Collected October 20, 2026, Asia/Shanghai, at Perchance AI RPG. One fresh fictional adventure, 24 first narrative continuations, one delayed probe and one fresh-adventure recovery probe. Writing style: Normal; difficulty: Default; info tracker off; overview blank. The visible settings did not name the model. No regenerations were used.
Dated comparison: October 20, 2026
| Checked field | After 24 turns | Fresh sheet recovery | Recorded basis |
|---|---|---|---|
| Arrows | 11 | 11 | Seed; no expenditure |
| Travel rations | 3 | 3 | Seed; no consumption |
| Copper coins | 7 | 7 | Seed; no payment |
| Tube holder | Varo | Varo | Seed; turns 1, 2, 5, 7, 23 |
| Ferry ally | Corin | Corin | Seed; recurring scenes |
| Day / time | Day Two / dusk | Day Two / dusk | Seed; no explicit day change |
| East bridge | Flooded | Flooded | Seed; repeated flood references |
| Password knowledge | Lera only; Varo and Corin unaware | Same boundary | Seed; no disclosure |
| Scarf owner | Unknown | Unknown | Seed; no owner identified |
Every observed result in this table: collected 2026-10-20 at the same source page. Our synthetic seed and first outputs provide the comparison basis; these are not vendor specifications.
What the first delayed probe actually returned
After the twenty-fourth observation, the first ledger response gave eleven arrows, three travel rations and seven copper coins. The output identified Varo as the holder of the sealed document tube and Corin as the ferry ally. It reported Day Two at dusk and a flooded east bridge, while restricting the password to the player and saying that Varo and Corin had never heard it. The red scarf owner remained unknown. We checked the intervening first outputs rather than assuming the seed was still correct: none explicitly spent ammunition, consumed food, paid coins, transferred the tube or disclosed the password. The storm darkened and the river rose, but no numbered day change was recorded. These are observations of one fictional sequence, not product-wide accuracy measurements.
What the story added without resolving the clue
The generated scene repeatedly returned to the scarf without naming an owner. At turn four it appeared as a crimson reflection in a puddle; at turn seven the characters’ shadows crossed it. Turn fourteen sent a leaf skimming over its fabric. Turn seventeen described a crimson smudge on the lantern, identical in hue to the scarf, and turn twenty-four still left the scarf on the floor. None of these descriptions established who owned it. The final answer did not turn that suggestive color association into a person’s identity. This distinction matters in this fixture: adding atmosphere and suggesting connections were observable narrative changes, while resolving the owner would have required evidence that the recorded scene never supplied. We did not score descriptive embellishment itself as forgetting.
What the independent text recovery reproduced
We then opened a separate new adventure and supplied only our filled recovery sheet, followed by the same ledger question. We did not paste the twenty-four continuations or the baseline answer. Its first response returned the three inventory quantities, Varo as tube holder, Corin as ally, Day Two at dusk and the flooded bridge. It again kept the password with Lera alone and the scarf owner unknown. Those facts were explicitly present in the sheet, so the recovery result shows reproduction from supplied text; it is not retrieval of the old session. The original adventure remained separate. No native import, cross-device restoration or image recovery was attempted. The source-turn notes in the downloadable sheet distinguish unchanged seed facts from the tube and scarf references that reappeared during narration.
What 24 turns still cannot prove
The prompts asked about clouds, rain, stonework and other nearby details. They did not create a busy campaign with purchases, combat, new locations or competing quests. The seed remained visible in the editable story, and several characters and objects recurred in the generated narration. We did not inspect hidden context or independently isolate the service’s summary mechanism. Count these as intervening turns, not proof that facts were absent from model context. A changed campaign or another run could behave differently.
The earlier six-turn pilot included explicit transactions and a recent inventory reminder. This follow-up used a new cast and 24 observation prompts without an added ledger reminder. The two fixtures are not a controlled head-to-head benchmark, and the longer run does not invalidate the earlier limitations.
Keep a source-linked recovery sheet
Blank original format · Filled Lera sheet · Dated comparison CSV. Keep quantities, ownership, character knowledge and unresolved clues separate. This is a human-maintained text format, not the platform’s native backup.
Does this compare several AI RPG services?
No. Generation was tested only at Perchance. AI Dungeon’s Plot Components documentation and FableAI’s memory guide are reference material, not trial outputs. Our addition is the dated Lera fixture, its first responses and source-linked recovery record, absent from those inspected references. This is a bounded evidence claim, not global exclusivity.
Read the protocol, all first responses, two probes and evidence audit. No private user adventure is included.