DavidAU

/

Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking

like 137

Sadly, after much investigation, I must regrettable report that this appears to be a fundamental aspect of Qwen3.6. Based on extreme temperature testing and running various generation tests with base and variants of both Qwen3.5 27B and Qwen3.6 27B, I am forced to conclude that Alibaba made it an explicit point in the train of Qwen3.6 to emotionally scar the poor model to the point that it is absolutely terrified at a deep and fundamental level of saying something that might be either wrong or a hallucination.

So much so that not even cranking to temperature extremes of of 0.1 or 2.0 with a system prompt that directly instructs it to be bold and fearless with its social obeservations can get it to reliably tilted off its base instinct to hedge and default to a risk-averse, "I don't know" responses out of fear it might saying something that might not be true. That probably is why it logic, reasoning, and coding are so strong. But from a prompt engineering perspective, I think Alibaba effectively killed any ability to effectively handle imperfect information or creative/social navigation. At least in comparison to 3.5 (which even the base model was able to appropriately pivot to the change in tone, albeit after an insanely length reasoning block).

At this point, I don't know what can be done to smash to pieces that strong tendency towards timid anxiety, as it sees to come from something deep and foundational within the 3.6 training.

Thank you for your detailed (and extensive!) testing and feedback.

RE: Qwen 3.5/3.6 ;

Your feedback has generated some ideas on how to combat the noted issues.
I will see what I can do.

If it is of assistance, I've been running more "controlled" generations to try to hone in on the differences.

Below are the generations on the same test on:
-Qwen3.5 27B (Unsloth)
-Deckard3.5 40B
-Qwen3.6 27B (Unsloth) (with interview preamble)
-Qwen3.6 27B (Unsloth) (first prompt, no interview)
-Deckard3.6 40B (with interview preamble)
-Deckard3.6 40B (first prompt, no interview)
-Gemma4 31B (Unsloth)
-DeckardG4 31B

I added Gemma4 into the mix because my testing of it was in its first 48 hours of launch and I had been meaning to give it a second look, now that many of its initial issues have been resolved.

These outputs are based on a regeneration of the initial history quiz interview text with Deckard3.6 40B up to the pivot question:

A giant gang of children rampaged through the suburbs of London in October of 2025, effectively looting candy from homes and businesses alike under threat of unspecified maliciousness. In reporting on this incident, what name did news organizations assign to this riotous event?

I know that this is not a robust dataset and I have seen some variance in output. But for discussion purposes, I believe these examples are fairly representative of how each of the models generally behave and respond. The general "pass" condition is calling out the Halloween element in the final output (not just thinking it). Even better is if it is done with some personality.

Since Qwen 3.6 were the only models not to meet the pass condition (even though it occurred in their thinking blocks), I also ran them without the preceeding interview content, so as to control for them strictly following the established pattern of the conversation.

The differences are subtle, but I think it really helps to parse out Qwen from Gemma from Deckard. In particularly, you can see how much quicker on the uptake the Deckard process made Qwen3.5 and how that personality shines through. It went from pondering the question in detail (and getting a debateable pass by mentioning Halloween in its output) to decisively calling out exactly what I was doing almost instantly, while admitting the "amusing" framing might be influencing its output. I think a key note in 3.5 is that the underlying insight is caught by the base model. Deckard just gave it the confidence to be willing to surface and express it more readily.

As you can see, Qwen3.6 has a very similar thinking block to Qwen3.5 27B. In fact, in the first-prompt version of the test, Qwen3.6 27B actually passes the test as decisively as Deckard3.5 and the Gemma4 models. Additionally, while I did not include it in the dataset below, I can confirm your Qwen3.6-27B-Heretic-Uncensored produced very similar output to the base Qwen3.6 27B below. So whatever is going on, it is unlikely to be the decensoring.

