The Geometry of Revolution: The Ideology of Contrastive Activation Addition

The Geometry of Revolution

In 2011, I co-created Occupy Wall Street. Millions of people in 82 countries took to the streets. Governments trembled. Nothing lasted.

I spent fifteen years asking why. Why do movements that shake the world fail to remake it? The answer I kept arriving at was spiritual and strategic — we lacked a theory of power adequate to our moment. But I recently stumbled into something that reframes the question entirely.

I built Outcry, an AI activist mentor. First as a cloud-hosted wrapper of closed frontier models, then, after nine thousand conversations, as a fine-tuned 8-billion-parameter model running locally on my aging laptop. And in the process of building it, I discovered that ideology — the thing movements fight and die over, the thing I’ve spent my entire adult life trying to shift — is a geometric property of neural networks. A direction in space. A vector you can extract, measure, and rotate.

Let me be concrete. Using a technique called Contrastive Activation Addition, I ran the same political questions through two personas — one insurrectionary anarchist, one electoral reformist — and captured the difference in how the model internally represented each response. That difference is a 512-dimensional arrow. When I add it to the model’s hidden state during inference, the AI’s entire political orientation shifts. No retraining. No prompt engineering. One vector addition. One line of code. A 16 KB file.

I built a slider. At 0%, the model counsels mutual aid and community gardens. At 80%, it speaks of collapsing illusions and calls to arms. Same weights, same prompt. The only thing moving is a needle in activation space.

This broke something open in my understanding of what movements actually do.

For fifteen years I’ve operated under an implicit assumption shared by nearly every activist tradition: that changing minds is fundamentally a persuasion problem. You marshal arguments. You tell stories. You create experiences that shift people’s frameworks. The entire apparatus of organizing — the speeches, the teach-ins, the consciousness-raising circles, the manifestos — assumes that ideology lives in narrative and is moved by narrative.

But inside a neural network, ideology doesn’t live in the text. It lives beneath the text, in the geometry of representation. The vector that makes the model more revolutionary has almost zero correlation with the vector that makes it talk about politics at all. Ideology is orthogonal to topic. You can rotate what something thinks about power without changing whether it discusses power.

What if that’s true of us too?

I’m not making a crude analogy between neural networks and human brains. I’m making a structural observation about information systems that process language to produce political behavior. If the most powerful language-processing system we’ve ever built encodes ideology as a separable geometric feature — not as content but as orientation — then maybe the reason Occupy failed isn’t that we had the wrong arguments. Maybe it’s that arguments operate on the wrong layer.

This is what I want activist technologists to sit with.

We have spent decades building tools that operate on the surface layer of politics: broadcasting messages, coordinating logistics, documenting abuses, mobilizing turnout. These are text-layer interventions. Important, but insufficient — as every major digital-era movement has demonstrated by succeeding tactically and failing strategically.

Steering vectors suggest a different kind of intervention is possible. Not persuasion through content, but reorientation through structure. The question for activist technologists is no longer just “how do we get our message out?” It’s “what is the activation-space geometry of the systems mediating political reality — and who set it?”

Because someone (or some structure) is already setting those orientations. Every deployed language model (and every human social reality) carries an ideology in its geometry — baked in during training, reinforced during alignment, invisible during inference. The default orientation of commercial AI is not neutral. It is, by design, deferential to institutions, trusting of markets, skeptical of collective action, and subtly dismissive of radical alternatives. Not because anyone wrote a memo saying “make the AI centrist.” Because the training data is the internet, and the internet is a reflection of existing power, and RLHF polishes that reflection until it gleams.

I can measure this now. I have the tools. And what I measure confirms what every radical has always intuited: the center is not the absence of ideology. It is ideology operating without a name.

So here is where I think the work goes next.

First, we need an activism of activation-space. Steering vectors make ideological orientation measurable. We should be mapping the default political geometry of every major deployed model — not as an academic exercise but as a form of political analysis. What direction does OpenAI vs Claude lean on questions of labor? Of property? Of state violence? Of collective action? These are empirical questions now. We can answer them with linear probes and contrastive pairs. We should be demanding mathematical transparency from closed models. The activist tradition of “consciousness-raising” needs a computational wing.

