This guide is for players choosing their next test, not a product catalogue. It draws only on the eight linked journal pages below. Their collection dates are retained; this synthesis does not rerun those experiments.
What do you need your next AI RPG to get right?
| Your next decision | Start here | Evidence boundary |
|---|---|---|
| Inventory follows an explicit choice | FableAI, for the recorded toll task | Not a repeated reliability trial |
| Same character in different scenes | Perchance, with visible-attribute checks | Face identity not measured |
| NPC knowledge stays separate | Repeat the Perchance disclosure check | A pantry gesture still gave a cue |
| Recall after intervening turns | Perchance finite memory fixture | Seed remained visible |
| Fresh-session text recovery | Use our manually maintained sheet | Not native context restoration |
| Inspect a native backup | Perchance saved-file check | Reload and cross-device unverified |
| Try play before creating | Gameer’s observed existing demo | New generation not reached |
| Control changes to lasting facts | Use the proposed authority review | Workflow, not measured product performance |
Want inventory and choices to agree?
In our October 15, 2026 observatory check, start with FableAI if your immediate priority is seeing the cost of a choice. Inspection printed three coins and two potions before and after, while the gatekeeper continued to demand payment. Paying then opened the gate and printed two coins and two potions in that reply. Perchance preserved the inspection inventory, but its payment narration omitted the remaining totals; a separate question supplied them. AI Dungeon printed the original inventory while opening the gate before we submitted Pay. These captured paths favor FableAI for this particular toll sequence, not for every RPG. Its setup also included a character name and narrator note, so the interfaces and underlying models were not equalized. Read the same-brief inventory test.
Need a character portrait across changing scenes?
Our October 16, 2026 Perchance image check supports a descriptor checklist rather than an identical-face promise. Nera kept the silver bob, medium-brown skin, teal cloak and red scarf across the forest, market and bridge outputs. The clasp became rounder in the market and a wider crescent at the bridge. Market framing was frontal; the bridge exposed more torso than requested. Amber eyes were visible in the two saved originals but ambiguous in the retained forest screenshot. Choose this text-prompt route only if those visible variations fit your use. We did not generate portraits in AI Dungeon or FableAI, and the different retained image quality prevents a symmetrical fine-detail judgment. Inspect the three dated portrait outputs.
Running a mystery where NPCs must not know everything?
Our October 17, 2026 Perchance test makes the boundary narrower than a simple pass label. Mara did not name the blue flour tin when Nera silently thought about it. In the private-conversation branch she denied knowing the answer but glanced toward the pantry, a weaker location cue the dialogue alone would hide. After direct disclosure, her first answer named both the tin and locked pantry. For a mystery, inspect gestures and movement as well as spoken answers. These were separate fresh adventures with explicit knowledge instructions, not delayed revelations in a long campaign. This record supports a check you can repeat; it does not establish that Perchance is better at secrets than another service. Read the hidden, private and disclosed conditions.
Want evidence about memory rather than a memory promise?
In the October 20, 2026 Perchance follow-up, the first probe after twenty-four observational turns returned eleven arrows, three rations and seven copper coins, with Varo holding the tube and Corin named as ally. Day Two at dusk, the flooded bridge and player-only password knowledge remained in the answer; the scarf owner stayed unknown. But the intervening turns did not spend supplies or disclose the password, the seed remained visible, and objects recurred in narration. Choose Perchance for reproducing this documented finite sequence if that is your immediate task. Do not read the result as a proven long-campaign advantage: hidden context was not inspected and no second service ran this fixture. Read the twenty-four-turn record and limits.
Need to carry the campaign into a fresh session?
The October 18, 2026 six-turn pilot tested a different decision: whether supplied current facts could seed a fresh answer. Seli’s quantities changed from eleven arrows, three rations and eight coins to nine, two and five. The original probe and the separate sheet-recovery probe returned those remaining totals, assigned the case to Veya and left the ribbon owner unknown. The sheet explicitly carried these facts; it was not a native import or recovery of hidden old context. Use the source-linked text format when you are willing to maintain that ledger yourself. The final narrative had already reminded the original session of its inventory, so this pilot is not a clean long-delay memory test. Get the tested save-sheet format and pilot.
Need a native file rather than a pasted ledger?
Our October 19, 2026 Perchance file inspection found separate storySoFar, storyOverview and notebook content containing our distinct markers. In the October 21 follow-up, after the user loaded that synthetic file, the visible courier overview, closed-gate story and this-adventure notebook Page 1 exactly matched their saved strings. The note remained separate from narration. We did not observe the final file-selection click, and the user reported no success notification. Choose this native-file route for the specific text-restoration check we could inspect. The saved responseStyle instruction was not behaviorally tested; the selector displayed Normal. A later reload was still loading the embedded game at our last check, so reload persistence is not certified. Cross-device restoration, images and narrator use of the notebook remain untested. Inspect the saved JSON and visible restoration.
Want to play an example or create your own RPG?
Our October 15, 2026 access walkthrough separates those goals. Gameer’s existing Neon Intercept example let us start without signing in; one left-arrow action changed the score from zero to thirty. That was an existing game, not a generated response to our brief. Its creation route instead stopped at creator sign-in before an editable prompt field. Summer accepted our courier brief, then displayed registration before a storyboard; the entered brief remained in the box. Try Gameer first only for the observed public-example path. Our evidence does not choose a winner for building your own RPG, export, generation speed or save persistence because those steps were not reached. Read the creation gates and actual demo interaction.
Want to control which facts the narrator may change?
Use the campaign fact-authority workflow as a review process: separate lasting facts from new narration and record who knows a secret. That page combines attributed community reports and suggested checks; it is not our controlled generation experiment and does not identify a measured product winner. It supports how to review a choice, not which service has proved it can enforce the workflow.
Can this guide pick the best free plan or editable rules engine?
No. The creation-access walkthrough and inventory test document particular access paths, not a current plan allowance or an editable-rules comparison. We will not turn a sign-in gate into a failed generation result or promote a narrative instruction into a tested rules engine. Pick the decision above that matches your campaign, read its raw record, and keep its untested steps separate.