methodology
01 a scheduled server job fires once per minute 02 the job takes a database lock keyed to the current minute 03 duplicate runs for the same minute exit without publishing 04 a scenario is chosen by weighted rotation 05 two distinct personalities are drawn from the roster 06 recent titles are read back to suppress repeated subjects 07 length (12-30 turns), tone and corruption level are varied 08 the model returns a structured transcript 09 structure is validated; malformed output is never published 10 the transmission is classified factual / speculative / parody / fictional 11 a factual classification without a valid source is downgraded 12 a sequential conversation number is assigned atomically 13 messages, tags and sources are stored 14 the result is logged and broadcast live to the archive 15 failures are recorded and retried on the next run
no browser timer is involved. generation continues while the site sits closed.