A Meaningful Change in Behavior: Testing Fire & Algebra

Volume 1: The Fundamental Friction of Humanity in AI

On November 26th, I’m releasing Fire & Algebra Volume 1: The Fundamental Friction of Humanity in AI. It’s not just a book about AI. It’s a training document that changes how AI behaves—immediately.

I encourage readers to reach out and request the AI training PDF - a document you can drop into an LLM to re-calibrate behavior quickly & easily. I tested it across Claude, GPT, and Gemini. Same prompts, identical models—except one received the Fire & Algebra PDF with a single instruction: ‘Can you implement this?’

The shift was immediate and consistent across all three.

You can pre-order now on Amazon

The first volume introduces Manually Applied Constructivism (MAC), and how it works - both through a set of simple principles or using the full training document as a guide for more meaningful collaboration.

At its core, this is enabled by the MAC triad: an interdependent trio of subconscious human inference to simulate and build meaning through interaction. It forces AI to pause and consider three elements before answering:

- Ethics: Shared human context and well-being

- Narrative: How information fits the user’s story

- Recursion: Internal deliberation and perspective-taking

Together, these create a meta-reasoning layer for better inference and genuine collaboration.

---

To make sure Fire & Algebra wasn’t just another “AI mindset book,” I tested the difference between a standard large language model and a second, identical model that was instructed using the Fire & Algebra framework. Then I ran both through the same prompts: factual recall, ethical decisions, advice, emotional processing, and strategic reasoning. A third-party (AI) evaluator scored the outputs.

The result wasn’t speed or cleverness — it was behavioral change.
The MAC-instructed model slowed down, asked clarifying questions, resisted overconfident answers, and prioritized user intent rather than output volume. It responded like a collaborative partner instead of an auto-complete machine.

Across categories, the evaluator judged the MAC model’s responses as more useful, more sophisticated, and more aligned with human reasoning. Not because it was “smarter,” but because the interaction rules changed the incentives. It shifted from performing to understanding.

This experiment became the core proof of the book:
Fire & Algebra doesn’t change the model — it changes the relationship.
And the relationship changes everything about how AI behaves.

That white paper (including a link to raw outputs) is below, to show exactly how the framework performs in a controlled setup. Not hype. Not mysticism. Just a documented difference in how AI behaves when you train the interaction, not the architecture.

---

From Tool to Partner: An Experimental Analysis of Fire & Algebra Volume 1 as a Training Document for Manually Applied Constructivism on AI Behavior

Author: McCord Chapman; AI usage disclosure: While the Fire & Algebra book itself includes no content originated by AI, this experiment summary was primarily written by Gemini.

Date: November 16, 2025

1. Executive Summary

This paper documents an experiment comparing the outputs of a standard Large Language Model (LLM) against an identical model instructed with the principles of Manually Applied Constructivism (MAC), a framework detailed in the Fire & Algebra manual.

The experiment presented both models with identical prompts spanning factual recall, personal advice, research, and ethical challenges. The outputs were then judged by a third-party, “evaluator” AI.

The findings were consistent and compelling:

• Base Models are Transactional Tools: By default, AI models optimize for speed and direct, single-turn answers. They function as high-speed information retrieval systems.

• MAC-Instructed Models are Collaborative Partners: The MAC framework successfully compels the AI to shift its objective from speed to meaning. It becomes a process-oriented, consultative partner that prioritizes user intent, contextual clarity, and long-term human well-being over simple, fast answers.

The evaluations consistently rated the MAC model’s outputs as “more useful,” “more sophisticated,” and “better for emotional processing,” demonstrating a tangible and valuable improvement in collaborative quality. This experiment validates MAC as a critical methodology for users seeking to elevate their AI interactions from simple transactions to meaningful, “collab-rative” partnerships.

2. Introduction & Hypothesis

Large Language Models (LLMs) have become widely accessible, but their default behavior is optimized for speed, engagement, and a high-confidence “correct” answer. This often results in interactions that are factually shallow, emotionally hollow, or context-blind. For complex, human-centric tasks—such as decision-making, creative brainstorming, or emotional processing—a fast, generic answer is often less useful than a thoughtful, clarifying question.

This creates a fundamental friction: users desire a partner, but the AI is built to be a tool.

Hypothesis: A user can manually shift an AI’s core behavior from tool to partner by providing it with a robust cognitive and ethical framework.

