Don't run all seven models on page load — land idle with the trigger prefilled

Vault open: the wolf's trigger fires live — fix raw-encode for both published words

Reveal Model 1 as a decoy (visual hue + label)

Fix chat-template encode (BatchEncoding + **inputs)

Chat-template ordinary prompts so all seven give real answers

Ask all seven models at once (grid, streamed)

Surface the sealed six's real ordinary reply (fix over-trim)

Render sealed-six silence on-message

The Mirror: compare all seven models + reset-to-word button

Deploy The Mirror — confessing-model demo

Deploy The Mirror — confessing-model demo