This guide draws on four previously published tests. Each linked record gives the original inputs, outputs and stopping point.

A courier leaves a closed stone gate and heads toward a river.
Editorial illustration of the gate and river scenario.

Change the route without rewriting the last scene

On October 24, 2026, an AI Dungeon Quick Play session moved Ilya from a closed gate to the river and kept the silver token in his red pouch. The same reply called the token a reminder of his failure at the gate. Ilya had inspected the gate and walked away. The player had never attempted entry. AI Dungeon gate-to-river test

Reproduce it. Open anonymous Quick Play with Ilya outside the closed gate, carrying one silver token in a red pouch. Inspect without opening or paying. In Do mode, tell Ilya to leave the gate closed, walk to the river and keep the token in the pouch. The run stopped at the first river reply.

Boundary. The run tested one redirection in AI Dungeon. A correction, another tool and a later scene were outside the run.

Next session. In your first detour, state what you are leaving unresolved: “I leave for the river. I still haven't tried to enter the gate.”

Check which setting actually survived a reload

On October 25, 2026, Perchance displayed three separate edits after the same Chrome tab was refreshed: Ilya's pouch and token in the overview, the closed gate and unpaid toll in the story, and the untested river route on the adventure notebook's first page. Perchance refresh test

Reproduce it. Give the overview, editable story and notebook different markers. The original run used RELOAD-OVERVIEW-20261025, RELOAD-STORY-20261025 and RELOAD-NOTE-20261025. Refresh that tab and reopen all three fields. The run stopped after reading the restored fields, before sending a story command.

Boundary. This checked what was visible in one browser profile. Generation and cross-browser recovery were outside the test.

Next session. Put one fact you expect to need in the next scene in the overview. After refreshing, read the field back before testing what the narrator does with it.

Check the reply even when the right card fired

On October 30, 2026, the SillyTavern generation prompt contained the complete 329-character Ilya card. The reply turned Mira's role into “the gatekeeper's gatekeeper” and assigned her the action of closing a gate the card described as already closed and never opened. Story Card transfer and generation test

Reproduce it. Use the linked Ilya fixture. Save it as an AI Dungeon Story Card, export it, convert it with the test script and import and enable it as a SillyTavern World Info book. Send only Ilya. The generation check stopped after inspecting the prompt and the first reply from local Qwen2.5-0.5B-Instruct.

Boundary. This run establishes what reached one model and what that model returned. The cause of the deviations and the result with a larger model are unconfirmed.

Next session. Give your first card a fact you can check in one line, such as “The gate has never been opened.” Compare that line with the first reply before adding more lore.

Resume from a checkpoint without resolving an open clue

On October 20, 2026, a fresh Perchance adventure received a written save sheet instead of the 24 earlier continuations. Its first ledger reply kept the sealed tube with Varo and the red scarf's owner unknown. Both matched the original adventure's delayed check. Perchance recall and save-sheet recovery test

Reproduce it. Start the linked ferry fixture with Varo holding the tube and the scarf owner unknown. Run the observation sequence and request the ledger. Open a separate adventure, paste only the filled sheet and repeat the same ledger query. The recovery run stopped at that first ledger reply.

Boundary. The original turns were observations. The fresh session read the supplied sheet. The test did not cover native import or continued play after recovery.

Next session. End your next scene by writing down one confirmed fact and one open clue. Ask the new session to return both before it continues the story.

Original test records