diff --git a/docs/brainstorms/mathematical-foundations.md b/docs/brainstorms/mathematical-foundations.md
new file mode 100644
index 0000000..56e133b
--- /dev/null
+++ b/docs/brainstorms/mathematical-foundations.md
@@ -0,0 +1,121 @@
+# mathematical-foundations
+
+## 2026-08-09T04:51:00Z
+
+```capsule
+topic: mathematical-foundations
+date: 2026-08-09
+
+seed (owner, paraphrase): the project's math is extensive — query algebras,
+Laplacians, curvature, sets and memberships, type correctness. What axioms do
+they rely on; can we "derive from first principles"? Would knowing that empower
+us — is foundational math the bridge to other branches, a mini Langlands? When
+we capture an idea it behaves like a theorem (or a proposed definition) — how
+far does the rabbit hole of its implications propagate? Is this where a
+proof-based language comes in?
+
+decisions:
+  - framing: the palace's math is finite in substance. The Laplacian is a
+    concrete matrix; Ollivier-Ricci curvature on a finite graph is a finite
+    linear program (W1 transport); memberships form a finite Boolean algebra;
+    embeddings live in finite-dimensional inner-product spaces over R. The
+    foundational reliance sits far below ZFC — classical logic + finite sets +
+    real arithmetic. Axiomatic consistency is not where correctness risk lives.
+  - framing: the working bridge to other mathematical machinery is
+    structure-instantiation, not axiomatic descent. Name which structure each
+    component is a model of (Boolean algebra, semiring, PSD operator,
+    metric-measure space) and that structure's theorems transfer for free. The
+    Langlands analogy lands as correspondences BETWEEN structures (graph <->
+    operator, curvature <-> transport, membership <-> logic), not as shared
+    axioms at the bottom.
+  - framing: falsifiers before proofs. Algebraic laws land as property tests
+    first — mechanized falsification matches the house epistemology (ratify
+    falsifiers, not proofs). A proof assistant (Lean 4 + Mathlib) proves the
+    math, never the Python; the translation gap means it belongs at
+    design-note level, if anywhere.
+  - the machine betrays the axioms: float addition is not associative, so the
+    exact laws we rely on are the ones that must be tested with tolerance
+    bounds. The epsilon-gap between R and float64 is the real foundational
+    risk, not Russell's paradox.
+
+parked:
+  - decision: adopting a proof language (Lean 4) for design-note-level math
+    default: no proof assistant; laws live as property tests in the suite
+    re_entry: a design note whose central claim a property test cannot falsify
+      (e.g. a convergence or spectral bound), or a wrong-math finding that a
+      machine-checked proof would have caught
+
+open_questions:
+  - which structure does each component actually instantiate? query algebra —
+    lattice, monoid, or semiring? memberships — finite Boolean algebra (finite
+    Stone: every finite BA is a powerset algebra)? Laplacian — PSD operator,
+    zero row sums? curvature — W1 on a finite metric-measure graph? Naming
+    these precisely is the laws-ledger question.
+  - how far does a captured definition's deductive closure propagate — should
+    implication-tracking (what a ratified definition forces elsewhere) be an
+    explicit artifact-chain mechanism, or is the gate discipline (findings
+    re-enter only through the gate) already the control on propagation?
+  - does semiring provenance (one algebra, many query semantics by swapping
+    the semiring) fit the query algebra as prior art? [FROM MEMORY — verify
+    before relying on it]
+  - is mypy + type_gate already the Curry-Howard layer (types as propositions,
+    the checker as a weak proof assistant), and how far can refinement-style
+    newtypes push it before diminishing returns?
+
+next_steps:
+  - candidate: a laws-ledger sweep — per mathematical component, the claimed
+    structure + its laws + one named falsifier (property test) per law;
+    memberships (in flight) is the cheapest first target — Boolean-algebra
+    laws are nearly free to test
+  - if the ledger lands, fold the structure-claim into the build-plan §8 math
+    field-guide so every new component names what it instantiates at mint time
+
+references:
+  - core/stores/memberships.py + tests/unit/test_memberships.py (staged, in
+    flight — the first candidate surface)
+  - dn-vector-membership-store, dn-core-graph-instruments (the structures
+    under discussion)
+  - Curry-Howard correspondence; Lean 4 + Mathlib [FROM MEMORY — verify
+    Mathlib coverage before citing]
+  - Green, Karvounarakis, Tannen — "Provenance Semirings" (PODS 2007) [FROM
+    MEMORY — verify before any book-grade citation]
+```
+
+## 2026-08-09T05:12:00Z
+
+```capsule
+topic: mathematical-foundations
+date: 2026-08-09
+
+seed (owner, paraphrase): the capture-as-theorem framing prompts a second move —
+should we ADOPT research-community terminology and formats outright? Refresh the
+templates and skills so framing and terminology reflect the research community.
+Three motives, in the owner's order of weight: (1) personal — a way to learn
+mathematical/CS research language properly; (2) credibility; (3) most important,
+accessibility — it makes the project legible to technical research communities.
+And the deskcheck could BE the research artifact: a LaTeX PDF formatted like a
+journal/conference paper.
+
+decisions:
+  - graduate immediately: owner directed the full design treatment this session
+    (ultracode) — a design note as the artifact, fully audited, PR-ready, with
+    issues raised alongside. This capsule is the warrant trail; the substance
+    lives in the note (dn PR to follow, cites this file).
+
+open_questions:
+  - carried by the design note rather than duplicated here (claim ladder, where
+    the lexicon lives, deskcheck-paper mechanics, the still-unwritten owner-only
+    research Question that a research idiom makes conspicuous).
+
+next_steps:
+  - land the design-note PR; file residual questions/risks as GitHub issues per
+    the issue skill.
+
+references:
+  - the 2026-08-09T04:51Z capsule above (the first seed: axioms, bridges,
+    falsifiers-before-proofs)
+  - docs/templates/deskcheck.md (the artifact a paper format would evolve —
+    same dc- lifecycle, never a parallel ritual)
+  - docs/design-notes/track-board-and-deskcheck-gate.md (the ratified gate the
+    evolution must stay coherent with)
+```