ADR-0034: MoE for the Σ₀ core — adopted as soon as feasible

Status

Accepted (Alex Place, in-session, 2026-07-24: "we are using MoE — as soon as is feasible", followed by "approve adrand start the gate tooling"). This ADR records that decision and defines feasible precisely, so the commitment is real and the safety story stays honest. It does not enable an MoE core today; gate tooling construction started the same day (src/sigma0/switched_gate.py).

Context

Every prior 2026-07-23 document held MoE OUT of v1 behind a named admission gate:

• The design of record (§3.1) defers MoE because a routed loop is a **switched dynamical

system**, and the Collapse Certificate's Part I stability results (the ρ(J)<1 machinery, the JSRR gate's theory basis) are proven for a FIXED update map only — cert §1.2.2 explicitly voids them for routed/MoE loops. Quoting §1 numbers for a routed core is forbidden.

• The serving memo rejected Option F (MoE-lite) as an "uncertified switched system."
• A 2026-07-23 grep confirmed the switched-system certification tooling does not exist.

The operator has now decided the destination: MoE is the capability path (sparse capacity is how a ≤4GB-footprint model punches above its active-parameter class). The gate does not disappear — it becomes the critical path.

Decision

1. The Σ₀ core adopts a mixture-of-experts architecture as soon as it is feasible.
2. Feasible means the MoE admission gate exists and passes — nothing else unlocks it:
   • (a) Per-expert contraction receipts. JSRR-style acceptance per active expert

composition: every expert path the router can select carries its own measured ρ with the same verdict machinery the dense loop has today.

• (b) Dwell-time certification. Switched-system stability needs more than stable

pieces: either a common Lyapunov certificate across experts, or a measured minimum dwell time between router switches meeting the average-dwell-time condition. The serve path must MONITOR dwell time and reject generations that switch faster than the certified bound.

• (c) Router canary. Router-entropy / expert-churn added as canary axes (a router that

thrashes is the switched analog of the collapse the NIS canary watches for).

• (d) External verification unchanged. Held-out exec verification, the M1 gate, Σ_θ

promotion gating, and the ≥60% anchor mandate apply to an MoE core exactly as to dense.

1. Until (a)–(c) exist and pass, the dense ≤3B looped core remains the product, and Part-I

stability numbers may not be quoted for any routed configuration (cert §1.2.2 stands).

1. The gate tooling moves to the FRONT of the parallel R lane (design of record §9): it is

now the critical path to an operator-decided destination, not a deferred curiosity.

Consequences

• The envelope is unchanged: ≤4GB footprint, CPU-viable; MoE spends the budget on sparse

capacity (total params may exceed 3B if the ACTIVE set and memory stay in the envelope — the envelope binds on footprint, which is what the operator's size rule protects).

• The depth-cascade training design (research/2026-07-24-depth-cascade-training-design.md)

composes: its contraction objective must hold per active-expert composition, and its monotone-depth kill test carries over unchanged.

• Honest sequencing: P0 (JSRR on the default dense path) still comes first — the dense gate

is a strict prerequisite skill for the switched one.

• Risk on record: switched-system certification at this scale is research, not engineering;

if the dwell-time story cannot be made measurable, feasibility is never reached and this ADR's commitment stays unexecuted rather than executed unsafely.

Evidence