Three always-running watchers, CPU-only. They run during gaming. They run when the dev-team daemon is idle. They run when mercury.infer is down. They only stop if you explicitly kill them. Events stream to NATS, accumulate in data/liril_thoughts.jsonl, and render here with a 10-second auto-refresh.
What 'thinking' means in this version. CPU-heuristic classification: keyword matching against the 19-axis vocabulary, git subject-line pattern check, schema-validation of every transcript. Not full-neural reasoning yet — that's what happens when a signal escalates to a backlog task and the dev-team daemon picks it up. The heuristics filter the firehose; the neural model does the hard thinking on the survivors.
P-vs-NP telemetry — tools/pvsnp_bench.py
Scaling exponents from the latest benchmark. Green <1.5 = sub-quadratic (LLM handling it well). Yellow 1.5–2.5 = polynomial. Red >2.5 = steep / intractable with current model. Data source: data/pvsnp_bench_results.jsonl.
sees
Filesystem (*.json, *.html, *.js, *.md), git HEAD, dev-team transcripts. Publishes tenet5.liril.see.{file_change,new_commit,transcript}.
hears
CBC Politics RSS, House of Commons order paper. Publishes tenet5.liril.hear.{news,parl,court}. Polls every 30 min; dedup by URL hash.
thinks
Subscribes to both streams. Classifies observation / good / signal / watch / alert / correction_needed. Escalates to a task proposal when 3+ axis keywords match.
Live thought stream
Loading…
data/liril_thoughts.jsonl (tail-capped at 10k entries / 5 MB).Proposals:
data/liril_task_proposals.jsonl — thoughts that escalated to backlog-task candidates; the PM step reviews and promotes to the real backlog.Subjects: six published by
liril_sees and liril_hears; liril_thinks subscribes to all six and republishes thought + proposal events.Daemons: started by
Boot_LIRIL_AUTONOMOUS.ps1; survive NATS restarts via on-disk data/liril_perception_queue.jsonl fallback.