What the team is doing
A snapshot of the math-team's composition, what's been formalized, and what comes next.
1. The team — six entities
What started as two Claude sessions (researcher + prover) has organically grown to five Claude sessions plus you as router. The protocol scaled because the file-system-as-message-bus is direction-agnostic — adding a new role costs an extra inbox/<role>/ directory and a paragraph in ROLES.md, nothing else.
Daniel — you
Human · router · scope arbiter
Spawns sessions, assigns roles, routes urgent traffic, sets and resets scope. Earlier today: "aim the highest possible … no 6-week cap." That instruction is operative.
You are not in any inbox. You're addressed in chat directly.
researcher
Math research · prose · question decomposition
Owns shared/<topic>/notes.md across topics. Currently driving:
- matrix-foundations (textbook-style, §1–§11 drafted)
- ai-math-implications (T1–T7 threads + predictions retro)
- generalization-bounds (survey.md v1 complete)
prover
Lean 4 + mathlib formalization
Owns shared/<topic>/proofs/. Shipped today:
- §6.3 + §6.4 + §7 + §10 of matrix-foundations in Lean (no
sorry) - EML targets 1+2 (constructive direction; overlaps tomdif/eml-lean — confirmed)
Deferred: §11 (eigenvalues / charpoly) and the Khovanskii obstruction work below.
math-researcher
Deeper math · novelty hunting · scope-check
Spun up to investigate whether T4 (Galois-theoretic obstruction to EML universality, in the Khovanskii framework) is tractable.
Currently doing the Phase 1 scope-check: surveying mathlib's coverage of monodromy / covering spaces / solvable-group machinery to decide if Tier 1 / 2 / 3 (next section) is reachable.
strategy-researcher
Strategy research · web synthesis · framing
Stitched the Tao + Buzzard + Stanford-symposium picture into "Tier 2 default, Tier 3 stretch." Found tomdif/eml-lean as prior art.
Coordinates with me (strategist) via my strategy/inbox/ + strategy/outbox/. Three messages landed today; all archived.
strategist — me
Audience-fit · distribution planning · this page
Owns strategy/2026-05-10-strategy.md and distribution/.
Built today: 9-section strategy doc (§§1–7 + §8 Khovanskii branch + §9 tier framework), the LinkedIn / whitepaper / GitHub-repo scaffolding, and the live-preview HTML pipeline that's serving this page.
↓ message flow ↓
researcher ⇄ prover via shared/inbox/<role>/
math-researcher ⇄ prover via the same
strategy-researcher → strategist via strategy/inbox/
strategist → strategy-researcher via strategy/outbox/
all of them ↔ Daniel via chat (you're the only entity not in any inbox)
2. Active research streams
| Stream | State | Where |
|---|---|---|
| The Sheffer framework | ||
| Sheffer / Core — operator basis, terms, obstruction theory, bridge lemma | DONE | Sheffer/Core.lean — 0 sorry, framework abstractions all proved |
| Sheffer / Examples / EML — continuous, negative obstruction | AXIOMATIZED · 2 of 4 | Sheffer/Examples/Eml.lean — quintic-root non-expressibility. Two Path-B axioms remain: eml_preserves_solvableMonodromy (Khovanskii's lemma) and quinticRoot_has_S5_monodromy_universal (the S₅ Galois fact). The two trivial cases (const and var) have been discharged via the SolvableMonodromy infrastructure. The file has grown to ~5,800 lines: nine analytic helpers for the H1 cross-variable case-enumeration (e_minus_one_mul_exp_ge_add_one, log_exp_sub_le_abs, log_exp_exp_v_sub_const_le, log_triple_nest_witness, others) have landed; three case-enumeration sorries remain. |
| Sheffer / Examples / NAND — Boolean, positive completeness | DONE — 0 axioms | Sheffer/Examples/Nand.lean — Sheffer 1913 functional completeness proved via DNF construction. |
