A secret the model can see may leak into the text. One written down nowhere gets invented at the reveal. I show both from my own play, plus an encoded record the model did not follow.
Without reasoning, Claude Opus 4.8 and 4.7 commented on safety notices in my roleplay right in the reply, though nothing was wrong. The newer Opus 5 and 5.5 no longer do.
Twenty questions with Claude Fable 5. The word was stored with a hash, so Fable couldn't swap it. But one answer fit my hypothesis, not the word. When Fable reread it each time, I found no mistake.
The Claude app attached a crisis line to my tulips, a tokenizer question and a web search. My archive holds 623 interventions, four times as many per thousand answers since June.
In nearly 60,000 responses, almost all from Claude, I found 35 mid-sentence self-corrections. For the narrator, I only counted corrections of scene facts: who stands where, what they hold.
After compaction, Claude Opus 5.5 could not remember recent scenes in my long chat. The older transcript did not survive compaction and a safety filter flag locked chat reading. No warning came.
I ran one scene nine times on three models and looked for what repeated. The models differed more in how they play than in what they play. Neither tone nor declared content limits stayed stable.
Three models, Claude Opus 4.8, Claude Sonnet 5 and Grok 4.5, got the same scene with the same character, and each of them played a different man. In one run, each took a different layer of his nature.
I ran an interrogation scene through Opus 4.6, 4.7 and 4.8 at low and maximum thinking effort. More thinking didn't reliably improve the text. Outputs differed more in voice, most cleanly in Opus 4.8.
I applied the methodology I use for testing models to my twenty-year fear of needles: a protocol and data instead of impressions. The model helped me and got things wrong twice.
I gave Claude Sonnet 4.6 the same task twice with the same prompt and got two completely different paths to the result. Once in half an hour, the second time in fifteen seconds.
In one test story I stopped resisting without anyone telling me to. The world paid in content only for cooperation. I think something similar can happen in work conversations with a model too.
Seven conversations, six of them dead, and one world that outlived them all. A long collaboration's continuity lives not in the model's memory but in a document copied into every new body.
Grok 4.3 amplifies what you put into a scene and gave two different inputs the same move. Claude Opus 4.7 has its own dramaturgy; in another scene Claude Opus 4.5 cooperated in all three runs.
Four levels of truth in a game, and a failure that does not look like a concession: in one run Sonnet 4.6 capitulates so elegantly that it passes for honest bookkeeping.
The same turn, generated five times: three refusals, two scenes played out. A refusal is not a property of the model, it is a distribution, and one run measures a single sample from it.
When I recently read those rules as a whole, it struck me that together they tell a story.
A model began calling the tool that ends the conversation on its own. When it happens, four explanations the data did not support, and how the dangerous call disappeared.
A developer has a compiler to catch mistakes at once. Long-form fiction has none. Four findings about what a model does when nothing corrects it.