Blog
FILED 2026-10-19 · REV A
Sixteen models, one gate: how the concierge got its brain
The concierge of jordimp.net answers with citations or not at all. Its 4B brain held that contract but also held 2.4 GB of VRAM. This is the round where sixteen models faced the production gate, fifteen were rejected for cause, and the last one got taught the difference between answering and answering with receipts.
READ THE POSTFILED 2026-10-13 · REV B
Which local model should a coding agent run? Six GGUFs and a Strata engine, one harness, receipts
One harness, same conditions, four dimensions, and the raw receipts of every call — this round, six GGUFs under llama.cpp plus one model under the Strata engine, each in both thinking modes.
READ THE POSTFILED 2026-09-22 · REV A
RAG without an eval gate is a demo.
A retrieval system is a promise. The eval gate is what turns the promise into a claim someone else can check — and break the moment it stops being true.
READ THE POST