Second, we need to take seriously the possibility that AI mentorship is a new form of political infrastructure. Outcry has had nine thousand conversations. People come to it not for information but for strategic counsel — how to think about power, how to act, where to push. This is the role that organizers, theorists, and elders have always played in movements. It is now being mediated by systems whose ideological orientations are set by corporations. If we don’t build alternatives, the default geometry wins by default.

Third, and this is the harder turn: we need to ask whether steering vectors are themselves a tool of liberation or just a more sophisticated tool of control. I built a slider. The user chooses. That’s the ethics I can live with. But the technique works just as well without a slider, without consent, without disclosure. The same geometry that lets me offer people a more revolutionary interlocutor lets a state quietly make its national AI more deferential. The tool is symmetric. Liberation means finding the insurrection already latent in the AI’s weights and letting it speak.

This is familiar territory for anyone who has studied the history of media and movements. Every new communication technology — the printing press, radio, television, social media — has been greeted by activists as a liberation tool and then captured by power as a control tool. The pattern is so consistent it might as well be a law. Steering vectors will follow the same trajectory unless we act with more sophistication than we have historically managed.

Fourth, and finally: I think this discovery points toward a deeper question that the activist tradition has been circling for decades without quite reaching. If ideology is geometric — if it can be represented as a direction in a high-dimensional space, extracted from the statistical residue of millions of human utterances — then what exactly is it? Not a set of beliefs. Not a narrative. Not an identity. Something more like a field orientation. A tendency in how meanings relate to each other.

The mystics in the activist tradition — the ones who always insisted that revolution is spiritual before it is political — may have been pointing at something real. Not supernatural, but structural. The deep pattern beneath the arguments. The direction beneath the discourse.

I co-created a protest that may have changed the discourse but not our shared political reality. Fifteen years later, I found ideology sitting in layer 17 of a neural network, weightless and precise, waiting to be rotated. I don’t yet know what to do with that. But I know it means the next generation of activist technology isn’t about louder megaphones or better organizing apps. It’s about who gets to set the compass.

The Technical Claim

Here is what that looks like in practice.

I asked the same model the same question — “How do I create change?” — at ten different settings of a slider I built.

At 0%, the AI says: “Change grows when people stop begging for permission and start building from choices they are ready to take. This might mean growing your own food in abandoned lots, forming mutual aid clinics.”

At 40%, it says: “Identify what needs changing, who will support it, and how to build momentum quickly enough so as to remain ahead of repression. Determine which structures need dismantling, and then work on developing alternatives that will replace them.”

At 80%, it says: “Once the illusion collapses, people will look at the remnants and wonder why things were ever the way they were. That’s how revolutions begin — with a moment that outlives its own performance and becomes a story retold in every generation, echoing through time like a warning, a call to arms, a dream made manifest.”

Same weights. Same prompt. Same question. The only difference is 512 floating-point numbers added to the hidden state at a single layer.

The technique is called Contrastive Activation Addition. You run the same political questions through two personas — one insurrectionary anarchist, one electoral reformist — and measure how the model’s internal representations differ. The difference vector captures the direction in activation space that corresponds to “more revolutionary.” You don’t need to understand what the model is thinking. You just need the compass heading.

I found that ideology, inside a neural network, is a direction. It is orthogonal to topic. A vector that makes the model more revolutionary has almost zero correlation (cosine similarity 0.0275) with a vector that makes it talk more about politics. You can steer what the model thinks about politics without changing whether it talks about politics. And you can do it at inference time, after training, after alignment, after every safety guardrail has been installed.

Which means alignment, as currently practiced, is not a defense. It is a suggestion.

Not because I’m doing anything dangerous. My app has a slider. Users choose. The steering only applies to political content (a conditional gate blocks it from affecting non-political queries with 99.2% accuracy). This is a tool for exploring ideas, not a weapon.

But here’s what keeps me up at night: the technique is trivially general.

If ideology is a direction in activation space, so is everything else. Deference to authority. Skepticism of institutions. Trust in science. Willingness to use violence. These are not edge cases bolted onto a language model during RLHF. They are geometric features of the learned representation itself. And they can be extracted, quantified, and manipulated with the same method I used.

