan corpora carry massive cross-file redundancy (re-scans, editions, embedded fonts, shared plates) that per-file and solid modes both miss beyond one archive. Kill: measured chunk-store ratio on a real multi-GB PDF corpus must beat solid+omni by ≥10% at acceptable index overhead, or the tier stays unbuilt.

• F3 — amortized neural cold tier [design; the E1 verdict flips]. E1 killed

model-as-compressor atMB because the 2.87 GB model dwarfed the data. The canon's own threshold (arXiv:2601.02875): LM coding pays past ~100 GB. At LoC scale aGB shared model is 0.00005% overhead — the DeepMind class (arXiv:2309.10668) becomes honest arithmetic. Kill (pre-stated, K4 pattern): on a ≥100 GB held corpus, model+coder must beat solid+omni including model bytes at a decode throughput a cold tier can live with — else it stays research-front. This is the same claim E1 tested, at the scale where the math says it can finally win; nothing about the kill is relitigated below its threshold.

4. Corpus gap (checked, as asked)

The local arXiv corpus (arxiv_query.js; AI/ML abstracts 2025-07+, control-eng + survivorship tranches) has no compression tranche — only two adjacent hits (diffusion-based lossless transmission arXiv:2606.06273; information-theoretic tokenization arXiv:2512.16975). Action row: harvest a cs.IT/compression tranche via scripts/arxiv_harvest.py so the next frontier pass grounds locally; until then AITDCC-2026 (Q2) remains the external instrument.

5. Falsification table

6. Honest scope

• No entropy claims: generative members are description-length wins on recipe-bearing

data; F3 is amortization arithmetic, not new coding theory.

• "Universe/multiverse/tesseract" are labeled metaphors over one shipped mechanism:

deterministic generators + verified materialization + (next) sliced observation.

• All tiers are opt-in extensions of the one canonical module (no new subsystem); LoC-scale

is an ambition benchmark for F2/F3, not a roadmap commitment.