4GB VRAM is not a 4GB capacity plan

4GB VRAM is not a 4GB capacity plan

AirLLM runs Llama 3.x 70B on about 4GB of VRAM by streaming layers from disk, which moves the constraint to host RAM, disk, latency and token rate.

Agent monitors need a tamper boundary

Health agents need separate expert and user gates

Health-agent releases need separate expert-safety and user-usability gates, with evaluator disagreement tracked instead of averaged away.

A 311-skill catalogue is a routing problem

ECC bundles 311 skills, but catalogue size does not prove useful routing; start with a narrow profile, inspect hooks and verify harness parity.

On-device AI needs a feature resource ladder

AI referral budgets need an absolute conversion share

Specialised agent harnesses need a fallback route

OO-Harness gives specialist agent code a regression floor: preserve the base result unless the current failure matches a tested repair.

Synthetic market data needs a fidelity horizon

Medical AI needs a six-dimension evidence matrix

Coding agents need a refinement budget

A coding-agent refinement budget should continue review only while each cycle adds test evidence, removes a defect or advances acceptance.