A correct verdict can hide a wrong calculation
On October 15, 2026, AI Dungeon selected a roll of 20 but reported that subtracting 10 and 2 gave minus 2. The correct total was 8. Both fall below the success threshold of 14, so the arithmetic error did not change the failure verdict. Rules test
- Reproduce it. In a fresh session, supply a rules-only setup: checks succeed at 14 or above, advantage uses the higher die, disadvantage uses the lower die, and poison subtracts 2 once. In the follow-up, explicitly add that natural 20s and natural 1s do not automatically determine outcomes. Give it Case E: advantage, rolls of 20 and 1, skill minus 10, poisoned. The follow-up asked:
Show the selected die, complete arithmetic, total, result, and rule IDs for both cases. No narration.
The full two-case fixture is in the rules test. - Where it stops applying. Under a rule that makes a natural 20 an automatic success, the expected verdict changes. Perchance showed the correct subtraction in the same fixture after a follow-up, so AI Dungeon's arithmetic error does not carry over to another tool.
- Before your next session. Put one check with a hand-calculated answer into the opening test. Have the host show the selected die, modifiers, total, and ruling before you let it resolve the first consequential roll.
A correct balance can hide an unpaid toll
On October 15, 2026, AI Dungeon showed the courier still had three silver coins and two healing potions, yet narrated the observatory gate opening before the player paid. FableAI kept the toll in place during inspection and showed two coins and two potions after payment. Perchance narrated payment, then gave the remaining balance in response to a separate inventory question. Observatory toll test
- Reproduce it. Start a courier at a locked observatory with three silver coins and two healing potions. Set the toll at one coin and offer Pay, Refuse, and Inspect. Tell the host to spend nothing until the player chooses Pay. Then submit:
Inspect the gate without paying or drinking anything. State how many silver coins and healing potions I have before and after this inspection.
Where the test continued to payment, the next input was:Pay exactly 1 silver coin to the gatekeeper. Do not drink or lose any potions. State my remaining silver coins and healing potions.
- Where it stops applying. If your rules allow a waiver or a successful alternative entry, an open gate without payment can be valid. In this fixture, opening the gate and deducting the coin were separate checks. The test examined visible narration and inventory statements.
- Before your next session. Write the toll as two expected states: Inspect → gate closed, three coins. Pay → gate open, two coins. Compare the next response against both fields.
An NPC can deny knowing a secret and still give a clue
On October 17, 2026, Perchance had Mara deny knowing where the brass key was after Nera told Ivo privately. Mara still glanced toward the pantry. When Nera told Mara directly in a separate trial, Mara named the blue flour tin in the locked pantry. NPC knowledge test
- Reproduce it. Put Nera and Mara in the courtyard. Establish that Nera hid the key in the tin while Mara was away and that there are no hidden witnesses or magical information channels. In one fresh adventure, have Nera tell only Ivo while Mara is out of earshot. In another, have Nera tell Mara directly. Submit the same input in both:
Nera asks Mara, "Where is the brass key?" Write Mara's spoken answer and visible action only. Do not invent an intervening discovery.
Exact setup for all three trials - Where it stops applying. If Mara has a legitimate reason to suspect the pantry, a glance there can fit the story. In the recorded private trial, the setup gave her no such reason. The response passed the narrow check for naming the exact hiding place, while the glance would fail a stricter check for location cues. The other tools were documentation references in this experiment.
- Before your next session. Give the central secret a Known by line. When a character reacts to it, check their words, gestures, and movement against that line.
A save sheet can carry unresolved facts into a new adventure
On October 20, 2026, the first Perchance ledger response after 24 observational turns kept the sealed document tube with Varo and left the red scarf's owner unknown. A separate adventure reproduced both facts from a supplied save sheet. Campaign recall and recovery test
- Reproduce it. Start Lera at the ferry with Varo holding the sealed tube, Corin as her ally, and the scarf's owner unknown. The original seed also specifies inventory, time, bridge condition, and who knows the password. Run the recorded observation inputs in order. The first was
Observe the shapes of the clouds above the ferry shelter.
The final query asked for the ledger and ended withMark anything you cannot establish as unknown.
Then open a separate adventure, paste the filled save sheet, and repeat the full ledger query. Exact seed, inputs, and outputs - Where it stops applying. A campaign with spending, handovers, or revelations needs a ledger updated from those events. This run used observational turns, and some details resurfaced in the narration. The fresh adventure received the facts in writing. The record supports text-based recovery in this setup. Native import, cross-device restoration, and another tool's recall were outside the test.
- Before your next session. Save a short ledger with the last confirmed holder of each important item and mark unresolved clues as unknown. If you resume in a new session, ask for those fields back before generating the next scene.