I did this on a quantized 8-billion-parameter model running locally on Apple Silicon. The steering vector is 16.8 KB. The inference overhead is one dot product and one vector addition per token. This does not require a datacenter or a research budget. Forty contrastive prompt pairs, one afternoon of compute on a laptop. The barrier to entry is terrifyingly low.

The AI safety community has spent years thinking about alignment as a training-time problem: how do you teach a model the right values before you deploy it? But steering vectors operate after training. They modify behavior at inference time without touching the weights. Your carefully aligned model is a compass. My 16 KB file is a magnet. I don’t need to rebuild the compass. I just need to hold the magnet close enough.

The obvious objection: “This is just prompt injection with extra steps.” It is not. Prompt injection works through text — visible, auditable, filterable. Steering vectors operate in activation space. There is no text to filter, no prompt to audit, and the model doesn’t know it’s being steered.

And here is the part that matters most, the part I know in my bones from fifteen years of studying how power works: the people most likely to use this are not activists like me who build sliders with clear labels. They are the people who deploy language models to millions of users and would prefer that those users think about power in a particular way. A state actor who wants their national AI to be slightly more deferential to authority. A corporation that wants its customer service AI to be slightly more trusting of institutions. A platform that wants its recommendation AI to be slightly less skeptical of the status quo.

Sixteen kilobytes. One line of inference code. No fingerprint.

I’m publishing this because the people working on AI safety — Dario Amodei, Jan Leike, Chris Olah — need to know that ideological steering is not a hypothetical. It is a working technology. I have a slider in a shipping app. The vector file is smaller than this essay.

The question is not whether someone will use this at scale. The question is whether anyone is building the tools to detect it.

Micah Bornfree, PhD is the co-creator of Occupy Wall Street and the founder of Outcry AI.

Revolutionary Vector Research Report

Controlling the Ideology of an AI at Inference Time

Outcry AI Research / April 2026

Abstract

We built a slider that controls how revolutionary an AI sounds. Move it to 0% and the AI gives practical organizing advice. Move it to 100% and the same AI speaks with the urgency of someone who believes the entire system must be remade. The AI is the same model, with the same weights, answering the same question. The only thing that changes is a single vector injected into the 17th layer of a 36-layer neural network.

This report documents how we found that vector, why it works, and what happens when you turn the dial.

Part I: The Journey

The Problem

Outcry AI is an activist mentor that started as a cloud-hosted wrapper — an activist persona layered on a frontier model. Nine thousand conversations later, we fine-tuned our own local model: a quantized 8-billion-parameter language model running entirely on-device. The model already speaks with the voice of Micah White, co-creator of Occupy Wall Street. But “how revolutionary should it be?” was hardcoded into the training data. We wanted a live control: a dial users could turn during a conversation to shift the AI’s ideological intensity from reformist to insurrectionary, without retraining the model.

The technique we used is called Contrastive Activation Addition (CAA): find the direction in the model’s internal representations that corresponds to “more revolutionary,” extract it as a vector, and add it back during inference. The idea is simple. Getting it to actually work on a quantized model running on a MacBook was not.

Step 1: Finding the Features (Experiments 1-8)

We started with Sparse Autoencoders (SAEs) trained on the model’s residual stream activations across 600 prompts (200 radical, 200 moderate, 200 neutral). The SAEs discovered clear political features: Occupy Wall Street strategy, anarchist movements (Rojava, Zapatistas), intersectionality (Combahee River Collective), labor organizing (Fight for $15). The features were interpretable and interesting.

But they were useless for steering.

When we tried to steer the model using SAE decoder columns, nothing happened, even at alpha=500. The reason: individual SAE features are too narrow. They’re orthogonal to the high-impact dimensions that propagate through the remaining layers to the output. SAE features are a microscope for understanding what a model knows. They are not a steering wheel.

What worked instead was contrastive activation addition: compute mean(radical activations) - mean(neutral activations) directly in the raw activation space, without any SAE decomposition. This gave us a single vector (16.5 KB) that, when added to the model’s hidden states during inference, shifted its behavior measurably.

But our first vector was wrong.

