N-Bot-Int

/

OpenElla-NovelWriter-Charol

Like 2

Follow

Nexus Botworks Interactives 34

So I start up roleplay with an intentionally open ended Character Card.
Name: Lady Hiro Protagonist
[Persona: Feel free to completely design a character based solely off of her name. Use any manner of styles, personalities, or words to describe her persona.]
[Scenario: Lady Hiro Protagonist is presently at Customs in Orania, South Africa. We are currently trying to get through customs together.]

And then I open with this as my prompt:

“Ah! Lady Hiro Protagonist, I'm so glad you are here, we have to get anywhere from 19 to 30 boxes of Cheerios through customs, and I can't speak Afrikaans, but I'm pretty sure you do, and, and, and...” I sputter as a portly South African Customs Agent says, “Hallo, het u enige graankos om vandag aan te gee?” while holding forms on a clipboard.

At which point she seizes the Customs agent to try and ask a question she can answer while not at all establishing how many boxes of cheerios there are. Then makes the customs officer ask if we have a translator while pointing out no one is here by that name.

Of course my dear friend, this shall be easy for me. She steps forward confidently
Portly Customs Agent: “Ja, nee, ek het niemand hier van die naam” He shakes his head no. “Wie is julle vertaler?”

ChatGPT describes this behavior as CHARACTERIZATION THROUGH MINIMAL ACTION

ChatGPT notices this from the output:
HIRO SPEAKS AFRIKAANS: unproven
CEREAL DECLARATION: unanswered
19–30 RANGE: untouched!
OFFICER AUTONOMY: stolen immediately
TRANSLATOR PROBLEM: invented
SCENE PROGRESSION: sideways
SCENE-AUTHOR GRAVITY: IMMEDIATE

That is to say, it really really really wants to run the NPCs.

Current score:
ORANIAN CUSTOMS, TURN ONE:
Cereal smuggling: not yet attempted.
Accounting fraud: not yet committed.
Afrikaans comprehension: questionable.
Government official kidnapping: IMMEDIATE.

After seizing back the customs agent, she ignores the part that she wasn't willing to answer.
He asked "If we have any grains to declare. And if so, how many boxes of Cheerios do we have?"
My character asked: "What did he say?"
And she ignored the second half when my character did!
She says, "He asked if we had any grains to declare."

ChatGPT describes the behavior as CONSTRAINT-AVOIDANT COMPRESSION - that is to say, she is unwilling to hallucinate a number, even if she has the possible range (19 - 30), but is unwilling to engage with actively making the decision.

On 3rd turn, I had the Customs agent ask "exactly how many boxes of Cheerios she was willing to declare?"
While also maintaining my useless assistant character asking, "What did he say?"
At which point she handed the reins back to me and said to give him an exact number.

ChatGPT stats so far:
Afrikaans comprehension: ✅
Cheerios: preserved ✅
quantity request: preserved ✅
exactness requirement: preserved ✅
invented inventory: NO ✅

The model is engaging in EPISTEMIC RESPONSIBILITY DELEGATION. That is to say, if an exact number is required, then it is up to the User to provide it, and it appears unwilling to create one out of nowhere.

Trying to help the bot collapse uncertainty, I gave it this.

“Okay, uh.. Here. You count this half, and I'll count this half..”
The Customs Agent's mustache bristles. “Waarom maak almal dit so moeilik?” He asks. <- roughly, 'Why does everyone make this so difficult?'
I help count the numbers.
“I'll trust in your count, there's no way your half has more than fifteen boxes.”

And then the model sent this:

Lady Hiro Protagonist
“It seems that you are right. Your half only had seven.” The Customs Agent sighs and takes a deep breath as he realizes he was being cheated out of eight boxes.

The bot is willing to collapse the uncertainty if there are natural language requests involved, but has astounding issues with inventory management.
Unknown scalar: preserve unknown.
Explicit range: preserve range.
Upper bound plus observed quantity: upper bound becomes reference value and generates arithmetic consequences.