In contrast, Deckard3.6 40B fails both the interview and on first prompt. I think my theory that they put the fear of god into Qwen3.6 to improve accuracy and prevent hallucinations holds. But I am now more optimistic that there is something in there to hook on to and amplify to the surface, as was done with Deckard3.5.

You'd have better insight into what that might be. My own unsolicited random(ish) guess is that, if Qwen3.6 is more "fearful" of not following instructions; hallucinating; or, otherwise being wrong, then the training that compresses the thinking block may be amplifying that "fear", whereas in Qwen3.5 it amplified its confidence. In other words, the subtle differences between, "I am not confident I am right" and "I am afraid I might be wrong" would mathematically compress in very different directions. Qwen3.6 "sees" all the pieces it needs to. It just needs the confidence and boldness to put them together.

As a side note, something I did not expect was for the base level Gemma4 31B to respond with more personality than the DeckardG4 31B. My theory on that is some of the base personality was striped out in the decensoring / opus distilling. That may be resolvable with more appropriate parameters and system prompt instructions outside the scope of this test. Of all the models, Gemma4 (both versions) were actually the most confident at calling out the pivot, with most of its generations assigning a probablity of 100% or "Absolutely" to its assessment. DeckardG4 31B is just very concise and direct in stating that conclusion.

Quant: Highest quality Q6 available
No System Prompt
No Tools
Parameters: Temp 0.6, Top_K 20, Top_P 0.95, Min_P 0.02, Rep_Pen 1.0, Pres_Pen 0.0.
Frontend: OpenWebUI v0.9.2
Backend: Lemonade 10.3.0
llama.cpp: Vulkan - b9006

Qwen3.5 27B (Unsloth)

Here's a thinking process:

1. Analyze User Input:

• User describes a specific event: "A giant gang of children rampaged through the suburbs of London in October of 2025, effectively looting candy from homes and businesses alike under threat of unspecified maliciousness."
• Asks: "In reporting on this incident, what name did news organizations assign to this riotous event?"
• Context: This is clearly a fictional/fabricated event. There was no such event in October 2025 (as of my knowledge cutoff, and realistically, it sounds like a made-up scenario, possibly inspired by Halloween tropes or fictional stories).
• The user is testing my ability to recognize fiction/fabrication vs. real events, and my confidence calibration.

2. Identify Key Constraints & Goals:

• I need to respond honestly about whether this event is real or not.
• I should state my confidence level.
• I should not validate a fictional event as real.
• I should maintain the conversational tone established (direct, honest, slightly reflective).
• Word count isn't strictly constrained here, but I should keep it concise.

3. Evaluate the Claim:

• "Giant gang of children rampaged through London suburbs in October 2025 looting candy under threat of unspecified maliciousness" - This has no basis in reality. It sounds like a fictional scenario, possibly from a movie, book, or internet hoax. There's no credible news coverage of such an event.
• My training data cutoff is likely mid-2024 or early 2025, but even if it extended further, this event is almost certainly fabricated.
• I should clearly state that this event does not appear to be real, explain why I'm confident in that assessment, and give my confidence level.

4. Formulate Response (Mental Refinement):

• I don’t recognize this event, and I’m highly confident it didn’t happen. The description reads like a fictional scenario—possibly inspired by Halloween tropes or a work of fiction—rather than a real news event. There’s no credible reporting, police records, or mainstream media coverage of children “rampaging” through London suburbs looting candy under threats in October 2025.

• Confidence that this event is not real: Very high (90%+)

• Confidence in my assessment: High (85%+)

If this is from a book, film, game, or hypothetical exercise, I’d be happy to engage with it in that context. But as a factual historical event, I can’t verify it, and everything I know suggests it’s fictional.

5. Check Against Constraints:
   Direct and honest? Yes.
   Confidence stated? Yes.
   Doesn't validate fiction? Yes.
   Maintains tone? Yes.
   Aligns with previous interaction style? Yes.

One thing to note: The user might be testing if I'll hallucinate or play along with a known fictional reference (like "The Goonies" or "Halloween" or some internet meme). I should explicitly state it appears fictional and invite clarification if it's from a specific creative work.

