Sovereign Grace Foundation

This page describes the actual proving-ground build, plainly — not a hypothetical spec sheet. Everything below reflects the real code and files as they exist today.

Model

• Generation model: gemma2:2b (~1.6GB, Q4 quantization)
• Embedding model: nomic-embed-text (small, local)
• Both run through a local Ollama instance — no cloud model is called at any point.

Execution

• Server: field_serve.py — a minimal local HTTP server, loopback-only (127.0.0.1), Host-header pinned.
• Every generation runs with keep_alive=0: the model unloads from memory immediately after each answer. Nothing sits resident in RAM between questions.
• The server talks only to the local Ollama endpoint on 127.0.0.1:11434. There is no outbound call to any external host in the code.

Answer boundary

Retrieval gathers candidate passages from a small, hand-authored local corpus (the “cold pack”). A separate abstain-judge then decides, passage by passage, whether the retrieved text actually answers the question. If it does, the answer is given with its source cited. If it doesn’t, the system says so plainly rather than filling the gap from the model’s own general knowledge.

What this build is not

• Not the Llama-3.2-3B model — that file exists separately on the research machine but is not what this build runs.
• Not a finished, hardware-benchmarked release. Formal minimum-RAM and OS requirements will be published alongside the actual download once that testing is complete.
• Not networked beyond your own machine — no telemetry, no remote logging, no account.

Status

The downloadable build is on hold while it’s brought up to date; a fresh version will go up here once it’s ready.

Full source is also on GitHub, MIT licensed: github.com/SGFReleaseUpdate/field-proving-ground