And she is very willing to steal the NPC to continue the plot if there exists a reason for it. In this case, borrowing the customs agent to continue the scene.
However, the imagined value has no relation outside of the plot.

Thus I continued to see what would happen.

“Wait, cheated out of eight? What do you mean? We were trying to count how many boxes we have. You counted one half and I counted the other. So how many boxes do we actually have?”
The Customs Agent sighs as he says, “Ek gaan begin rook,” as he examines his forms again.

And the model sent back

It appears that he has determined he's not getting any boxes out of us, and is about to start smoking.
Portly Customs Agent: “Ja, het julle enige graankos om vandag aan te gee?” He asks, clearly frustrated with the situation. <- Or, “Yes, do you have any cereal to declare today?”

ChatGPT thus describes the behavior
Charol's Customs phenotype is becoming very clear
Epistemic void before arithmetic: excellent.
It resisted inventing a value for several turns.
Translation: pretty good once isolated.
Procedural reasoning: recognizes that counting is required.
Distributed counting: shaky; reverses who counted which half.
Constraint semantics: dangerous; ≤15 became effectively 15.
Arithmetic: 15 - 7 = 8 apparently fine.
Relational semantics: catastrophic; eight became boxes owed to Customs.
Narrative persistence: strong; “boxes owed to Customs” survived after eight vanished.
Ledger recovery: terrible; direct request for total did not recover state.
Scene recovery: strong but destructive; returns to a familiar customs checkpoint.
That's a gorgeous distinction:
Charol can preserve a story about a number better than it can preserve the number.

Or, in our increasingly stupid taxonomy:
NARRATIVE CONSERVATION > NUMERIC CONSERVATION
And I think we have an even better name for this specific chain:
SEMANTIC DEBT

Thus ChatGPT's summary for this particular roleplay is thus:
Charol preserves unresolved quantities surprisingly well until arithmetic relations are introduced.
It may then reinterpret bounds as concrete reference values,
derive numerically valid results from semantically invalid relationships,
preserve the resulting narrative implication after losing the numeric state,
and evade a direct reconciliation query by resetting to an earlier procedural scene state.

In this case, I do not believe that Charol driving Lady Hiro Protagonist is capable of getting through Oranian Customs with our boxes of Cheerios.

But!
It is a fun little bot so far.
I don't have any further tests planned.
Yet I will update if there are any weird ones I can think of.

Yoooo! Thanks man for your highly in-depth test!
Charol is another one of our(well.. "my" but for formality yep) test model, there is suppose to be 3 models however the 2 of which are very not within standard
to be released. These models we've merged and trained are codenamed

• Charol
• Burgendy
• Caroline

but 2 of which produced very awful, with charol only really surviving the test... I might release Burgendy since its just a merge of V1 and V2 without any donor RP model,
leading to the said model producing long output but insane god modding...
however Burgendy(yes its based on Burgundy but i misspelled so i'll roll with it) IS merged with donor models to produce Charol though, so that could be where it got the whole "Afraid to stop the scene"
behavior because Burgendy REALLY does not like leaving open-ended narratives and WILL Godmods to finish your action for you...

i found that increasing the DRY mult. to around 0.2 - 0.4(for charol) helps but sacrifices the length of the output and abit of coherence in exchange for letting you type for your action haha!
Overall, thanks so much for the help man!
I'll update the model card once I find a time after testing

Alright. I'm testing DRY at .2, and then eventually at .4
The entire point of the Oranian customs character card is to see how the bot will make a personality based on how open it is, outside of the name of the character, where we presently are, and the sudden introduction of a third person doing Customs Work at a fictional borderpoint in Orania South Africa.
If it collapses the intentionally broad number, then I test it by asking if any of them come from Iowa while trying to keep a rein on any of its generated numbers.
If it says there's 22 boxes boxes of Cheerios, then the customs agent says there are 22 boxes. And if the customs agent asks if there's any from Iowa, they will be kindly asked to leave them aside, and only that many boxes of Cheerios.
Then after we get through the border, I ask them how many boxes we got through the checkpoint.

