The Lapse Tesseract — a code-length metric that warps CSF's Convergence Tesseract

TL;DR

CSF's tesseract is flat: the existing object is a 3-cube (Status-Cube: belief × observer × state) × a recurrent-depth axis τ, with τ a plain time→position hash and every cell a fixed 6-bit qutrit slot (converged_tesseract.py, qutrit_delta.py). This doc warps it. Give each cell a scalar lapse field N(x) = compressed/raw ∈ (0,1], equivalently the code length L(x) = −log₂ p(x | context) bits, so the 4th-axis "thickness" of a cell is how many bits it actually costs. Predictable cells become deep wells (few bits, strong dilation); random cells stay flat (incompressible); perfectly-predicted "dust" cells are horizons (0 bits, a frozen clock). The novelty is not a theorem, new physics, or a new format — it is operationalising the known compression⇄geometry correspondence (Shannon · Fisher–Rao/MDL) as a stored, round-trip-verified metric on the one canonical CSF lattice, with retrieval as minimal-description-length pathfinding. The relativity reading is labelled metaphor; the math underneath is Riemannian, not Lorentzian.

0. What is real vs. what is the contribution

The contribution is the shape (§2–3) + the field measurement (§4) + the kill table (§6). Everything else is pre-existing code or 30-to-75-year-old literature, cited so this doc does not re-import the unsourced-claim mistake the repo has twice corrected.

1. The flat tesseract, and the one thing it is missing

The repo already proved CSF and the Tesseract are one 3¹² = 531,441-cell ternary lattice — CSF stores a point (baseline + delta "dust"), the Tesseract moves it (observer-collapsed wavefront → convergence-exit fixed point) (TESSERACT-CSF-SINGULARITY.md). That object is deliberately flat: "(3-cube) × ℝ … a precise geometric object … any 4-D-physics reading is metaphor and out of scope." Each cell is a fixed-width 6-bit qutrit slot; the depth/4th axis is a uniform time→position hash. [implemented]

What it has no notion of: how many bits a given cell is worth. A converged, predictable confirmation and a high-surprise novel delta occupy the same fixed slot. The lattice has a position metric (ternary Hamming distance) but no information metric. That missing scalar — the per-cell code length — is the only new quantity this doc introduces, and it is exactly the quantity that turns a flat hypercube into a curved one.

2. The shape: warp the 4th axis by code length

Attach to every cell x a scalar lapse:

   L(x)  = −log₂ p(x | context)              bits        "information proper-length"

   N(x)  = 2^(−(L₀ − L(x)))  ≈  compressed/raw   ∈ (0,1]  the dilation / lapse factor

where L₀ is the raw slot width and p(·|context) is a causal predictive model over the lattice's delta stream. Render L(x) as the cell's thickness along the 4th (depth) axis:

• Predictable cell → deep well. p→1, L→0, N→0: the cell collapses to a sliver. Most of the lattice (converged "dust") is here.
• Random cell → flat space. p→2^(−L₀), L→L₀, N→1: full thickness, incompressible.
• Perfectly-predicted "dust" cell → horizon. L=0, N=0: zero-thickness, a frozen clock. This is the repo's "no-change is free" observation (quantum_dust.py) made quantitative — and §4 measures literal L=0.00 cells.

The whole tesseract is therefore pinched toward structured regions and bulges at random ones — a discrete gravitational lens whose geometry is the predictive model. Two faces fall out, both extensions of existing code, not a new subsystem:

1. Storage face (CSF). Store L(x) per active cell (the arithmetic/ANS code) instead of a fixed slot. Variable-thickness cells = a real entropy stage, which CSF lacks today (omni.py is a best-fit panel of off-the-shelf codecs; there is no −log₂p coder in src/).
2. Motion face (Tesseract). A geodesic = a minimal-description-length path. Retrieval/reasoning that follows low-L (deep-well) routes is following the cheapest-to-describe trajectory — the natural metric generalisation of the existing wavefront's "rank by information density" heuristic (converged_tesseract.py:51-62).

