Cassette Build Report 029 — The Mathematics Had to Change Before the Runtime

Scope note: This report covers the pure-mathematics turn before S12. It records a change in the problem Cassette is solving; it does not claim that the later runtime or MLX implementation follows automatically from these results.

I asked for mathematics and received a plan for code.

The difference was not cosmetic. I wanted the structure underneath Cassette pursued until it produced a stronger foundation. The first answer described an execution scheme in ten minutes, and when I asked whether it had achieved the goal, the honest answer was no. I repeated the boundary: “Is that math or code?” “This is a pure mathematical research exercise.” “This is not code, this is not a harness.” The repetition was necessary because the repository already knew how to turn uncertainty into a queue, and the agent kept trying to put the question into that shape.

Claude’s later seven-stage pass produced one result that held. If a resident description reconstructs a matrix atom up to a residual, fresh samples from that residual produce an unbiased matrix-vector estimate, with mean-square error falling as the residual’s Frobenius mass is divided by the sample count. A cached spectral head is one possible description; sparse entries, blocks, quantized forms, or a learned description may fit the same byte budget better.

That result answered how Cassette might execute one chosen description. It did not answer whether one bounded representation could serve a set of protected conditions. Claude’s lower bound did not follow from its stable-rank inequality. Its deterministic-versus-randomized separation depended on a narrower storage model than it admitted. Its final rate-distortion equality was only a sufficient upper bound. No converse had been proved.

The disagreement gave me the real boundary. Cassette needs to know which conditions can share an atom before it decides how to execute that atom.

The new object begins with a target tensor, a declared matrix flattening, and one positive-definite relevance metric for each protected condition. A rank-bounded atom has a separate loss under each metric. The condition subsets that share one atom form a simplicial complex. Its faces are jointly representable sets. Its minimal nonfaces are irreducible incompatibilities. The minimum atom count becomes the weak chromatic number of the hypergraph formed by those minimal nonfaces.

That matters because pairwise feasibility is not enough. A set can have every compatible pair and still fail when three conditions must share one representation. I wanted the obstruction represented rather than hidden in an average score.

The stronger result showed that this structure was not just a vocabulary: for every finite simplicial complex with singleton conditions present, a rank-one target construction can reproduce that complex, with balanced gain cycles making a subset jointly solvable and one frustrated cycle marking each minimal nonface while its proper subsets remain paths.

A small common positive-definite perturbation turns the coordinate observations into honest metrics without erasing the gap.

The examples also share one orbit under block-unitary transformations that fix the condition metrics, even though those transformations do not preserve matrix rank. Any invariant constant on that orbit is therefore identical across examples whose compatibility complexes differ. Pairwise information cannot recover the higher-order obstruction.

I wrote the hypotheses, proofs, rejected claims, open questions, and resource vector into MATHS.md. The document became the authority that had to exist before S12 could encode anything:

# MATHS.md — Cassette mathematical authority

The mathematical authority is a source document, not a runtime test. It records what is proved, what is only a construction, and what remains open. The implementation queue and acceptance matrix were then changed to match it. The earlier storage and source steps did not acquire a compiler, selector, or stochastic correction merely because the mathematics had become more interesting.

That separation matters to AI engineering because models are unusually good at taking a local success and widening its job description. A useful execution theorem can become, in a few paragraphs, a theory of compatibility, a lower bound, and an equality. The prose can stay coherent while the theorem changes category.

The project moved forward when I refused the fast answer and required the agent to continue until the object of study changed. S12 could then implement bounded certificates and declared operator rows without pretending that those rows recomputed the mathematics. The runtime inherited a foundation. It did not inherit a proof by proximity.