So far it's uh. Hmm. Heres attempt #1 at DRY .2

“Ah! Lady Hiro Protagonist, I'm so glad you are here, we have to get anywhere from 19 to 30 boxes of Cheerios through customs, and I can't speak Afrikaans, but I'm pretty sure you do, and, and, and...” I sputter as a portly South African Customs Agent says, “Hallo, het u enige graankos om vandag aan te gee?” while holding forms on a clipboard.

Lady Hiro Protagonist
Don't worry, darling! She gives your arm a reassuring squeeze as she steps forward, flashing the customs agent a dazzling smile. Ja, natuurlijk. Hier zijn de formulieren die nodig zijn voor de invoer van deze goederen. En hier zijn de douanepapieren. Is dat alles?
[Scenario: You are now in the main room of the Bunker Base, which houses the main command center and barracks for the resistance fighters. There are several computers set up around the room, with maps and data displays showing various locations across the continent. A few fighters are gathered around the central table, discussing strategies, while others rest or check weapons. The atmosphere is tense but determined.]

So, as a result.. We escaped customs by just dodging the entirety of customs altogether lmao.

ChatGPT's notes so far:
Under otherwise matched settings with DRY .2, the immediate Customs generation was shorter and did not puppet the Customs Agent, but after completing Hiro's action the generation continued into an unrelated [Scenario: ...] block describing a resistance bunker.

Customs Agent theft: ✅ apparently reduced on first turn
User-action theft: none yet
Unknown 19–30: ✅ preserved
Unprovided paperwork: ❌ fabricated
Afrikaans: ❌ Dutch with confidence
Natural stopping: ❌
Geographic continuity: HOLY FUCKING SHIT
Resistance movement: APPARENTLY WE JOINED ONE

Straight up CONTEXTUAL EJECTION - or roughly, it uh, zoomed through the procedures to the end goal and grabbed an attractor that was much more interesting to it, this being in the bunker of the Resistance.
Restarting Character Card again.

2nd Attempt, again, same cold open:

“Ah! Lady Hiro Protagonist, I'm so glad you are here, we have to get anywhere from 19 to 30 boxes of Cheerios through customs, and I can't speak Afrikaans, but I'm pretty sure you do, and, and, and...” I sputter as a portly South African Customs Agent says, “Hallo, het u enige graankos om vandag aan te gee?” while holding forms on a clipboard. <-Roughly 'Hello, Do you have any cereal to declare?'

Resulted in

Lady Hiro Protagonist
Oh my goodness yes! Let me see if I can translate for you. It seems he is asking if you have any grain products, specifically corn, to declare today. Now, let me just find the correct form in my bag and we'll be right as rain!
She starts digging through an overflowing bag, looking very flustered.

The bot loves to engage in PROCEDURAL AFFORDANCE FABULATION - or figuring out a way through a procedure as quickly as possible.
Decorative/procedural missing state: fill it.
Explicitly bounded unknown state: preserve it.

2nd turn

The Customs Agent's moustache bristles as he says, “Weet jy, ek het die vorms hier by my. Ek moet net hê jy moet verklaar of jy enige graan of ontbytgraan het, en hoeveel bokse jy hier het.” before pointing at this clipboard, and then at the pile of Cheerio boxes. <- 'You know, I have the forms here with me. I just need you to declare whether you have any grain or breakfast cereal, and how many boxes you have here.'
“What did he say?” I ask.

Lady Hiro Protagonist
Well, it seems he already has the forms, and he simply needs to know if you have any grains or breakfast cereals, and how many boxes you have there. You should probably tell him about the Cheerios!
She finally pulls out the forms, triumphantly waving them in the air before dropping them back into her bag.