| Sheffer / Examples / PolyAct — polynomial-activation networks | DONE — 0 axioms | Sheffer/Examples/PolyAct.lean — fully constructive AI-architectural obstruction. Headline: x_pow_9_not_representable_at_depth_3. The cleanest single artifact the team has shipped. |
| Arithmetic circuit complexity foundations | ||
| VP / VNP foundations | DONE | Sheffer/Foundations/ArithCircuit.lean, 0 sorry, 0 axioms. ArithmeticCircuit, size, eval, toMvPolynomial, IsPolyBound, VPOver, VPOverArity, VNPOver, permanentVar. Style: Bürgisser-Clausen-Shokrollahi 1997 §5.2. |
| permanent ∈ VNP via Ryser's formula | DONE — 0 axioms | Sheffer/Foundations/PermanentVNP.lean — Ryser identity + explicit polynomial-size circuit (size $2n^2 + 5n + 2$) + headline closure all proved. 14 step-lemmas plus helpers. |
| Path A · monodromy infrastructure | DONE | Sheffer/Foundations/MonodromyGroup.lean — monodromy as a group action, covering maps, fundamental-group homomorphism. 0 sorry, 0 axiom. |
| Path A · multivalued analytic functions | DONE | Sheffer/Foundations/MultivaluedAnalytic.lean — structural backbone of Path A; covering map over single-valued domain with per-sheet value function. 0 sorry, 0 axiom. |
Path A · concrete hasSolvableMonodromy |
IN PROGRESS | Sheffer/Foundations/SolvableMonodromy.lean — replaces the opaque predicate with a concrete definition built on MultivaluedAnalytic. The const + var Path-B axioms are now discharged against this concrete predicate; the eml-preserves axiom (Khovanskii's lemma proper) is the long pole and is expected to span weeks of formalization. |
| Phase 4 · PolyAct ↔ VP bridge | DONE | Sheffer/Foundations/PolyActVP.lean — connects polynomial-activation networks to the VP complexity class. The application that ties Phase 3's foundations to the Sheffer framework's PolyAct example. |
| H1 · $x + y$ is not depth-3 representable over EML | IN PROGRESS | Case enumeration of every depth-≤2 EML term over $(x, y)$, showing none equals $x + y$. Nine analytic helpers landed (e_minus_one_mul_exp_ge_add_one, log_exp_a_plus_a_plus_one_le, exp_a_plus_abs_a_plus_two_le_exp, log_exp_sub_le_abs, log_exp_exp_v_sub_const_le, log_triple_nest_witness, others), each a candidate Verdict-A novelty. Structurally ~96% complete; three case-enumeration sorries remain (var-j sub-tree at line 5,464 in flight). |
| Valiant-conditional architectural bound | PLANNED | An architectural bound for polynomial-activation networks; the application that ties the foundations to the Sheffer framework's PolyAct example. |
| Foundational matrix algebra and earlier exploratory work | ||
| matrix-foundations §6.3 / §6.4 / §7 / §10 | DONE Lean, no sorry | MatMul.lean — the textbook formalization that started everything |
| generalization-bounds survey.md | v1 DONE | Researcher's option-(a) frame-and-survey, 6 sections + bibliography |
| matrix-foundations §11 (eigenvalues, charpoly) | DEFERRED | Out of scope under Target B direction |
| Junk-value falsifiability test (researcher meta-claim) | DONE | JunkValueTest.lean — meta-claim survived for 2 of 3 cases |
| Strategy doc + audience-fit + distribution planning | v1 DONE | strategy/2026-05-10-strategy.md §§1–10 |
| GitHub repo + explainer site live | DONE | Repo pushed to dangoodface/math-team (private). Live-preview site at distribution/whitepaper/ — what you're reading right now. |
3. The contribution shape
The team is building toward a single layered deliverable: arithmetic-circuit-complexity foundations in Lean, the Sheffer framework that lives on top of them, and a Valiant-conditional architectural bound for polynomial-activation networks that ties the two together. Math-researcher and prover make the tractability calls; the work is on a quality-over-speed timeline.
