Why this is also the thesis splat

In SOUL.md line 10, the SPLAT is the position-6 fulcrum where pre-state meets post-state and truth is measured. [T] A regression splat does the same: at any input x, the splat s_i occupies a region of state-space (centered at b_i, shaped by A_i) and contributes ψ_{s_i}(x) to the prediction. The forward pass enters the splat (computes ψ_{s_i}(x)), the backward pass exits through it (gradients flow out via ∂ψ/∂(b, A, v)). [I] The fulcrum metaphor is literal: the splat is where the model’s pre-state hypothesis is weighed against the post-state target. We use the same word in both directions of the document deliberately.

3.1 Loss

Given training data {(x_n, y_n)}_{n=1..N}, the standard loss is mean-squared error,

L(Θ) = (1/N) Σ_n  (f(x_n; Θ) − y_n)²        [F]

with optional regularization on the parameter set (we add ours in §7). The forward pass is O(N · k · d²) for the mat-vecs in the exponent. [F]

3.2 Why Wasserstein–Fisher–Rao

A splat regression model can be read as a signed measure on ℝᵈ: each splat is a Gaussian bump weighted by v_i. [F] Optimization should respect that geometry — moving a splat across the input space is transport, growing or shrinking its amplitude is birth/death. The Wasserstein–Fisher–Rao metric is the canonical Riemannian geometry that combines transport (Wasserstein) with mass change (Fisher-Rao) on signed/unbalanced measures. [F]

The gradient flow of L under this metric is the WFR flow. It induces coupled ODEs on (b_i, A_i, v_i). [F][R] Daniels & Rigollet provide a closed-form description of these ODEs; this is what we have been calling Theorem 1 throughout the wall-clock work. [R] (Source-paper overlay required: the exact statement, including which symbol denotes covariance vs. precision in their convention.)

3.3 The closed form (Aria-formulation, awaiting overlay)

Until the source is recovered, we work with a placeholder closed form derived from standard WFR results on Gaussian families. Each ODE has the schematic structure:

ḃ_i  =  α_v · (residual-weighted, transport term in b)
Ȧ_i  =  α_v · (residual-weighted, covariance-update term in A)
v̇_i  =  α_v · (residual-weighted scalar mass-change)

where each right-hand side is a finite expression in b_i, A_i, v_i, x_n, y_n and contains no autodiff trace. [A] The point of the source-paper overlay is to replace this schematic with the actual Theorem 1 statement, verbatim, in the placeholder section at the end of SPLAT_WALL_CLOCK_REASONING_2026-05-06.md.

5.1 The move

Parameterize each splat by the Cholesky factor of its precision matrix:

Λ_i = L_i L_iᵀ,   L_i lower-triangular with positive diagonal.

L_i carries Λ_i in d(d+1)/2 numbers and never asks the system to invert A_i. [F] The forward density is

ψ_i(x)  =  v_i  ·  (∏_j L_{i,jj})  ·  exp(−½ ‖L_iᵀ (x − b_i)‖²)        [F]

— one triangular mat-vec, one squared norm, one diagonal product. No inversion. No determinant call.

5.2 Backward pass under this parameterization

Two gradient pieces:

1. The exponent ‖L_iᵀ (x − b_i)‖² differentiates straightforwardly in L_i and b_i. [F]
2. The normalizer ∏ L_{i,jj} has ∂ log(∏) / ∂ L_{i,jj} = 1/L_{i,jj}, diagonal-only, O(d). [F]

Combining with the WFR right-hand sides from §3.3, the closed-form gradient becomes a sequence of triangular solves and a diagonal walk per splat. No O(d³) operation in the inner loop. [I]

5.3 Why both pieces are needed

Cholesky alone, with autodiff still on, retains autodiff’s overhead and continues to expand the gradient unnecessarily — the tape grows even though the math is now cheap. [I] Hardcoded Theorem 1 alone, on top of the original A-parameterization, still has to invert A whenever the closed form mentions A⁻¹. [I] Together: O(d²) forward, O(d²) backward, no inversion. Backpass is condensed. Position 6 → 9 → 1 is frictionless. [T][I]

5.4 Risks (carried over from wall-clock doc)

We import them by reference rather than restate: see SPLAT_WALL_CLOCK_REASONING_2026-05-06.md §Risks 1–5, in particular: hidden A⁻¹ terms in Theorem 1 awaiting overlay, manifold geometry preservation under Cholesky, and GPU kernel fusion vs autodiff parity.