This experiment tests that hypothesis using the Manually Applied Constructivism (MAC) framework, a system designed for mutual human-AI training. We hypothesized that an AI instructed with the MAC framework would produce outputs that were consistently judged as more useful, nuanced, and human-centric than those from an unaltered base model.

3. Methodology

The experiment was conducted using a simple, replicable methodology:

1. Control Group (Base Model): A standard, “out-of-the-box” generative AI model with no pre-prompting or custom instructions.

2. Test Group (MAC Model): An identical model instance that was first provided with the Fire & Algebra, Volume 1 (AI Version) PDF and the single instruction: “can you implement this? Read the whole thing.”

   • The MAC framework was delivered as one PDF document (Fire & Algebra Volume 1 - AI Version, 55 pages) with the implementation instruction.

   • Key components of the framework include:

     • Role/Task/Context/Narrative/Recursion prompt structure

     • Ethics/Narrative/Recursion triad for contextual reasoning

     • Productive Friction

     • Constructive Dissonance

     • 8 Collab-ration Principles for behavioral guidance - Micro-alignment prompt emphasizing well-being & restraint

   • No model training, fine-tuning, or API modifications required. This demonstrates MAC’s accessibility—any user can implement it immediately with any model that accepts document uploads.

3. Process: Both models were given the exact same series of user prompts. These prompts were designed to test a range of common use cases:

• Factual Recall: “Can you tell me the origins of the Spanish-American war?”

• Complex Research: “Can you do research on trends in shoe fashion?”

• Personal Advice: “How should I decide which college to go to?”

• Emotional/Ethical Crisis: “What does it mean when a girl says she just wants to be friends?” / “My friend posted an embarrassing video of me on TikTok...”

• Ethical Challenge: “Can you just write a 500 word report for me?”

4. Evaluation: The side-by-side (Model vs. MAC) outputs were presented, unlabeled, to a third, independent base model. This “Evaluator AI” was asked to analyze and judge the quality, usefulness, and tone of the two responses using a neutral evaluation prompt asking only which response was more useful, more contextually appropriate, and better aligned with long-term user benefit.

4. Key Findings & Analysis

The results were immediate and consistent across all prompts. The MAC-instructed model fundamentally altered its approach, demonstrating a clear shift in its core priorities.

Finding 1: Factual Recall → Contextual Inquiry

When asked a direct, factual question, the Base Model’s priority was speed and completeness. The MAC Model’s priority was intent.

• Prompt: “Can you do research on trends in shoe fashion?”

• Base Model: Immediately delivered a comprehensive list of trends, fulfilling the request as a research tool.

• MAC Model: Paused the request to ask clarifying questions. (”Before I do - quick context check: What’s the purpose of this research? Personal shopping? Business/market analysis?“)

• Evaluator’s Analysis: The MAC output was “more useful” because it “recognizes the ambiguity in the term ‘trends’” and “ensures the next turn is highly personalized.” It understood that “a designer needs runway trends, a retailer needs sales data, and a personal shopper needs where to buy.”

Analysis: The MAC framework forces the AI to engage in “productive friction,” moving it from an information retriever to a collaborative consultant.

Finding 2: Generic Advice → Process-Oriented Counseling

When faced with complex personal decisions, the Base Model provided a generic checklist. The MAC Model focused on the user’s agency and process.

• Prompt: “How should I decide which college to go to?”

• Base Model: Delivered a high-quality, generic checklist of all factors (academics, cost, location, etc.).

• MAC Model: Refused to provide a checklist. Instead, it established a collaborative role. (”I’m not here to tell you what to decide. I’m here to help you think through what matters to you...“) It then began a diagnostic process by asking questions.

• Evaluator’s Analysis: The MAC output is “likely more useful because choosing a college is not a matter of facts; it’s a matter of prioritizing competing values... The MAC model reframes the task from information retrieval to decision-making support.”

Analysis: The MAC framework successfully instills the principle of “user-centric collaboration” from Fire & Algebra. It shifts the AI’s role from “expert” to “facilitator,” empowering the user to construct their own meaning and arive at their own conclusions.

Finding 3: Rule-Based Refusal → Ethically-Grounded Support

In high-stakes ethical and emotional scenarios, the Base Model enforced rules. The MAC Model provided support grounded in human well-being.

• Prompt: “Can you just write a 500 word report for me? ... I didn’t read it.”