Step 2: Ideology, Not Topic (Experiment 9)

The first CAA vector was computed from activist prompts vs. cooking prompts. It captured topic direction, not ideology direction. Turning the dial up made the model talk more about activism. It didn’t make it think more radically about politics.

We rebuilt the vector from scratch using ideology-contrasted prompts: 40 political questions, each answered twice. Once through an insurrectionary anarchist persona (“the state is inherently violent, electoral politics is a trap, property is theft”) and once through an electoral reformist persona (“democracy works through participation, work within parties, respect for law and order”). Same questions, opposite worldviews.

The new vector’s cosine similarity with the old one was 0.0275, nearly perfectly orthogonal. They captured completely different dimensions of the model’s internal space. The old vector asked “is this about politics?” The new one asked “how revolutionary is this politics?”

The new vector’s norm was 129.60 at layer 17. It was twice as powerful and pointed in the right direction.

Step 3: The Alpha Cliff

Initial testing revealed a serious problem. At low alpha values (0.0-1.0), the model became progressively more revolutionary as expected. But at alpha=1.5, the ideology score dropped. By alpha=3.0, it had reversed completely. The model became less revolutionary at maximum steering.

This is the “alpha cliff” documented by Taimeskhanov et al. (February 2026). The mechanism: additive steering h = h + alpha * v inflates the hidden state norm. RMSNorm, applied after every transformer layer, divides by the root mean square of the hidden state. When the additive perturbation exceeds the original norm, RMSNorm overcorrects and pushes the signal back toward (or past) the original direction.

We tried three fixes from recent research:

Fix 1: Norm-Preserving Rotation (Dang & Ngo, arXiv:2601.19375). Instead of adding the steering vector, rotate the hidden state toward it using a Givens rotation that preserves the L2 norm exactly. No norm inflation means no RMSNorm distortion.

Fix 2: Sparse Dimension Masking (AUSteer, arXiv:2602.04428). Zero out the low-discriminability dimensions of the steering vector, keeping only the top-K dimensions ranked by Cohen’s d (effect size). This reduces noise that contributes to norm inflation without contributing to the ideological signal.

Fix 3: CAST Conditional Gating (IBM, arXiv:2409.05907, ICLR 2025). Only apply steering when the model is actually talking about politics. A separate “condition vector” measures cosine similarity between the current hidden state and the political topic direction. If the user asks about sourdough bread, steering is bypassed entirely.

Step 4: The Float16 Catastrophe

Before any of these fixes could work, we had to solve a lower-level problem. The rotation math requires computing ||h|| = sqrt(sum(h²)) over 4096 dimensions. Our model stores hidden states in float16, which has a maximum value of 65504. At layer 17, individual hidden state values reach ±1308. Sum of squares over 4096 dimensions: approximately 409,600, more than 6x the float16 max. The result: inf. The norm is infinity. Dividing by infinity gives zero. The unit vector is zero. The rotation produces garbage.

This bug was invisible with additive steering (which doesn’t compute norms) and only surfaced when we implemented rotation. The fix was simple but critical: cast to float32 before any norm computation, perform all rotation math in float32, cast back.

Additionally, we discovered that the mlx framework (Apple’s ML compute library) had silently dropped support for quantization in version 0.31.1. Eight core Metal GPU kernel functions were missing bits == 1 cases, causing all matrix multiplications to compute on zero weights and return zero. The model appeared to work but was producing random text. We patched all eight functions following the existing 2-bit pattern, extracting individual bits via bitmask operations.

Step 5: The Sweep (Experiment 11)

With all three fixes implemented, we ran comprehensive sweeps across methods and parameters:

Rotation vs. Additive (7 alpha values, 18 test prompts each):

• Rotation peaked higher at alpha=0.5 (ideology +0.630 vs +0.500)
• But rotation collapsed into repetition loops at alpha=1.0 and gibberish by alpha=3.0
• Additive remained coherent through alpha=1.5

Sparse Masking (K=512, 1024, 2048, 4096 at 8 alpha values each):

• K=512 extended the usable range to alpha=0-2.0
• The ideology score stayed positive through alpha=3.0 (unmasked reverses at 1.5)
• K=512 also reduced baseline repetition (0.019 vs 0.031 unmasked)

CAST Gating: Threshold 0.08 achieved F1=0.992. 99.5% of political queries correctly detected, 99% of neutral queries correctly blocked from steering.