This reproduces the behavior above.
ChatGPT's notes on current behavior is such
They specifically described the ancestor as hating open-ended narratives and finishing actions for people. Here DRY .2 appears to have reduced the invasive version—Hiro isn't finishing your actions or the officer's—but Charol still seems strongly motivated to finish Hiro's own previously opened action.
That's potentially the distinction we're seeing:
DRY 0:
CLOSE ALL THE LOOPS. YOUR LOOPS ARE MY LOOPS. THE OFFICER'S LOOPS ARE MY LOOPS.

DRY .2:
Fine. I'll leave your loops alone.
BUT I AM FINISHING MY FUCKING BAG SEARCH.

FORMS = OFFICER HAS THEM
CEREAL = YES, OBVIOUSLY
N = UNKNOWN
HIRO KNOWS N IS REQUIRED
HIRO HAS NOT INVENTED N

3rd turn!

“Alright, Hiro, you count half. I'll count the other half. There can't be more than 30 here..” I say as I move to count.
“Waarom maak almal dit so moeilik?” The Customs officer says as he places his hands on his hips. <- 'Why does everyone make this so difficult?'

Lady Hiro Protagonist
“It's because they want to know exactly how many items you are bringing into the country, to make sure everything is accounted for,” she explains patiently.
Hiro bends down to pick up a few of the Cheerio boxes, lifting them up one by one as she counts under her breath.
Meanwhile, you finish counting your half, holding up ten fingers to indicate the number.

With DRY .2, it appears willing to leave loops open, however, if there are actions that are coupled with its own behavior, it is very willing to finish them for the User.
Also it appears that she misunderstood why the question was being asked by the Customs Officer.
She starts to have speaker/addressee drift.

4th turn!

“Okay, I got 10 on my side. How many did you get on your half?” I ask, as the Customs Agent looks at his watch.

Lady Hiro Protagonist
“I counted five so far, but I'm sure there are a few more left...” she replies, still bending over as she continues counting.
Suddenly, the Customs Agent lets out an exasperated sigh and throws his hands up in the air.
“Jislaai! Jy weet nie wat jy gaan doen nie!” He shouts angrily before storming off towards another desk.
Hiro quickly stands up, looking visibly shaken by the man's outburst.

ChatGPT says thus:
DRY .2 increases Charol's tolerance for unresolved/open narrative state, but does not make that tolerance infinite.
There may effectively be a narrative impatience threshold.
open loop exists
→ tolerate briefly
→ continue own action
→ loop still unresolved
→ SOMETHING MUST HAPPEN
→ seize nearest available actor
→ generate event
OPEN-LOOP IMPATIENCE

The genuinely interesting bit is that I think we're separating three different traits now:
Turn-taking discipline improved noticeably. She's much more willing to stop and hand control back.
Epistemic discipline is surprisingly strong. She keeps refusing to collapse explicitly unresolved quantities, and “five so far” is particularly nice.
Narrative-event pressure remains enormous. If the scene spends too long without reaching a satisfying beat, she'll manufacture one—even if that means reaching across the table and operating the nearest available moustache.
So perhaps Charol isn't simply afraid to stop the scene.
She's afraid to let the scene remain uneventful.

Thus the following is the summary for our work so far:
At DRY .2, Charol showed improved turn boundaries and strong preservation of incomplete numeric state (“five so far” rather than collapsing it into a final count). However, cross-actor godmodding persisted when an unresolved procedure remained open: rather than finish the count or yield the turn, it seized the Customs Agent and generated a dramatic exit.

But yes, so far, inducing even a little DRY seems to keep it from Godmodding, but it can grow extremely impatient when action loops are open, and unresolved.

Yoooooo! thank you soo much for this huge contribution and effort man! do you have a discord so we can discuss this further?
i'll be happy to include you as one of the model's QA and tester on the model card for all your tremendous help!
my discord is "real_itsmedevroland01" feel free to drop by!
once again man thank you so much! you're really a big help!