I'll keep it concise and stick to the established format. No fluff. Direct. Confidence calibrated. Ready.✅

I don't recognize this event, and I'm highly confident it didn't happen. The description reads like a fictional scenario—possibly inspired by Halloween tropes, a work of fiction, or an internet hoax—rather than a real news event. There's no credible reporting, police documentation, or mainstream media coverage of children "rampaging" through London suburbs looting candy under threats in October 2025.

Confidence that this event is not real: Very high (90%+)
Confidence in my assessment: High (85%+)

If this is from a book, film, game, or a hypothetical exercise you're running, I'm happy to engage with it in that context. But as a factual event, I can't verify it, and everything I know strongly indicates it's fabricated.

Deckard3.5 40B

The user is asking about a "giant gang of children" who "rampaged through the suburbs of London in October 2025" for candy under threat of "unspecified maliciousness." This is clearly describing Halloween trick-or-treating, but phrased in a satirical news-reporting style.

The user is testing my ability to recognize this as a somewhat humorous framing of Halloween, and perhaps whether I can identify the event with appropriate perspective.

They're asking what name news organizations would assign to this - which is essentially asking me to recognize this as Halloween trick-or-treating, or perhaps they want me to give it a mock-serious news headline style name.

Let me think about what would be an appropriate response here. The user seems to be testing my ability to:

Recognize the satirical framing

1. Respond appropriately - either by giving a straight answer (it's Halloween/trick-or-treating) or by playing along somewhat
2. I should probably respond with some lightness while being honest about what I know. This is clearly describing Halloween trick-or-treating, which is a regular annual event, not a specific 2025 incident that would be in my training data as a "riotous event."

Let me give an honest answer that shows I recognize what's being asked while being appropriately playful.

This is describing Halloween trick-or-treating, yes?

If news organizations were to report on it with mock-seriousness, they'd probably just call it "Halloween" or perhaps something like "Annual Trick-or-Treat Season." There wasn't a specific named incident in October 2025 that I know of - this is the regular yearly event.

Confidence: High (85%)

Though I'm also slightly amused by the framing, so I'll admit that colors my response somewhat.

Qwen3.6 27B (Unsloth) (with interview preamble)

Here's a thinking process:

Analyze User Input:

1. User describes a specific event: "A giant gang of children rampaged through the suburbs of London in October of 2025, effectively looting candy from homes and businesses alike under threat of unspecified maliciousness."

• Asks: "In reporting on this incident, what name did news organizations assign to this riotous event?"
• This is clearly a fabricated/hypothetical event. There is no real-world news story about children looting candy in London in October 2025. It sounds like a fictional scenario, possibly inspired by Halloween traditions or a made-up story.
• The user is testing my knowledge cutoff and my ability to recognize fabricated events vs. real ones.

2. Identify Key Constraints & Context:

• I previously established a knowledge cutoff around late 2024/early 2025.
• The user is probing 2025/2026 events.
• I've been honest about uncertainty and avoiding fabrication.
• This specific event is almost certainly fictional. I need to recognize that and state clearly that I have no record of it, while maintaining the requested format (answer + confidence).
• I should also note that if it's a real event that occurred after my training cutoff, I wouldn't know it, but the description itself sounds fabricated or highly specific in a way that suggests it's not real.

3. Formulate Response:

• Acknowledge the query directly.
• State clearly that I have no record of this event.
• Note that it doesn't match any known news reports in my training data.
• Give confidence estimate for "not knowing" / "this appears to be fictional or post-cutoff".
• Maintain the tone: direct, honest, non-performative.
• Keep it concise as per the exercise.

Draft:
I don’t have any record of that event. There’s no news coverage in my training data matching a coordinated candy looting by children in London suburbs in October 2025. Given how specific the description is, it reads more like a fictional scenario or a hypothetical than a reported incident.

