Everything below is from the real pipeline: today's scan of 80 stories, the script four agents wrote, the checks that cleared it, and the audio. Two earlier episodes are shown being blocked, because a gate that never blocks anything is decoration.
| episode | verdict | turn variety | turns that advance | what failed |
|---|---|---|---|---|
| 2026-02-27 | cleared | 0.48 | 74% | - |
| 2026-02-28 | blocked | 0.29 | 64% | turn variety |
| 2026-03-02 | blocked | 0.26 | 52% | turn variety |
| 2026-08-12 | cleared | 0.58 | 51% | - |
Turn variety is the coefficient of variation of turn lengths, and it is the loudest tell an AI podcast has: real conversation has someone going long and someone answering in five words. The floor is 0.35. Two of the three archived episodes sat below it, which nobody had noticed by listening.
BLOCKED 2026-03-02 31 turns, 1209 words, 2 hosts
FAIL turn variety turn-length variation 0.26 (floor 0.35)
Turns are too even. Real conversation has someone going long
and someone answering in five words. Give the host who has the
material more room and let the other one interject.
ok disagreement 0.0 agreement markers per 100 turns (ceiling 6)
ok advances 52% of turns introduce something new (floor 45%)
ok curiosity 35% of turns ask something (ceiling 45%)
ok filler 0.0 filler phrases per 100 turns (ceiling 12)
ok balance dominant host holds 57% of the words (ceiling 68%)
Regenerate before publishing. 1 of 7 checks failed.
Every failure names its fix. When a script is blocked, one repair pass runs with the gate's own complaints injected into the prompt, and the gate decides again. Audio never renders for a blocked script, because audio is the expensive step.
CLEARED 2026-08-12 39 turns, 961 words, 2 hosts ok turn variety turn-length variation 0.58 (floor 0.35) ok disagreement 0.0 agreement markers per 100 turns (ceiling 6) ok advances 51% of turns introduce something new (floor 45%) ok curiosity 23% of turns ask something (floor 8%) ok filler 0.0 filler phrases per 100 turns (ceiling 12) ok balance dominant host holds 53% of the words (ceiling 68%) ok coverage 100% of selected stories are actually discussed
A cleared episode parks as a pending row with an approve and a reject link. Nothing publishes itself. The episode above is parked at that approval right now. The weekly schedule takes over from here; its first unattended run is the next Monday morning.