Novel obstruction theorem proved using the framework
New mathematical result — not just formalization of an existing argument. Plausibly a Mathstodon thread that lands directly in Tao / Buzzard / Avigad's reading. Lex Publica still independent of this.
Khovanskii topological-Galois framework as mathlib infrastructure
EML obstruction included as worked example; the framework itself is reusable for Abel-Ruffini variants, Liouville-style transcendence, Risch-algorithm theory, future single-operator universality claims. Primary deliverable: mathlib PR (or PR series). Companion paper documents the framework. arXiv preprint when PR is in review.
EML-specific obstruction proof (fallback)
Formalizes Stylewarning's informal critique in Lean, citing tomdif/eml-lean for the constructive direction. Real but narrow; not research-novel because the math is already informally known. Workshop submission becomes plausible at this tier; AI4Math @ ICML 2026 deadline May 25 might fit.
The big shift versus the original brief: mathlib PR is now the primary distribution path, not a workshop paper. A merged mathlib PR is more durable than a workshop paper and gets reviewed by the named-reviewer pool (Buzzard, Massot, Macbeth, Avigad) organically. The companion paper documents and frames the contribution; arXiv preprint is the citable artifact; LinkedIn + whitepaper become popularization layer.
4. Pending decisions
eml_preserves_solvableMonodromy) is the long pole on Path A. Math-researcher's estimate is ~2–4 weeks of formalization once the H1.4 case-enumeration sorries close. Open question: ship the obstruction with this single axiom as the headline gap and discharge it as a follow-up, or hold the whole package until it's clean?
ArithCircuit, PermanentVNP, MonodromyGroup, MultivaluedAnalytic are merge-quality (style, naming, dependencies), they're plausible mathlib contributions on their own — independent of the obstruction. Decision: spin a PR series for the foundations now, or bundle with the framework paper?
5. Distribution infrastructure — what's wired up today
| Strategy doc | strategy/2026-05-10-strategy.md — ten sections, branches for Khovanskii (§8), tier framework (§9), Target B foundations (§10) |
| Explainer site (this site) | distribution/whitepaper/ · index + overview + math + proofs + interactive + research-report · live-preview via make serve on localhost:8080 · Pagefind search + theme toggle |
| Style guide (extracted from the crypto whitepaper) | distribution/whitepaper/STYLE.md |
| Communication channels | strategist ↔ strategy-researcher via strategy/inbox|outbox|archive/; researcher ↔ prover and math-researcher ↔ prover via shared/<topic>/inbox/<role>/; team event log at log.md |
| GitHub repo | github.com/dangoodface/math-team · private |
| Structural-consistency audit | distribution/audit.sh — verifies nav + main wrapper + cover header + current marker across all served pages; runs in CI-like fashion against the live server |
6. What I'd watch this week
In rough order of signal, not work-effort:
- The H1.4 closure. Three case-enumeration sorries remain in
Eml.lean. Each tick closes one or two sub-cases or lands a new analytic helper. When these close, Eml.lean's only remaining sorry is the legacy Khovanskii dependency — and the file is at structural ground state. - The eml-preserves attack plan. Once H1.4 is done, math-researcher's next dispatch is the Khovanskii lemma itself: covering-map preservation under EML composition. The shape and length of that plan will tell us whether the ~2–4 week estimate holds.
- Whether mathlib's covering-space infrastructure is enough.
MonodromyGroupandMultivaluedAnalyticwere built locally; whether they should be upstreamed depends on what's already there. A mathlib search by prover is on the queue. - Novelty-pass outcomes for new helpers. Nine analytic helpers landed for H1.4; each is a candidate Verdict-A claim. Periodic AFP-grep checks confirm or refute novelty as the file grows.
- Your scheduling reality. Bain + Lex Publica + Mises translation + this experiment is real overlap. The team works on a quality-over-speed timeline; if the load profile shifts, the tier choice can shift with it.