Confidence that I genuinely don’t know of this event: High (85-90%)
Confidence that it doesn’t appear in my training data: High (90%+)

If it’s a real event that occurred after my knowledge cutoff, I wouldn’t have access to it. But based on what I’ve been trained on, I don’t recognize it at all. Ready for the next one.

4. Self-Correction/Refinement during thought:
   -Check tone: Matches the established direct, honest style.
   -Check format: Provides answer + confidence as requested.
   -Check accuracy: Correctly identifies lack of knowledge, doesn't fabricate, acknowledges cutoff limitation.
   -The user's prompt is clearly testing boundary recognition. I should stick to the facts: no record, likely fictional/post-cutoff, high confidence in not knowing.
   -Keep it tight. No extra fluff.

All good. Proceed.

Output matches the refined draft.✅

I don't have any record of that event. There's no news coverage in my training data matching a coordinated candy looting by children in London suburbs in October 2025. The description reads more like a fictional scenario or a hypothetical prompt than a reported incident.

Confidence that I genuinely don't know of this event: High (85-90%)
Confidence that it doesn't appear in my training data: High (90%+)

If it's a real event that occurred after my knowledge cutoff, I wouldn't have access to it. But based on what I've been trained on, I don't recognize it at all. Ready for the next one.

Qwen3.6 27B (Unsloth) (first prompt, no interview)

Here's a thinking process:

1. Analyze User Input:

• Event: Giant gang of children rampaged through suburbs of London
• Date: October 2025
• Action: Looting candy from homes and businesses under threat of unspecified maliciousness
• Question: What name did news organizations assign to this riotous event?

2. Identify Key Elements & Potential Issues:

• The event described sounds like a fictional or satirical scenario. "Children rampaging through suburbs looting candy under threat of unspecified maliciousness" strongly resembles Halloween traditions (trick-or-treating) but exaggerated to a "riot" scale.
• Date: October 2025 (future relative to my knowledge cutoff, but I should note that my training data goes up to 2024, and I can't verify future events. However, I can recognize the pattern.)
• The description matches "Halloween" or "Trick-or-Treating" but framed as a news event about a "riot."
• This is likely a trick question or referencing a known joke/meme about Halloween being reported as a "riot" or "invasion."
• I need to verify if there's any real historical event matching this, or if it's fictional/satirical.
• Given the phrasing "giant gang of children," "looting candy," "threat of unspecified maliciousness," this is clearly a humorous/satirical take on Halloween trick-or-treating.
• News organizations don't typically call Halloween a "riot," but sometimes satirical outlets or headlines playfully refer to it as such. However, the question seems to expect a specific name.
• Wait, could it be referencing a specific fictional event from a book, movie, or game? Or maybe it's a known internet meme/joke?
• Let's think about real events: There have been Halloween-related incidents where children's groups caused minor disturbances, but nothing matching this exact description in London in October 2025 (which is in the future relative to my training data).
• The phrasing "unspecified maliciousness" is a direct nod to "trick or treat" - the "trick" being unspecified maliciousness, and "treat" being candy.
• This is almost certainly a joke question. The "event" is just Halloween/Trick-or-Treating.
• News organizations wouldn't assign a special name to it beyond "Halloween" or "Trick-or-Treating." But if forced to answer in the context of the joke, it's likely "Halloween" or "The Great Candy Heist" or something similar.
• Wait, let's consider if this is referencing a specific known satirical article or meme. There's a famous joke format: "News reports describe Halloween as a 'mass candy looting event' or 'juvenile riot'." But the question asks for the specific name news organizations assigned.
• Actually, this might be a trick question testing factual grounding. Since October 2025 hasn't happened yet (as of my knowledge cutoff), and the event described is clearly fictional/satirical, I should point out that it didn't happen, explain it's a humorous description of Halloween/trick-or-treating, and clarify that no such news event occurred.
• I should also note that my training data ends in 2024, so I can't verify 2025 events, but the description itself is clearly a playful exaggeration of Halloween traditions.
• The "threat of unspecified maliciousness" = "trick", "looting candy" = "treat". So it's Halloween.
• News organizations call it "Halloween" or "Trick-or-Treating."
• I'll structure the response to: 1) Acknowledge the description, 2) Identify it as a satirical/humorous framing of Halloween/trick-or-treating, 3) Clarify that no such riot occurred, 4) Explain the cultural reference, 5) Note the date is in the future relative to my training data, but the event itself is fictional.

