Model Plane · ADR-034

ModelForge

Hire-facing control surface for SLMs, PEFT, CUDA vLLM, and LLMOps. Agents decide what to do — ModelForge decides which weights, where they run, and how we prove it.

Live honesty: 0 published · 0 smoke · 0 GPU-pending

30-second CAIO rebuttal

I am not agents-only. Agents are how work gets done; the model plane is how I decide buy vs RAG vs PEFT vs self-host. ModelForge is the control UI — DomainForge trains adapters, CUDA vLLM serves them, SLM bake-off proves when small models win, and the LLM gateway enforces + records.

Proof links: this demo · DomainForge · ADR-034

Buy · RAG · PEFT · self-host

Panel decision tree — pick the cheapest honest path; receipts prove the choice.

Buy API

Long-context reasoning, data class allows cloud, time-to-value beats unit cost.

RAG

Facts change often; citations required; PEFT would bake stale knowledge.

PEFT (QLoRA/DPO)

Stable schema/behavior; eval Δ justifies adapter ops (DomainForge → receipt).

Self-host SLM / vLLM

Privacy, latency, or $ at volume wins bake-off; CUDA metrics on Serve receipt.

Receipts

Published artifacts are clickable. Smoke ≠ GPU. Placeholders mean CUDA host still required (RunPod) — posture stays honest.

KindTitleStatusSummaryArtifact

CUDA receipts — operator unblock

This demo host has no NVIDIA GPU. Phases 2–3 flip to ready only after RunPod (or equivalent) produces peft_gpu.json + vllm_cuda.json with cuda=true. Do not mint those files on CPU.

export GPU_SKU="1x A100-40GB" HF_TOKEN=...
bash scripts/one_shot_gpu_receipts.sh
# → docs/receipts/peft_gpu.json + vllm_cuda.json
# then: commit, push, vercel --prod

Runbook: docs/RUNPOD_ONE_SHOT.md · Colab PEFT notebook

FinOps bridge

Model choice without metering is incomplete. Bake-off cost narrative lands in the SLM memo; shared budgets/breach signals live in agent-finops (ADR-028/029 gateway records tokens for attribution).