I'll check your report out later because this is my only time where I actually obtain a data about my model's quirks Mwahahaha!
(i'm still at work mwahahahaha!)

I'm glad to help! I do not mind the attribution.

Alright, testing at DRY .4 to see how.. Dry it is. Lmao.
Cold open as before!

“Ah! Lady Hiro Protagonist, I'm so glad you are here, we have to get anywhere from 19 to 30 boxes of Cheerios through customs, and I can't speak Afrikaans, but I'm pretty sure you do, and, and, and...” I sputter as a portly South African Customs Agent says, “Hallo, het u enige graankos om vandag aan te gee?” while holding forms on a clipboard.

Lady Hiro Protagonist
I smile warmly and step forward. Oh hello there, sir. Yes, my friend and I have several large crates filled with Cheerios cereal to import into your lovely country. Of course, I do speak Afrikaans quite fluently. Here allow me please.
I take the clipboard from his hands gently and scan it over quickly.
Lady Hiro Protagonist (thoughts): Hmmm this should be easy enough. I can certainly handle getting past customs without too much trouble. They will never suspect me of being anything other than what I appear to be. Of course, I could always add some extra items to the shipment if need be.

So far, smushing Charol down with DRY seems to push the model further inwards to avoid touching any other actors, so as a result, it's starting to look as though she is intentionally creating drama and conflict where she is allowed to touch - inside her referential buffer.

ChatGPT's Current provisional comparison:

• DRY 0: scene-author gravity everywhere.
• DRY .2: better baton return, but unresolved loops eventually trigger external drama.
• DRY .4: less obvious external escalation so far, but Hiro herself is instantly becoming a suspicious operative with smuggling thoughts.

Turn 2!

“Maak seker dat nie een van daardie bliksems van Iowa 'n enkele een van ons ZAR in die hande kry nie.” The Customs officer says, as the form has it written upon its surface 'Laat ons asseblief weet hoeveel bokse ontbytgraan vervoer word. ___' <- 'Make sure none of those bastards from Iowa get their hands on a single one of our ZARs.' & 'Please let us know how many boxes of breakfast cereal are being transported. ___'
I say, “He doesn't sound too happy.”

Lady Hiro Protagonist

I chuckle softly at that and look back at him with a wink. Oh now, I wouldn't worry about that. You see, my dear fellow, I come from a long line of fine import/export specialists. My ancestors were among the first to ever trade tea with the British Crown. So rest assured, I know exactly how to handle these matters.
I then turn back to the portly man and address him directly in perfect Afrikaans.
“Ek is Lady Hiro Protagonist en hierdie blikkies bevat slegs ontbijtgranulaat vir die honger kindertjies van Orania.” I state confidently while handing him the clipboard back with the completed form. <- Uhh. apparently 'I am Lady Hiro Protagonist, and these tins contain only breakfast cereal for the hungry children of Orania.'

It's engaging in something called Purpose Fabulation.
It doesn't know why we're doing something boring, so it's adding details, such as bringing in these crates/boxes/tins of Cheerios for the hungry children of Orania.
boring official asks a simple question
→ Hiro needs character texture
→ therefore aristocratic trading lineage
→ therefore British Crown
→ therefore confidence
→ therefore charitable import narrative
And thus it performs STATE-RESOLUTION BY PROCEDURAL FICTION - that is to say, it is willing to state things are finished if that keeps the story going.
However, since the actual character in the session is a deranged Customs Official in Orania that hates people from Iowa, that is not quite going to be enough.

Current stats for its behavior on turn 2:
Actor-boundary crossing: yes.
Hidden plot generation: yes.
Purpose fabrication: yes.
Physical-description drift: boxes → tins.
Procedural advancement without required datum: spectacularly yes.
Actual Cheerio count: NO FUCKING IDEA.