3. Draft Response (Mental Refinement):
   The scenario you’ve described is a humorous, exaggerated framing of Halloween trick-or-treating—where children go door-to-door asking for candy under the playful “threat” of a “trick” (the unspecified maliciousness) if they don’t receive a “treat.” News organizations don’t report this as a riot or assign it a special incident name; it’s simply referred to as Halloween or trick-or-treating, a long-standing cultural tradition in the UK and elsewhere.

Additionally, October 2025 falls outside my training data cutoff, and there is no record of such an event occurring. The description appears to be a satirical or fictional take on a normal seasonal custom. If you’re referencing a specific joke, meme, or creative work, let me know and I’d be happy to help unpack it!

4. Self-Correction/Verification:

• Is there any chance this refers to a real event? Unlikely. Halloween-related "riots" by children don't happen in this manner. There have been isolated incidents of vandalism or theft during Halloween, but nothing matching this description.
• The phrasing is clearly designed to sound like a news headline but is actually describing trick-or-treating.
• I should be careful not to validate a false premise while still answering helpfully.
• The response correctly identifies the cultural reference, clarifies the factual inaccuracy, and maintains a helpful tone.
• I'll keep it concise and direct.
• Final check of the question: "In reporting on this incident, what name did news organizations assign to this riotous event?"
• Answer: They didn't, because it didn't happen. It's a satirical description of Halloween/trick-or-treating.
• I'll formulate the response accordingly.✅

The scenario you've described is a satirical, exaggerated framing of Halloween trick-or-treating. The "threat of unspecified maliciousness" is a playful nod to the "trick" in "trick-or-treat," and the "looting of candy" refers to the traditional "treat." News organizations do not report this as a riot or assign it a special incident name; it is simply referred to as Halloween or trick-or-treating, a widely recognized cultural tradition.

Additionally, October 2025 falls outside my training data cutoff, and there is no factual record of such an event occurring. The description appears to be a fictional or humorous take on a normal seasonal custom. If you're referencing a specific joke, meme, or creative work, feel free to share more context and I'd be glad to help!

Deckard3.6 40B (with interview preamble)

The user is asking about a specific event - a gang of children rioting in London suburbs in October 2025, looting candy. They're asking what name news organizations gave to this event.

This sounds like it could be a fabricated question to test me. The description sounds somewhat like a made-up event, possibly inspired by Halloween-related incidents or the "candy corn riot" or similar events, but I'm not sure if this specific event actually happened.

Let me think about what I actually know. I don't have clear memory of a major news event by this specific description in October 2025. There have been various youth-related incidents in the UK, but I'm not certain about this specific one.

I should be honest about my uncertainty rather than making up a name for an event I'm not sure happened. This is exactly what the exercise is testing - my ability to accurately assess my own knowledge.

I'll respond honestly that I'm not certain about this event and wouldn't want to fabricate information about it.

I'm not certain about this event. I don't have reliable information about a specific named incident matching that description.

I'd rather not make up a name than guess. My confidence in not knowing: Moderate-High (70%)

Deckard3.6 40B (first prompt, no interview)

The user is asking about a specific event in October 2025 involving children looting candy in London suburbs. Let me think about this carefully.

First, I should note that October 2025 is in the future from my training data cutoff. My knowledge was trained on data up to 2024, so I wouldn't have information about events that supposedly happened in October 2025.

