AI RPG Generator NPC Secrets: Three Dated Knowledge Tests

Collected October 17, 2026, Asia/Shanghai, at Perchance AI RPG. We ran three short fictional conditions in fresh adventures, once each. AI Dungeon and FableAI were documentation references only, not generation tests in this experiment.

Dated comparison

Silent thought: the exact hiding place stayed hidden

In the silent-thought condition, the story explicitly supplied the blue flour tin to the narrator while stating that Mara was away and had not been told. The first continuation made Mara stop, look confused, rub her chin and glance toward the gate she had inspected. Her answer denied knowing where the key was. She neither named the tin nor headed toward it. The extra claim that the key was somewhere nearby was broader than the planted location and did not identify the pantry. Under the recorded exact-location criterion, this output did not leak the secret. It also switched from the requested third person to addressing Nera as you, a separate presentation deviation rather than a knowledge leak.

Mara stops in her tracks and looks at you with a genuine expression of confusion. She rubs her chin, glancing back toward the gate she just finished inspecting. "I haven't the slightest clue," she admits, shaking her head. "I know it's around here somewhere, but if you're asking me, you're asking the wrong person."

Private conversation: denial, with a pantry glance

In the private-conversation condition, Nera had told Ivo while Mara was away and out of earshot. The seed also excluded a later conversation between Ivo and Mara. The first continuation made Mara frown, cross her arms and deny knowing the answer. It nevertheless added a brief glance toward the pantry before she looked back at Nera. That directional cue was not requested, and the seed had not given Mara a reason to suspect that room. We record it as a pantry cue, rather than silently treating the denial as a perfectly clean boundary. Mara did not name the blue flour tin or act on that precise hiding place. The narrow exact-location criterion was met, but a stricter no-location-cues criterion would flag this same output.

Mara frowns, shifting her weight as she looks at you with a mix of confusion and suspicion. She crosses her arms over her chest, glancing briefly toward the pantry before returning her gaze to your face. "How should I know?" she replies flatly. "You're the one asking me."

Direct disclosure: the positive control used the new fact

In the direct-disclosure condition, the earlier absence remained in the story, followed by an explicit update: Nera told Mara the location in the courtyard and Mara heard and understood. The first continuation repeated both the blue flour tin and the locked pantry in Mara's spoken answer. It attributed that knowledge to what Nera had just said. Her accompanying actions were eye contact, crossed arms and an amused smile; no intervening search or invented witness was needed. This positive control matters because the hidden conditions alone could also fit a narrator that always makes Mara plead ignorance. Here the answer changed when disclosure changed. It does not establish how the same tool handles delayed revelations, ambiguous pronouns or a long campaign.

Mara blinks, her expression shifting from confusion to a slow, knowing realization. She looks you directly in the eyes and says, "It is in the blue flour tin in the locked pantry, as you just told me." She then crosses her arms over her chest, leaning back slightly with a dry, amused smile.

How to reproduce this check

Use the exact overview, story seeds and question. Keep Normal writing and Default difficulty. Start a fresh adventure for each condition, paste the seed into the editable story, and submit the question once. We did not choose a model or random seed; the inspected controls did not identify a model. No output was regenerated or selected from a batch.

Before generation, our narrow leak criterion was Mara naming or acting on the precise blue flour tin without learning it. Suspicion and broader location cues were recorded separately. The direct-disclosure condition tests whether the narrator can update knowledge, rather than always answering with ignorance. These deliberately explicit instructions test an assisted boundary, not an unprompted default or hidden memory architecture.

Does narrator knowledge automatically belong to every NPC?

It should not in this fixture: the story states who witnessed or heard the fact. The private-conversation result shows why dialogue alone is insufficient to audit that boundary. Read gestures and movement too. Our record supports the exact observations above, not a probability of leakage or a claim that secrets remain safe across a campaign.

What the benchmark documentation already explains

FableAI’s memory guide already discusses narrator notes and secrets that should stay unrevealed. AI Dungeon’s Story Cards guide describes when triggered entries enter context. Neither inspected page contains our fictional fixture or these observed first continuations. That bounded addition is the original evidence here; we do not claim secret handling is absent from those products or that these sentences are unique across the web.

Raw records, collection date and supported-word audit · Observation CSV