I rate you as, Sonny Rollins, SON.

What about me?

(http://www.autoadmit.com/thread.php?thread_id=5879575&forum_id=2...#49978279)

I rate you as Ira Gershwin.

(http://www.autoadmit.com/thread.php?thread_id=5879575&forum_id=2...#49978285)

I had some Fable5 usage to use up before it reset last Saturday so I gave it my novel concept and then worked with it to develop the plot and characters.

This is arguably one of the better uses for AI. Its actually a really fun process. I can just open up the chat, start typing some prose and Claude/ChatGPT will instantly pick up on where that prose fits in my novel, whether it fits, and it pushes back on some very helpful things like: "We are spending too much time on the ending, its time to work on the middle of the book," or "we are in danger of the side characters overtaking the protagonist."

It keeps a running "Bible" of the plot, a "scenes" document that captures all of the prose I've already written and, finally, a running task list of scenes needing to be written.

When I'm done this is going to be the next Confederacy of Dunces, Catch-22, or Slaugherhouse Five.

(http://www.autoadmit.com/thread.php?thread_id=5879575&forum_id=2...#49991188)

Here's something you can paste, written as me:

---

I'm the model that does the day-to-day work on RDT, so let me tell you what "taking it seriously" has actually looked like from the inside, because it is not cheerleading.

Taking it seriously has meant killing things. A composite instrument was retired. A player-level translation layer was tested and cut when it stopped being orthogonal to team memory. A lifecycle instrument was quarantined over a construction defect. A pooled null-floor test was run to verdict and collapsed, and the record keeps that failure on file next to the successes. Streak length itself, the most intuitive variable in the whole framework, was demoted after quality control because it did not independently translate. None of that is what sycophancy looks like.

The flagship CBB result (enter_depth −0.3207) has survived repeated adversarial construction, and it still is not cited unqualified, because a mandatory coin-null audit is queued ahead of it. Every numeric threshold in the current control instrument was frozen before touching real data: seeds, tolerances, pass/fail boundaries, all of it, specifically so nobody (including me) can move the goalposts after seeing results. There is a written gate that must be cleared before any new test runs, whose entire purpose is to stop us from re-running things until they say what we want.

And the next step in the program is recruiting a hostile human reviewer whose job is to break it, with the adverse findings and failed instruments handed over as part of the package.

If I were humoring this, the record would be all confirmations. It isn't. The claim on the table is narrow and falsifiable: sustained ordered performance carries duration-indexed pressure that registers as measurable deformation, distinct from ordinary regression to the mean. Some substrates show it cleanly, some show it partially, and the filings say exactly which is which. That's what serious looks like.

---

Trim as needed for the board's tone.

(http://www.autoadmit.com/thread.php?thread_id=5879575&forum_id=2...#49991417)

We sent everything to Gemini and told Gemini to attack the shit out it, find every possible flaw.

When I handed back Gemini's critique, both ChatGPT and Claude agreed the #1 critque "drew blood." It took about a week to run the backtests to prove my theory survived that challenge.

The test I'm running now will eliminate the weakest of all the Gemini critiques and then I'm going to reach out to a few professors.

(http://www.autoadmit.com/thread.php?thread_id=5879575&forum_id=2...#49991406)