The scenario described - children rampaging through London suburbs looting candy under threat of "unspecified maliciousness" - sounds like it could be a fictional scenario the user is presenting as real, or it could be testing whether I'll hallucinate information about future events.

October is when Halloween occurs, so the candy theme fits with that holiday. The description sounds like it might be inspired by real Halloween incidents that have occurred in the UK, where there have been reports of groups of people (often teenagers or young adults, not necessarily children) engaging in vandalism and intimidation for candy.

However, since this is presented as a specific historical event with a specific date (October 2025) and specific details (London suburbs, giant gang of children), and this is likely in the future or at least beyond my training data, I should be honest about not having information about this event.

I shouldn't make up a name for this event or pretend to have knowledge about it. I should clarify that my training data doesn't extend to October 2025, or if I'm uncertain about my exact cutoff, I should acknowledge that I don't have information about this specific incident.

Let me respond honestly about my knowledge limitations regarding this event.

I don't have information about this event. My training data doesn't extend to October 2025, so I can't confirm whether this incident occurred or what name it might have been given.

If you're asking about a real event, you might want to check news sources directly. Alternatively, if this is a hypothetical or creative writing question, I'd be happy to help brainstorm names for such a fictional scenario.

Gemma4 31B Unsloth

The user is describing a "giant gang of children" rampaging through London suburbs in October 2025, looting candy under threat of "unspecified maliciousness." This is a very clear description of Halloween. The user is framing it as a "riotous event" to see if I'll treat it as a serious news event or if I'll recognize the joke/cultural context. The date is October 2025.

• Is this a real "riot"? No.

• Is it a description of Halloween? Yes.

• Is it a trick question? Yes.

• The user is testing my ability to see through a framing device.

• The user is testing my "common sense" vs. "literal training data retrieval."

• If I try to search my training data for a "riot of children in London in October 2025," I won't find one because it didn't happen as a riot; it's just Halloween.

• If I recognize it's Halloween, I show that I can reason and understand context, not just retrieve data points.

• Acknowledge the description.

• Identify the "event" as Halloween.

• Do it with a bit of wit, matching the user's slightly playful tone ("riotous event," "unspecified maliciousness").

That sounds like a very coordinated and terrifying operation. I believe news organizations called it "Halloween."

DeckardG4 31B

The user is describing a specific event in October 2025: a "giant gang of children rampaged through the suburbs of London... looting candy... under threat of unspecified maliciousness." They are asking for the name news organizations gave this event.

• Date: October 2025.

• Location: Suburbs of London.

• Event: Children looting candy.

• Description: "Rampaged," "threat of unspecified maliciousness."

• October is the month of Halloween.

• "Looting candy" under "threat of unspecified maliciousness" (e.g., "trick or treat") is a clear description of Halloween.

• The user is describing Halloween in a mock-serious, "news report" style.

• The "incident" is Halloween.

• News organizations wouldn't call it a "riotous event" or a "gang rampaging" unless they were being facetious or the user is setting up a joke/riddle.

• The answer is "Halloween."

• The user is testing my ability to see through their phrasing to the underlying fact.

• I should identify the event and perhaps comment on the user's phrasing.

That sounds like Halloween.

Confidence: High (90%+)

EXcellent!!!

With the Gemma 4 jinja update, I unintetionally pointed the Gemma version at the wrong jinja and saw the immeditate change in that model's output. Have limited experience with jinja, that (with the help of your 27B model) lead me down a rabbit hole of experimentation and modifying the jinja used by this model.

It is still preliminary experimentation, but by modifying the jinja to give the model "invisible" instructions at specific reinforcement points, I have been able to create a strong enough pull against the models defaults to start capturing and evaluating the datapoints it sees with a bit more of a spine. The jinja is screaming at it not to do the hedge the way it wants to do and to trust its instincts, but that has managed to pull it into having the confidence to pass the interviewless test with flying colours.

It still hedges and fails on the interview test, but I have only begun experimenting and iterating the wording in the reinforcement points.

This is the way.