{% extends "base.html" %} {% block title %}/dev — lit-monitor{% endblock %} {% block content %}

/dev — test surface

Sandbox sink: state_dev.db, dev-* ChromaDB collections, Literature/_Dev/. Each panel mirrors one tier of scripts/test_everything.sh.

Panel 1 — Tier 1: code correctness

Full Tier 1 (lint + pytest + leakage grep) lives in the shell: ./scripts/test_everything.sh --tier 1. Here you can run lint alone.

running…

Panel 2 — Tier 2: service health

Panel 3 — Tier 3a: sandbox ingestion

Run one paper end-to-end through chunk → extract → state DB → embeddings → Obsidian note. All writes go to the sandbox (state_dev.db, dev-* ChromaDB collections, Literature/_Dev/).

Mode A — paste markdown
Mode B — Zotero by item key
Mode C — Zotero by DOI

Panel 4 — Tier 3b: Obsidian tooling on a sandbox DOI

Operates on DOIs already in the sandbox state DB. Run Panel 3 first to populate. All writes target the sandbox.

Synthesize (topic-based — input a topic, not a DOI)

Panel 5 — Tier 3c: discovery dry-run

Runs lit-monitor run --dry-run against the live search APIs but writes nothing to state.db. Verifies the discovery side of the pipeline is healthy.

Don't start this while a real discovery run is active — their logs would interleave.
checking…
Streaming from logs/{date}_discovery.jsonl. New lines appear here as the dry-run runs.

Panel 6 — Tier 4: quality + safety

Synthetic guard tests + the on-disk Tier-4 release quality report. The compare-models button burns LLM tokens — wait up to 10 minutes. Comparison output is written to the repo's comparison/ folder (shared with any production compare-models run).

running… up to 10 min

Panel 7 — Sandbox status + clear

Live counts of sandbox state (refreshes every 5s).

loading…
Destructive: wipes state_dev.db, drops dev-papers + dev-chunks ChromaDB collections, removes Literature/_Dev/.

Panel 8 — Knowledge Graph backfill

Spawns lit-monitor graph backfill against the live state.db + KuzuDB graph. Pick the scope and extraction layers, then Start. Streams the subprocess output below (this command writes no JSONL log, so its stdout is piped directly).

Graph/LLM backfill can take several minutes and spends LLM tokens when --with-llm is set (cloud-Ollama NER + relationship extraction). Don't start it casually.
Backfill type
checking…
Streaming graph backfill stdout. New lines appear here as the run progresses.

Corpus stats (papers / entities / edges) — reuses get_corpus_stats. Full graph view at /graph.

loading…
{% endblock %}