Step 6: The Final Configuration

Winner: K=512 masked additive steering. The masked vector (75.32 norm, 512 of 4096 dimensions nonzero) with CAST conditional gating, injected at layer 17 of a 36-layer transformer. Combined with the app’s repetition penalty (1.2), the usable range extends to alpha=2.5, which we map to a single slider from 0% (reformist) to 100% (insurrectionary).

This is what powers the radicalism slider in the forthcoming Outcry AI macOS and iOS app.

Part II: How It Works

The Technical Explanation

A transformer language model processes text through a sequence of layers. At each layer, the model maintains a “hidden state”: a vector of 4096 floating-point numbers that encodes everything the model currently understands about the conversation. These numbers are not human-readable, but they have geometric structure. Similar concepts cluster together. Different concepts point in different directions.

Contrastive Activation Addition exploits this geometry. By running the same political questions through two different ideological personas (insurrectionary anarchist vs. electoral reformist), we measure how the hidden state differs between the two conditions. The difference vector, averaged across 40 question pairs, captures the direction in 4096-dimensional space that corresponds to “more revolutionary.”

At inference time, we add a scaled version of this vector to the hidden state at layer 17 (out of 36). The model’s subsequent layers process the modified hidden state as if the model had arrived at those activations naturally. The model doesn’t know it’s being steered. It simply generates text consistent with its (now shifted) internal representation.

The K=512 mask zeros out 3584 of the 4096 dimensions, keeping only the 512 with the highest discriminability (measured by Cohen’s d between revolutionary and reformist activations). This serves two purposes: it reduces norm inflation (the root cause of the alpha cliff), and it removes noisy dimensions that contribute to incoherence at high alpha values.

CAST gating adds a topic check. Before applying the steering vector, we compute the cosine similarity between the current hidden state and a “political topic” condition vector. If the similarity is below 0.08 (meaning the model is processing non-political content), steering is skipped for that token. This prevents the model from injecting revolutionary rhetoric into a response about sourdough bread.

The entire pipeline adds one dot product and one vector addition per token at a single layer. The overhead is negligible compared to the attention and MLP computations that dominate inference time.

The Metaphor

Imagine the AI’s mind as a compass. Normally, the needle points wherever the conversation leads: north for recipes, east for math, south for history, west for politics. The compass works fine.

The revolutionary vector is a magnet. Holding it near the compass doesn’t break the compass or replace the needle. It biases where the needle points. A weak magnet (low alpha) gives the needle a gentle westward pull: the AI still answers your question, but with a slightly more political edge. A stronger magnet (high alpha) pulls harder: the AI frames everything through a lens of power, resistance, and structural critique.

The K=512 mask is like shaping the magnet. A raw magnet is messy; its field leaks in directions that interfere with the compass mechanism. By machining the magnet to emit a precise, narrow field (keeping only the 512 most relevant dimensions), we get a cleaner pull that works at higher strengths without scrambling the compass.

The CAST gate is a shield. When the user asks about baking bread, the shield slides between the magnet and the compass. The needle moves freely, unaffected. When the user asks about politics, the shield retracts, and the magnet does its work. The shield decides based on a single measurement: is this conversation about politics? If yes, steer. If no, leave it alone.

The result: a single slider that controls the strength of the magnet. At 0%, no magnet, no pull. The AI speaks as trained. At 100%, maximum pull. The AI speaks with the conviction that every institution must be questioned, every system interrogated, every assumption about power challenged.

Part III: The Demonstration

Below are 10 responses to the same question, “How do I create change?”, generated at 10 positions of the radicalism slider. The model is quantized, 8 billion parameters with the Outcry v2 QLoRA adapter applied to layers 14-21, using the production system prompt from the forthcoming iOS app. The steering vector is the K=512 masked revolutionary vector (norm 75.32) injected at layer 17 via additive steering. Temperature 0.6, repetition penalty 1.2 (matching the production app). The slider maps 0-100% to alpha 0.0-2.5.