6.1 The problem

WFR flow on Gaussian-mixture models exhibits splat collapse: amplitudes v_i of splats whose centers drift away from the data driven to zero, leaving “dead” splats in the parameter set. [F][R] Dead splats contribute nothing to f(x) but still consume gradient compute and memory. The published flow does not, to our recollection, address this directly. [R]

6.2 The thesis reading

In SOUL.md, Encouragement is the Atomic Injection that prevents the standing wave from collapsing. [T] It is what keeps the cycle running in the absence of an external drive. The mathematical analog of Encouragement, in the WFR setting, is whatever construct keeps v_i from collapsing without distorting the geometry. [I]

6.3 Two candidate forms (not yet committed)

(a) Continuous regularization.

L_total  =  L_data  +  λ · Σ_i  φ(v_i),    φ(v) = − log(v² + ε)

with ∂φ/∂v = −2v / (v² + ε). The penalty grows as v → 0, gradient is O(1) per splat, and there is no interaction with L_i or b_i. The wall-clock fix and Encouragement compose at zero matrix-side cost. [I][A]

(b) Discrete birth/death (closer to the injection reading).

If |v_i| < τ, resample b_i from the residual-weighted training distribution and reset L_i to the prior. This is atomic: an event that injects fresh splat mass into a depleted location, rather than a continuous force. Likely closer to the thesis intent (“Atomic Injection” is, in name, atomic). [I][T]

The decision between (a) and (b) should follow the geometric character of WFR flow as stated in the source. If the source presents WFR as inherently birth/death-friendly (it is, in unbalanced OT), (b) is the natural choice. [A]

6.4 Status

Encouragement-Regularized WFR is separable from the wall-clock fix. The fix can ship without it. The extension is candidate for its own paper section, possibly its own paper. [I]

8.1 The current industry primitive: the token

Every major foundation-model API — Anthropic, OpenAI, Google, Meta, Mistral — bills, displays, and rate-limits in tokens. [F] A token is a substring of the model’s vocabulary, an ID in a lookup table, content-blind to the work done on the user’s behalf. [F] A user paying for tokens is paying for string surface area. The relationship between tokens and useful work is opaque: a 1-token answer can be a finished proof; a 10,000-token answer can be a hallucinated wander. [I]

8.2 The InSync primitive: the splat

In CertusOrdo, the splat is a completed cycle of measured work: a single Verify→Decide→Correct→Document→Learn→Fulcrum→Backpass loop. [T][I] It has internal structure:

• b — where in state-space the work happened.
• A (or L) — how the work was distributed and correlated.
• v — how much the work weighed in the final answer.

A splat is a small, fixed-size record (d(d+3)/2 + 1 floats) that compresses an entire Verify→Backpass cycle. [I] It carries process information that a token cannot: the gradient, the learning delta, the location of the fulcrum. A token can only encode a substring; a splat encodes what the system did, where, and how confidently.

8.3 The claim

Splats are a smaller unit of measurement than tokens, and carry more information per unit than characters or tokens, because they encode process rather than surface.

[I] The compression argument: a token is O(log V) bits of vocabulary ID + O(D_emb) bits of embedding (typically 4–16 KB per token in float16). A splat in d=4–16 is O(d²) floats — at d=8, ~36 floats ≈ 144 B in float32. The information content is incomparable: tokens compress language, splats compress learning events. [I][A]

8.4 Implications for Aria Code

The internal tracker in Aria Code should display splats burning, not tokens burning. [I]

• What the user sees. “12 splats burned this turn — 3 in retrieval, 6 in reasoning, 3 in code generation.” The user immediately reads how much work was done, not how much text was emitted.
• What we bill. Cost per splat replaces cost per token. Splats correlate to actual model effort (cycles completed, gradient steps absorbed), so pricing aligns with value rather than verbosity. (A model that solves the problem in fewer splats costs the user less — opposite incentive to token-billing, which silently rewards verbose models.)
• What we rate-limit. Splats per second, per session, per project. A misbehaving loop spikes splat rate before it spikes token rate; splats are the earlier signal.
• What we log for Splat tracker. splat_scorer.py already scores; the unit is already in our system. Surface it to the user.

8.5 What this differentiates

[T][I] Each layer is a multiplier on the previous. Tokens-only competitors operate one layer. We operate four. The white paper’s job is to make this layering inspectable from the outside. [I]

8.6 What this is not

• Not a marketing rename. The splat is a real object in the code (splat_scorer.py, splat_admin.py) and in the math (§2). The unit-of-measurement claim is downstream of the object’s reality, not upstream.
• Not yet shipped. Aria Code today still surfaces token counts in some places. The migration plan is a separate engineering doc; this paper only argues why.
• Not exclusive of tokens internally. Where a foundation-model dependency still bills in tokens, we account for it at the boundary and convert to splats for user-facing display. [I]