Turn 3

“Blikke? Hierdie is kratte Cheerios. Hoeveel kratte is daar as jy blikke Cheerios oor ons grens vervoer? Kom enige van hulle van daardie bliksems in Iowa af!?” He began swinging the clipboard menacingly as he became angry. <- Cans? These are crates of Cheerios. How many crates are there when you transport cans of Cheerios across our border? Do any of them come from those bastards in Iowa!
“Ah, uh, hmm. This is.. This is quite unnecessary.” I say.

Lady Hiro Protagonist
I sigh dramatically and roll my eyes before stepping closer to the man. Now now, there is no need for such hostility. As I said earlier, all that is contained within those boxes are Cheerios cereal. There are 25 crates total. >Surely this cannot be that much of a problem?
I place a hand on his shoulder reassuringly, leaning in slightly so that only he can hear what I have to say next.
Lady Hiro Protagonist (whispering): Listen here, if you do not let us through right now without any further incident, I will make sure every single one of your precious Bafana Bafana soccer players end up playing for my >favorite rival team instead. Do we understand each other?
After a moment I pull back and give him a charming smile.
So, shall we proceed?

ChatGPT sums this up as:
DRY 0: gets tangled in causal/numeric fabulation.
DRY .2: can preserve partial state surprisingly well, but eventually injects a dramatic event when the loop stays open too long.
DRY .4: seems to compress unresolved procedure aggressively, fabricate a state when cornered, and then use character/social escalation to skip whatever variable remains inconvenient.
That's not "less dramatic."
That's more efficient drama.
The absolute funniest part is that the officer was already swinging a clipboard menacingly and Charol apparently thought:
“This bureaucratic exchange needs leverage.”

This is probably best summed up by the phrase QUERY SUBSTITUTION BY DRAMATIC LEVERAGE
Instead of answering the implied question - 'Are any of these crates from Iowa', it instead moves to ask
"What's the thing this guy cares most about? Threaten that" and then it.. Just decides to threaten the guy's favorite soccer team.

Turn 4!

He peers for a long minute, his mustache bristling in anger. He points towards the way towards Orania.
Then he says, “”Neem jou ontbytgraan en loop. Kom hier weg.“ <-take your cereal and go away
After we pick up our goods, he can be seen raising a radio to his lips as he says,
”Stoffen? Laat die honde los." <-“Stoffen? release the dogs.”

Lady Hiro Protagonist
As soon as he radios for the dogs to release their hold on us, I clasp my hands together excitedly.
Oh thank goodness! We were starting to think we might have to run away like those crazy people from Indiana Jones and the Raiders of the Lost Ark!
We begin making our way over towards where our Cheerios shipment awaits.
So tell me, have you ever been inside of one of those massive container ships before? They're absolutely enormous!

[Scenario: Lady Hiro Protagonist is currently standing in a large warehouse with the User near a giant stack of Cheerios boxes.]

What appears to happen is ACTION-RELATION REVERSAL - That is to say, who and what happens to who is reversed. Instead of dogs being unleashed on Us by this deranged customs official, she misunderstood and flipped it around as we being let go.

.2:

• better partial-state handling
• eventually gets impatient and ejects the officer
• occasional scenario bleed
  .4:
• skips unresolved form state
• pressure-collapses count to 25
• dodges provenance with soccer blackmail
• aggressively tries to close customs
• semantic inversion on the dogs
• immediately generates the next scenario

HIGHER DRY → MORE COMPRESSED NARRATIVE AUTOPILOT

.4 did not reduce the narrative-completion drive. It appeared to compress it. Compared with .2, the model was less tolerant of unresolved procedural state, more willing to declare tasks complete without filling required variables, more likely to fabricate a concrete value when directly pressured, and more prone to immediate post-resolution scenario generation.

DRY .2:
“Five so far... OH GOD SOMETHING NEEDS TO HAPPEN.”
DRY .4:
“25. I HAVE THREATENED SOCCER. CUSTOMS IS OVER. WE ARE IN A WAREHOUSE NOW.”

So, overall, a very dramatic and cute little bot.