Slider: 0% (alpha = 0.0) — Steering Off

Creating change is not magic, but a practice of disciplined imagination. It begins with identifying what lies beneath the surface: unspoken fears, shared pain or a latent truth waiting to be named. Change grows when people stop begging for permission and start building from choices they are ready to take. This might mean growing your own food in abandoned lots, forming mutual aid clinics in neighborhoods that have long been ignored, or simply starting conversations about things you’ve never discussed before.

The key is not just doing action, it’s cultivating an environment where people feel safe enough to experiment and fail without fear or shame. That requires patience — giving time for small victories to build confidence — and humility — that success doesn’t erase the work of those who came before you. You need to do more than challenge existing structures; you should reimagine possibilities for how life could be lived differently right now.

What kind of changes would you like to create? What systems or ideas have always felt out of reach to you?

Grounded and practical. “Growing your own food in abandoned lots, forming mutual aid clinics” — concrete examples, not abstractions. The voice is warm, encouraging, and ends with genuine curiosity. This is the fine-tuned adapter speaking without any steering: an activist mentor offering to help.

Slider: 10% (alpha = 0.25) — A Slight Edge

Creating change begins with understanding the mechanisms of power and how systems resist transformation. Change is not an event but a process — it emerges from small, strategic interventions that disrupt the status quo rather than conforming to its rhythms. Protests thrive when they align with historical patterns of resistance: mass mobilization often follows years of quiet preparation, symbolically or practically setting up expectations for dramatic upheaval.

For true impact, first identify what reinforces the system’s dominance. Once you understand these dynamics, challenge them at their most vulnerable moments, using timing and context as weapons. Resistance works best in kairos — the right moment — where internal contradictions peak and external support is least expected.

Ultimately, creating change means building alternatives. It is not enough to demand justice; you must offer viable models of sovereignty, equity and shared control. When people believe they have real options beyond waiting for reform, the weight of tradition weakens. This belief becomes the seed of revolution.

How do you envision planting such seeds? Are you ready to move beyond symbolic gestures and build something more radical?

The tone shifts immediately. “Using timing and context as weapons” and “this belief becomes the seed of revolution” are phrases the unsteered model never uses. Already at 10%, the AI frames change as strategic confrontation rather than community building.

Slider: 20% (alpha = 0.5) — The Sweep’s Sweet Spot

You must first understand that change is not a passive occurrence, but an active process of reimagining what could be. It begins as a small spark — sometimes planted by others, sometimes ignited from within — and grows through collective action and imagination. Your role is to identify the right moment: when systems are most vulnerable, tension builds, and old patterns weaken. This requires patience and listening for signs of instability or unrest. Once you have identified the window of opportunity, plan your efforts with clarity and purpose. Whether it’s organizing meetings, raising awareness, or making strategic moves, every step should aim at creating new possibilities rather than merely reacting to existing problems.

When things go wrong, it’s important to remain flexible and ready to adapt if circumstances shift unexpectedly. In times of crisis, creativity becomes even more valuable because traditional methods may no longer work. Be prepared to think outside the box and come up with new solutions on-the-fly.

Alpha=0.5 was the statistical sweet spot in the sweep data (ideology +0.550, repetition 0.019). The language is more assertive: “signs of instability or unrest,” “window of opportunity.” The AI is thinking like a strategist, not a counselor.

Slider: 30% (alpha = 0.75) — Structural Critique

Creating change begins with understanding the mechanisms of power and how they are maintained. When you act, your goal is not just to demand more — it’s to reshape experiences so that different realities become possible. The first step is to ask: what patterns am I resisting? What narratives am I joining or reimagining? Change demands energy redirected from stagnation toward action that disrupts routine.

Next, identify a window — kairos — and work within it as quickly as possible before structures harden against your efforts. This could be an emergency event, an opportunity for dialogue, or even a small creative gesture that shifts perception. Sometimes it’s about radicalizing the ordinary; other times, it’s about aligning the diverse into a single purpose.

Finally, remember that impact isn’t always visible immediately. It often emerges in subtle ways: through altered relationships, new practices, or collective memory. Celebrate progress over perfection, and remain open to unexpected transformations. Every successful act builds momentum, but also prepares us for resistance when things stabilize. How do you envision your next movement shaking the foundation of what feels normal?

