SZL-Nemo RUNTIME-BOUND · NOT FINE-TUNED

SZL-Nemo is an identity-bound local recipe and governed orchestration runtime. Built on NVIDIA Nemotron 3 Nano 4B; the weights are NVIDIA's and are not SZL-fine-tuned. The differentiator is an auditable domain router: a query is routed to domain-expert heads by a Λ-governed router (Conjecture 1, advisory floor < 1.0) that reuses the active-flux router crossover and the RouteLLM Thompson posteriors — and every expert selection emits a signed DSSE receipt.

Honest by design. We did NOT train a foundation model from scratch — there is no 550B SZL model and no local Nemotron-Ultra (cloud tier only). OUR contribution is the governance layer: governed-MoE domain-expert routing, MTP / speculative-decode default, Reflexion + Voyager + τ-bench self-improvement experiments, exact-tag local serving, and tamper-evident signed receipts. sovereign:true only with a live per-GPU gpu_reachable probe; the cloud tier is always sovereign:false. Every number is MEASURED (live) or ROADMAP — never fabricated. Λ = Conjecture 1 (advisory, never a pass/fail oracle); trust < 100%; locked-8 @ c7c0ba17; 0 runtime JS CDN; effectors SIMULATED human-on-loop. Fonts self-hosted (0 runtime CDN of any kind).

CROSS-LINK: τ-bench & calibration J/token & sovereign posture Reflexion / Voyager loop Active-flux crossover

Base (open, cited)

—

τ-bench (MEASURED-by-SZL)

—

MTP speedup (S)

—

Sovereign-local

—

live gpu_reachable probe

Governed-MoE Domain-Expert Router the differentiator · signed every selection

"Experts" are domain heads (counter-uas / maritime / governance / code / finance) — an auditable MoE, not learned FFN experts. Routing fuses the Λ governance score (Conjecture 1) with the active-flux serving crossover (small/local ⇄ large/cloud) and the RouteLLM Thompson posteriors. Try a query:

Λ —

SIGNED DSSE RECEIPT (route decision)

—

verify—

a11oy Restraint code-frugality ladder · gates the code path before emit · signed

Before SZL-Nemo emits ANY code, the intended diff is routed through the governed, Ponytail-derived 6-rung frugality ladder (YAGNI → stdlib → native → installed-deps → one-liner → minimal). The chosen rung, the restraint: ceiling, a lines-saved estimate and a signed restraint receipt are attached to the Nemo response. Adopted from Ponytail (MIT), governed + Λ-scored here. R1 owns the ladder module; if it isn't live yet this reads honest PENDING (no rung or number fabricated).

rung — —

— (ask SZL-Nemo for code to see the restraint verdict)

SIGNED RESTRAINT RECEIPT (nested in the Nemo code receipt)

—

verify—

3D Router Dashboard GPU blocks · particle token-flows · signed MoE routing pulses · live J/token

Holographic view of the sovereign-inference fleet: each serving tier is a GPU block (color = where/sovereign, honest gpu_reachable probe), animated particle token-flows stream from the router to the selected domain-expert heads, and a signed routing pulse lights the chosen edge when a route receipt is signed. The J/token counter is MEASURED only on a live GPU power probe; with no meter wired it stays an honest ROADMAP estimate (never presented as measured). Renders on the shared 0-CDN WebGL2 kit; CPU/old-GPU falls back to a 2D canvas; the panels above are the full non-3D experience. Patterns: vLLM/SGLang, NVIDIA Dynamo, LiteLLM, RouteLLM.

J/token — tokens routed 0 signed pulses 0 Λ advisory —

3D is off (default). Click Enable 3D to render the holographic router on the live /nemo/tiers + /nemo/route + /energy/budget endpoints. The 2D dashboard above is always available as the fallback.

0 runtime CDN · WebGL2 baseline + honest 2D fallback · J/token MEASURED-or-ROADMAP (never fabricated) · sovereign:true only on live probe · Λ = Conjecture 1 (<1.0) · trust <100%.

MTP / Speculative Decoding inference default

Speedup S = (k+1) / (k(1−α)+1) (Leviathan et al. 2022). Draft model serves k tokens, target accepts at rate α. This remains ROADMAP and is not enabled on the verified Ollama path.

—

Self-Improvement Loop Reflexion + Voyager + τ-bench

Runs the REAL τ-bench (Dev B) for a deliberately weaker baseline, adopts rule-following (Reflexion), admits a Voyager skill, re-runs, and signs the measured delta. Score history lives inside receipts.

—

SIGNED DELTA RECEIPT

— (run an iteration)

Serving Tiers honest where / sovereign labels

sovereign:true ONLY with a live per-GPU gpu_reachable probe (Dev C). The cloud NIM frontier tier (Nemotron Ultra) is always sovereign:false — it needs ~768GB VRAM and cannot run on the 2 GPUs.

—

τ-bench — MEASURED-by-SZL Dev B real suite

SZL τ-bench-style tool-rule-following suite with negative controls; an always-pass agent scores 0, proving non-triviality. NOT the upstream leaderboard.

—

Model Card base cited · what's ours

—