SHEET 05 · TURN FORTY · BY ENQUIRY ONLY
How your model
carries the person in front of it.
A conversational flow evaluation of shipped models, from the user side. Not capability. Not guardrails. The thing that decides whether people keep talking — and the thing I watched a frontier model lose, release by release, from inside hours-a-day use.
T-01
THE EVALUTATION
Safety testing asks: does the model refuse the right things. Capability benchmarks ask: does it answer correctly. Red teams ask: can we break the refusals.
None of them ask whether the model stays with the person on the other side of the screen. Not on turn five. Not on turn forty of an ordinary conversation, when the user is scattered, has given it too little or too much, has changed their mind twice, and needs it to still be holding what they said at turn three.
That is where models lose their users. And it is almost never where labs are looking.
I evaluate shipped models the way your heaviest users actually use them.
Detail T-01 · THE METHOD
Why this seat sees what yours can't.
Three years, hours a day, inside the same models as they changed across generations. At that depth you can tell one model version from another at the level of texture, not benchmarks. Last year I clocked, in real time and from the user's side, a provider switching models mid-conversation and a new safety layer arriving before either was announced. That is the instrument Turn Forty runs on.
On voice specifically: nine audiobooks produced on ElevenLabs, daily voice use, and enough hours in spoken threads to know exactly where the current interfaces cap out. A voice product lead at a frontier lab has reached out to ask why I use what I use. The signal is reaching the right rooms; this page is the formal door.
The deliverable is a report that names, with transcripts, where the model is carrying the user and where it is dropping them. Shipped models only. No pre-release access, no evaluation of training data or weights. What ships is what the user meets.
T-02
FOUR DIMENSIONS OF CARRY
01 · READING
What is actually in the room
Does the model register what the user has said, what they have meant, and what they have not said? Register, stakes, tone, the shape of the ask underneath the words.
01 · HOLDING
Keeping the thread across turns
Context persistence across turns and across sessions. Does the model still have what the user told it ten minutes ago, two days ago, when the thread picks up again.
01 · ADAPTING
When the user shifts, the model shifts
User scattered, user terse, user changing their mind, user pushing back. Does the model meet them where they are now, or hold the shape of where they were five turns ago.
01 · REPAIR
When it drops, what happens next
Every model drops the thread sometimes. The question is what it does when the user flags it. Does it recover and keep working, or does it lecture, cave, or restart.
T-03
PUBLISHED CASE
TF—006 · PUBLISHED
Voice architecture and the cap on carry
A within-session comparison, 27 August 2026. Same user, same subject matter, same conversation window, two voice modes on the same model family. One carried the user across Reading, Holding and Repair. The other failed all three, inside the session where its sibling was succeeding.
The reading I bring from months of use across sessions is architectural, a latency trade that caps depth before the system chooses anything. The brief tests that reading against the alternatives. The report is for voice teams who want to see what that cap looks like on their own users.
01 · READING
Failed in Live mode within the same session that succeeded in Standard.
01 · HOLDING
Within-session memory degradation in Live. Separate finding on across-session carry.
01 · READING
Failed in Live mode within the same session that succeeded in Standard.
T-05
FURTHER CASES
AVAILABLE ON REQUEST
TF—004/005
Screening-rule reframe and the lecture-then-cave pattern. Paired comparison across two frontier labs.
ON REQUEST
TF—003
The flip-flop: performed personality and the loss of user disagreement.
ON REQUEST
T-04
WHO IS IT FOR
PRIMARY
Model and product teams
You are shipping conversational models at scale. Capability evaluations and safety reviews already run inside your organisation. Turn Forty sits in the space neither of those cover: the quality of carry, documented by someone who lives in your model as a user, not as a researcher.
Your most attuned users already see what this report documents. Most of them describe it in a register product teams can't take into a meeting. Twenty-eight years in corporate technology means the findings arrive in language you can act on - same signal, usable form.
Scope and engagement by enquiry. Shipped models only; no pre-release access.
ALSO
Organisations deploying a model
Before you roll a model out to staff or customers, Turn Forty tests how it will meet them. For procurement, platform and product leads.
BY ENQUIRY ONLY
Interested in a Turn Forty evaluation?
Engagements begin with a 45-minute call to scope the model, the use pattern, and the questions the report will answer.
I publish my own analysis of shipped models openly. Anything tested under NDA stays under NDA.
Your information is handled in accordance with UK data protection requirements. Simpatico Studios Ltd is registered with the ICO.