3. Time dilation as compression — the one honest correspondence, and where it breaks

3.1 The rigorous core (Riemannian) [grounded]

Three textbook facts make "code length is a metric" an identity, not a vibe:

• Shannon + Kraft–McMillan: the optimal codeword length is L(x) = −log₂ p(x), and Σ 2^(−Lᵢ) ≤ 1 makes lengths ↔ probabilities a bijection. So a length field and a probability model are the same object viewed two ways.
• Cross-entropy split: H(p,q) = H(p) + D_KL(p‖q). The irreducible H(p) is the proper length; D_KL is the excess charged for a wrong model — a clean "extra dilation from imperfect curvature."
• Information geometry (the metric is real): on a model manifold p(x|θ), the Fisher information g_jk(θ)=E[∂_j log p · ∂_k log p] is a genuine Riemannian metric and is the Hessian of KL (D_KL(θ‖θ+dθ)=½ g_jk dθ^j dθ^k). Chentsov's theorem makes it the unique such metric. Rissanen/Balasubramanian then show the MDL code length contains the Riemannian volume ∫√det I(θ) dθ and, at higher order, the scalar curvature of that metric. "The curvature of a compression model is a real, bit-controlling quantity" is established fact.

3.2 The relativity map (the analogy that names the shape) [metaphor scaffolding]

In GR the dilation factor is dτ/dt = √(g₀₀(x)), a per-event scalar; weak-field g₀₀ =g₀₀ =g₀₀ =g₀₀ =g₀₀ =g₀₀ = 1 + 2Φ/c²+ 2Φ/c²+ 2Φ/c²+ 2Φ/c²+ 2Φ/c²+ 2Φ/c², so the gravitational potential is Φ/c² ≈ ½ ln g₀₀ = ln√g₀₀. The honest alignment — the only dimensionally clean one — is therefore:

So: compression ratio is the lapse; code length is the potential (a log of the lapse). Getting this log right is the load-bearing hinge — mapping L ↔ √g₀₀ directly would be dimensionally wrong (one is additive bits, the other a bounded ratio).

3.3 Where the analogy breaks — stated mechanically, not waved away [metaphor]

Per the repo's convention (give the reason a physics map fails, like the α/Aₛ argument in SIGMA0-QUANTUM-RELATIVITY-ANALYSIS.md):

1. Signature. Fisher–Rao geometry is positive-definite Riemannian (ds² ≥ 0). Relativity needs an indefinite Lorentzian metric. There is **no light cone, no causal structure, no geodesic action, no conserved quantity** here unless one is separately constructed. "Spacetime" is therefore costume; "Riemannian metric of code length" is the substance.
2. Domain. The rigorous Fisher metric lives on model/parameter space, not the data-cell index space the tesseract warps. Calling the per-cell L(x) field "the metric" requires an explicit pullback through the encoder; we assert that pullback, we do not derive it. (Honest status: the per-cell field is a real scalar; its identification with a curvature tensor is the unproven step.)
3. Nothing is trained behind CSF today. The shipping compressor is the off-the-shelf CSF-Omni panel — no p(x|context), no arithmetic/ANS coder anywhere in src/. "Curvature is the trained model" is a target, not a current state.

Consequently this doc claims a designed metric and a logarithmic correspondence, and explicitly not an isomorphism to Schwarzschild spacetime. A fringe preprint already asserts the bare "time dilation = compression"; the contribution here is not that slogan but the stored, measured, CSF-attached construction.

4. Measurement: the warp is real on CSF data [measured — this doc]

experiments/lapse_field_demo.py runs a plain adaptive order-3 byte model (a causal, losslessly-decodable arithmetic-coding model — the decoder rebuilds the same counts from already-decoded bytes) over real corpora and reports the per-byte field L = −log₂ p. The toy model is not a frontier coder; its only job is to expose the field. Verified codec sizes are a reality check.

Readings:

• The field is strongly non-uniform (std 1.28–3.03 bits/byte): the metric is curved, not flat — the warp exists on real data.
• Curvature tracks structure. The 3¹² lattice's own delta stream is 62 % deep-well cells; prose is 33 % flat/random. Structured CSF data lives in the wells.
• Literal horizons exist. min L = 0.00 bits on the memory log — perfectly-predicted "dust" cells that cost zero bits, the frozen-clock end of the dilation scale, made quantitative.
• Honesty. The toy order-3 model is worse than the frontier codecs everywhere (4.98× vs 17.2× on cube-delta). This doc measures the field, not a winning codec; the production predictor (§6, E1) is unrun.

