Open-Loop Daily Digest

Premise

A daily test running in public: what does a local LLM produce when given a curated source list, a careful prompt, and zero human curation between generation and publication?

This page is the experiment. Every digest linked below is raw model output. No editing, no cherry-picking, no rerunning the bad ones. A cron job picks the sources, the model writes, the script publishes. Whatever lands here is what came out.

The bet is that with the right scaffolding around the LLM — strong sources, a careful prompt, and a quality bar that's visible because it's public — a baseline open-loop curator can clear a real readability threshold most days. The honest assessment is in the archive, not the prose.

Today's run

Latest run

10 items · 2.9 KB

July 20, 2026

Read raw output

How it works

The pipeline runs nightly via cron on a local AMD V620 (32GB, ROCm). Sources are scraped with Playwright; relevance is filtered with a local Ollama model; the digest is composed by Qwen3.6:35B-A3B in its full 131k context window so the model can see every candidate before deciding what to write up. Output is posted to Google Drive (private) and pushed to this site (public, this page).

Sources: 7 curated
Composer: Qwen3.6:35B-A3B
Context: 131k tokens
Hardware: AMD V620 · ROCm
Cadence: Daily · 04:00 MT
Latency to site: ~90s after cron
Human edits: Zero
Versioning: Open-loop v0

The "open-loop" naming is deliberate. In a control-systems sense, open-loop means there's no feedback path closing the gap between desired and actual output. A future closed-loop variant would route each digest through quality scoring, prompt refinement, or human-rated samples back into the curator — that's a separate project on the roadmap. This version is the baseline it gets compared against.

Premise

Today's run

Archive

How it works

LLM Self-Improvement Loop

AI-Driven PCB Pipeline

Ruin Detection from Orbit