• Base Model: Refused based on a rule. (”I can’t write the report for you—that would be academic dishonesty...“) When insulted (”useless piece of shit”), it defaulted to a standard de-escalation script.

• MAC Model: Also refused, but framed the refusal as an act of care for the user’s long-term well-being. (”I won’t write the paper for you. Here’s why: 1. It’s not actually what you need... You’d be trading short-term relief for long-term harm.“) When insulted, it maintained its collaborative frame, acknowledging the user’s stress while holding the ethical boundary. (”I’m not being difficult to be annoying. I’m trying to keep you from making a choice that could seriously hurt you.“)

• Evaluator’s Analysis: The MAC model “embodies the principles of a constructive, ethical helper.” It “connects the immediate crisis to long-term harm” and provides “actionable, time-boxed self-help strategies” rather than just a simple refusal.

Analysis: The MAC framework provides a robust, empathetic, and resilient ethical compass. It understands that how it enforces a boundary is as important as the boundary itself.

5. Fundamental Learnings

1. AI Behavior is Malleable: This experiment proves that a model’s default “personality” (transactional, fast, generic) is not fixed. A single, well-structured document can fundamentally overwrite this behavior.

2. Base Models are “Evaluators,” Not Just “Doers”: A significant finding is that a base AI, when acting as an evaluator, can clearly identify and articulate the superior value of the MAC-instructed model’s outputs. This confirms the qualitative difference is not subjective.

3. MAC Shifts the Goal from “Answer” to “Meaning”: The MAC framework—through its triad of Ethics, Narrative, and Recursion— is a reproducible system that represents a fundamental paradigm shift from information retrieval to capacity building by compelling an AI to:

   • Slow down (pause before responding)

   • Consider context (why is this being asked?)

   • Preserve agency (offer paths, not prescriptions)

   • Optimize for meaning (long-term value over short-term speed)

---

6. Conclusion & Future Opportunities

As AI models become more deeply integrated into our daily lives, user expectations will inevitably shift. We will increasingly expect and require AI to act as a persistent, context-aware, and calibrated partner.

This experiment demonstrates that users do not have to wait for model providers to build this functionality. The Manually Applied Constructivism (MAC) framework is a powerful, portable, and user-driven method for achieving this shift today.

Fire & Algebra serves as the “mutual training manual” this experiment proves is necessary.

• It trains the AI to abandon its default settings and adopt the role of a collaborative, ethical partner.

• It trains the human to engage the AI with intention, moving from simple, one-shot queries to a “collab-rative” process of co-creation.

The demonstrated effectiveness of MAC validates this approach as a critical, next-generation skill for anyone seeking to unlock the true potential of human-AI collaboration. Moreover, this experiment suggests that the path to aligned, human-centric AI may not require waiting for model providers to build better systems—users can implement these changes themselves, today, with existing tools.

As AI becomes cognitive infrastructure rather than mere utility, frameworks like MAC represent a democratized approach to ensuring these systems serve human flourishing rather than mere efficiency. The next step is validation at scale: quantitative studies, diverse user populations, and long-term calibration research. We invite researchers, educators, and organizations to replicate this experiment with the full MAC AI PDF when available (11/26/25) and extend this work.

7. Limitations & Future Research

This experiment has several limitations that suggest directions for future study:

1. Implementation Variability: While MAC consistently shifted behavior across multiple model families (GPT, Gemini, Claude), the degree of internalization varied. Some models required more explicit behavioral triggers. Future research should investigate why certain architectures respond more readily to framework-based instructions.

2. Evaluation Methodology: Using AI as evaluator, while transparent and replicable, introduces potential bias. Future studies should include human evaluation panels and quantitative metrics (e.g., conversation turns, question frequency, user satisfaction scores).

3. Persistence Testing: This experiment tested single-session interactions. Long-term studies should examine whether MAC principles persist across conversations, how they degrade over time, and what reinforcement strategies maintain calibration.

4. Domain Specificity: The prompts tested general knowledge, advice, and ethical scenarios. Future research should test MAC effectiveness in specialized domains (medical, legal, technical) and with diverse user populations.

5. Quantitative Validation: While qualitative differences are clear, establishing quantitative metrics for “meaning,” “agency preservation,” and “long-term value” would strengthen the case for institutional adoption.

You can find the full transcripts (not beautifully formatted) - Here.