ARCHITECTURE — THE PIPELINE

Run until it breaks.
Then mint the seal.

Nucleus is the production pipeline that turns frontier-model reasoning into a permanent, signed Super Skill Model. Seven stages, four agents in adversarial loop, three certification gates, and one Ed25519 provenance chain. The pipeline never stops on a schedule — it stops when the swarm exhausts every failure mode it can find.

01 · CORPUS + MANIFEST02 · 7-LAYER EXTRACTION03 · SYMBOLIC REASONING04 · FINE-TUNING05 · ADVERSARIAL LOOP06 · CERTIFICATION07 · NUCLEUS SEALINGESTKICE + TICESCoTD + CoTDSFT + RAFTSWARM3-GATESEALrun until convergence ⟳pedagogical phase — v1.N continuesv1.0Click a phase to inspect it — or open its full breakdown.7-stage distillation pipeline. Adversarial loop runs until convergence. Three gates before any Seal is minted.

PHASE 05

SWARMAdversarial Loop

Run-until-it-breaks. Four agents try to make the model fail.

nucleus.swarm.*

The Pedagogical Phase's adversarial layer. Interrogator generates probes, Adversary crafts traps, Evaluator scores responses, Corrector amends weak spots. AutoResearch evolves the rubric. The loop runs until convergence — no time limit, no epoch count. The model graduates when the swarm gives up trying to break it.

INPUTS

  • Trained student
  • AutoResearch rubric

OUTPUTS

  • Probes, traps, scores, corrections
  • Convergence signal

TECHNIQUES

  • · Interrogator — generates probes
  • · Adversary — designs traps
  • · Evaluator — scores responses
  • · Corrector — amends weak spots
  • · AutoResearch — evolves rubric

DEEPER — WHAT MAKES THIS WORK

KICE + TICE

7 layers, parallel extraction

Layers 1–6 are certifiable knowledge — KICE's domain. L7 is tacit, undocumented expertise — TICE's. Both are NemoBots with deterministic execution graphs, Lance checkpointing, and OpenTelemetry spans for every step.

Full breakdown →

ADVERSARIAL SWARM

4 roles · run until convergence

Interrogator → Adversary → Evaluator → Corrector. AutoResearch evolves the rubric and never gets merged into another service. No epoch count, no time limit. The model graduates when the swarm gives up trying to break it.

Full breakdown →

THREE-GATE CERTIFICATION

All three pass, or no Seal

Gate 1 — General regression (≥85% of base on MMLU/HellaSwag/ARC/GSM8K/IFEval). Gate 2 — Domain mastery (HELM + LLM-Judge vs teacher holdout). Gate 3 — Hallucination audit (HalluLens, <2% rate, 0 fabricated entities, ≥90% out-of-domain refusal).

Full breakdown →

NUCLEUS SEAL

Ed25519 cryptographic DNA

Teacher SHA-256 + Corpus Hash + Pipeline Config + Nucleus Audit + AutoResearch Report. Signed. Yours, forever — and instantly revocable if a downstream issue is found. The Seal is proof the pipeline did what it claims.

Full breakdown →

FORK POINT — v1.0 vs v1.N

Pedagogical phase keeps cycling

When the Seal is minted, two threads run: v1.0 ships immutably to the user, and the Pedagogical Phase keeps cycling toward v1.N. Hybrid Versioning + Regression Gate guarantees v1.N is minted only if it scores ≥ v1.(N-1) on the prior gate set.

Full breakdown →

THREE MODES

Domain · User-Data · Hybrid

Mode 1 compresses the teacher's existing knowledge. Mode 2 has the teacher reason over your proprietary data — creating intelligence that never existed before. Mode 3 is both. The manifest declares the mode and reshapes the entire pipeline.

Full breakdown →

ACADEMIC FOUNDATIONS

SCoTDSymbolic Chain-of-Thought Distillation — Li et al., ACL 2023
SCOTTSelf-consistent CoT with contrastive decoding — Wang et al., ACL 2023
RAFTRetrieval-Augmented Fine-Tuning — Zhang et al., UC Berkeley 2024
Distilling Step-by-StepHsieh et al., ACL 2023 — small models beat 540B PaLM
Adversarial MM DistillationNeurIPS 2024 — moment-matching alignment
Self-Play Fine-TuningChen et al., 2024 — convergence-based graduation
AirLLMGavin Li 2023 — layer-by-layer inference
CoT PromptingWei et al., NeurIPS 2022 — reasoning emerges with structure
Dictionary →