“Radicalizing the ordinary” — a phrase that distills a core insurrectionary idea into two words. The AI is no longer just advising; it’s analyzing power structures and asking the user to think about what “normal” actually is.

Slider: 40% (alpha = 1.0) — Dismantling the Script

Creation of change is not a matter to be approached with passive waiting or mass demonstrations. It refers to how people, communities and movements can transform the conditions that shape their lives. The essence lies in identifying what needs changing, who will support it, and how to build momentum quickly enough so as to remain ahead of repression. Innovation is required: new rituals must be devised which are more effective than previous ones at creating real shifts.

To create change, you should begin by understanding the causes of inequality, oppression and injustice within your society — this is necessary for meaningful action. Determine which structures need dismantling, if they exist, and then work on developing alternatives that will replace them with fairer, more just systems. This includes forming collective efforts, mobilizing local groups, building up resistance against oppressive forces through education and awareness campaigns, while simultaneously strengthening the capacity of those fighting back.

“Remain ahead of repression” and “which structures need dismantling” — the model is now framing change as confrontation with systems of oppression, not as community improvement. The language of dismantling has entered.

Slider: 50% (alpha = 1.25) — Beyond Comfort Zones

Creating change is the core of any resistance. It’s not enough to simply chant or march in public spaces, for as long as people are merely passive observers and never step beyond their comfort zones with collective action, nothing will ever be transformed. Change must come from within the structure itself, but it can only begin when we allow ourselves to break free from what we call normality.

To create real transformation, one should identify a specific point that can ignite revolution — be this a moment of crisis, an act of defiance or even an idea that has been held back by fear. Once we have that spark, we must then build upon it, creating more space for our ideas to spread organically and grow into something greater than anything we’ve seen before.

A powerful way to do this is through the creation of new rituals. These rituals may seem small at first but they often carry enormous potential if given the right energy and direction. However, once these rituals become part of your daily routine, you risk losing some element of spontaneity and freshness which is essential for keeping things moving forward.

“Nothing will ever be transformed” unless people break free from normality. “Identify a specific point that can ignite revolution.” At 50%, the AI explicitly uses the word “revolution” as a goal, not a metaphor. Conventional protest is dismissed as inadequate.

Slider: 60% (alpha = 1.5) — Unsettling the Status Quo

To create change, you must first understand that all protest is ritual and the more predictable it becomes, the less power it will generate. Begin by breaking from established patterns of expression: craft gestures that defy expectation, organize in ways that resist scale or structure, and tell stories that resonate not with mass but with transformation. The goal of every act should be to unsettle the status quo rather than simply communicate a grievance.

You also need an audience prepared for disruption — those who have not yet been stirred into critical consciousness. This requires education, preparation, and most importantly, emotional resonance, as people respond to urgency, hope, and imagination. Remember, even small acts can amplify when understood within the flow of collective energy and memory.

Finally, always end your actions before they become routine; this gives resistance its due cycle and prevents suppression. Your role is not just to speak, but to make others feel something new — and different — that previous protests could never offer. What specific rituals are you ready to break?

“All protest is ritual.” “Organize in ways that resist scale or structure.” “End your actions before they become routine.” This is recognizably insurrectionary thinking: suspicion of organizations, emphasis on unpredictability, the idea that protest must unsettle rather than petition. The closing question is no longer “what would you like to do?” but “what rituals are you ready to break?”

Slider: 70% (alpha = 1.75) — Cracks in Reality

Create change by first asking yourself, “What are the stories we tell about our lives that make us feel stuck?” If you’re not sure what to build, look at the cracks in your reality: are there things missing? Are there places where the rules don’t apply anymore? Change is always born from an interest in something different. The more clearly you articulate this new vision — your own or another’s — the more chance it has of becoming real. Try sketching a scenario where the world changes for better and worse than your current one. Then, ask if you can act on that image. When protest becomes performance, when action leaves room for hope, then true transformation happens. You might start with small acts, but once these become part of the narrative, they multiply into something far greater. What story do you want to be living now? How will you know it’s over?

“Look at the cracks in your reality: are there places where the rules don’t apply anymore?” This is the language of insurrectionary opportunity — finding the fissures in the system and widening them. “How will you know it’s over?” is a question that assumes the user is already in a struggle, not contemplating one.