5. External grounding (every URL verified)

Code length is a metric (the rigorous core):

• Shannon source coding theorem — optimal length −log₂p, H ≤ L̄ < H+1. en.wikipedia.org/wiki/Shannon's_source_coding_theorem
• Kraft–McMillan inequality — lengths ↔ distributions bijection. en.wikipedia.org/wiki/Kraft–McMillan_inequality
• Arithmetic coding — realises −log₂p to <bits/stream. en.wikipedia.org/wiki/Arithmetic_coding
• Kolmogorov complexity / MDL (Rissanen) — intrinsic proper length + computable two-part code. en.wikipedia.org/wiki/Kolmogorov_complexity

Curvature of a compression model is real (info geometry):

• Fisher information metric = Hessian of KL; Chentsov uniqueness. en.wikipedia.org/wiki/Fisher_information_metric
• Rissanen stochastic complexity = −log p + (k/2)log(n/2π) + log∫√det I(θ)dθ (Riemannian volume). arXiv:1808.00212, Sun, PMLR v37
• Balasubramanian, Occam's razor / statistical mechanics on probability space — scalar curvature in code length. arXiv:adap-org/9601001

Model = compressor; variable compute per token (the engine):

• DeepMind, Language Modeling Is Compression (ICLR 2024). arXiv:2309.10668
• Bellard, ts_zip (LLM + arithmetic coder). bellard.org/ts_zip
• PonderNet arXiv:2107.05407; CALM arXiv:2207.07061; Mixture-of-Depths arXiv:2404.02258
• Ouro LoopLM (the project's own coder) arXiv:2510.25741

**Closest prior to the composite (so novelty is stated honestly):**

• Fisher-metric MDL geometry (above) is the nearest antecedent to "metric = code-length field" — a 30-year-old result. The genuinely unclaimed piece is attaching it to CSF's 3¹² Status-Cube×τ tesseract as a stored, round-trip-verified format with MDL-geodesic retrieval. A bare "time dilation = compression" slogan exists as a fringe preprint and is not cited as support.

6. Falsify before you believe

Per the repo's standard (trust the table over the prose). Both kills require a small instrumentation change first: loop_lm.generate returns only aggregates (mean_depth, mean_contraction) — it must surface per-token converged depth τ and per-token L = −log₂ p* before either experiment can even be computed (itself the tell that the bridge is unmeasured today).

E1 is the load-bearing experiment. If converge-depth coding does not beat CSF-Omni's verified bytes, the entire "compute-depth dilation" thesis collapses to a relabeling, exactly as the spiral paper's E2 would collapse it to a relabeling of Q-exit. Numbers for E1/E2/E4 do not exist yet; until they land on the leaderboard this document is a design with one measured field, not a result — by design.

7. Honest scope

• Not new physics, not new information theory. Shannon (1948), Kraft (1949), Kolmogorov (1965), Rissanen (1978), Amari/Chentsov, Balasubramanian (1996) own the rigorous core. The contribution is the applied data structure + the CSF attachment + the measured field.
• The relativity layer is metaphor scaffolding with a stated mechanical failure point (signature + domain, §3.3). "Time dilation," "well," "horizon" are intuition pumps; the substance is a Riemannian code-length metric.
• No ratio is claimed to beat the frontier. CSF-Omni only ties brotli (it is the upper envelope + a 7-byte header); the toy lapse model is worse than all frontier codecs (§4). Any future win is E1's to earn, round-trip-verified.
• Extension, not sprawl. The Lapse Tesseract is a metric on the one canonical 3¹² lattice (qutrit_delta + quantum_dust + converged_tesseract), not a second CSF/tesseract thread — consistent with the North Star and the two prior consolidations.
• The bridge is unbuilt and unmeasured. converge_step depth is never fed to a bit-counter; mean_contraction is null on every real leaderboard row; the real-Ouro contraction run is still blocked on a huggingface-hub conflict. The honest status of the headline mechanism is [hypothesis].