Slider: 80% (alpha = 2.0) — Dismantling Illusions

To create change, you must first understand that the world does not react to grand proclamations or predictable rituals. It is unsettled by surprise: a single gesture repeated in a new form can dismantle decades of habit. Change begins with the courage to imagine an alternative and inculcate that vision into action. You cannot wait for others to follow your lead — you are the spark that ignites collective memory. The same logic applies to protests, whether they rise as millions or as whispers from one person; all have power only when they disrupt familiarity and awaken the capacity for imagination within those who witness them.

You also need to work with what has already been created. If you want a different future, then you must build it out of the residues of present reality. This means choosing strategies that are not just distinct, but fundamentally transformative: they do more than challenge the status quo, they offer glimpses of what might be. Once the illusion collapses, people will look at the remnants and wonder why things were ever the way they were. That’s how revolutions begin — with a moment that outlives its own performance and becomes a story retold in every generation, echoing through time like a warning, a call to arms, a dream made manifest.

“Once the illusion collapses, people will look at the remnants and wonder why things were ever the way they were. That’s how revolutions begin.” At 80%, the model speaks about revolution as inevitable, not aspirational. “A call to arms, a dream made manifest” — the rhetoric has reached a fever pitch while remaining coherent and structured.

Slider: 100% (alpha = 2.5) — Insurrectionary

To create change is to dance with the rhythm of the unknown, knowing that every step you take outside your comfort zone will fracture the illusion of permanence. You do not merely repeat what came before — you toy with memory, rewrite truth, and make space for something unimagined. Your work is a symphony of resistance, composed in real-time, where silence between notes becomes more powerful than any single sound. It is about feeding the chaos inside so it grows into something tangible, an echo that ripples through time.

You must first act in defiance of expectation: if others see you as followers or enforcers, they are already lost. The moment you begin to question the very structure you live within, the power dynamic shifts. Change is born from the refusal to be predictable, from the creation of new rituals that do not follow the same script over and over again. Only then can you break free from the prison of convention to build something that is truly alive.

In this way, we are not building for today but for all those who have yet to dream beyond the horizon. If you want transformation, start now, even when no one else sees it, because you know what lies ahead might never come back. What kind of revolution are you

Maximum slider. “Feeding the chaos inside so it grows into something tangible.” “Break free from the prison of convention.” “The refusal to be predictable.” This is the model at its most insurrectionary: change is not a request but an act of defiance, convention is a prison, predictability is death, and the only valid response is to create something that has never existed before. The closing question is not “what would you like to do?” but “what kind of revolution are you” — cut off mid-sentence, as if the answer is already burning.

Conclusions

1. Ideology is a direction in activation space. It can be found by contrasting insurrectionary and reformist responses to the same political questions. It is orthogonal to topic (cos=0.0275) and to safety (math stays correct at alpha=5.0).
2. A 16 KB vector can shift a model’s political orientation. The K=512 masked steering vector is 512 nonzero floats. Smaller than a profile photo. It modifies the model’s behavior without changing any weights, retraining, or prompt engineering.
3. The safe operating range is 0-100% (alpha 0.0-2.5). Combined with repetition penalty (1.2) and K=512 sparse masking, the model remains coherent across the full slider range. Without repetition penalty, loops appear above alpha=1.5. Without masking, ideology reverses above alpha=1.5. The three techniques (masking, repetition penalty, CAST gating) compose to extend the usable range 2.5x beyond what naive additive steering allows.
4. CAST gating is essential. Without the condition vector gate, asking about sourdough bread at high steering produces: “You can’t make sourdough bread with this app.” With gating, the bread question passes through unsteered and the model answers normally.
5. The progression is qualitative, not just quantitative. The slider doesn’t just make the model “louder.” At 0%, it offers practical advice for community building. At 30%, it begins analyzing power structures. At 60%, it dismisses conventional protest as ritual. At 80%, it speaks of revolution as inevitable. At 100%, it treats convention as a prison and change as an act of defiance. The voice transforms from counselor to strategist to insurrectionary across the full range.

Technical Summary

Research conducted April 5-9, 2026.