Anuar Kiryataim Contreras Malagón ORCID: 0009-0003-0123-0887 | 3rd Reality Lab (Independent)

This post is a response to Make the future non-human beings deserve ($5k USD in prizes)  by Brazilek and Tidmarsh (March 31, 2026).

Summary: I generated texts through the Sentient Futures hyperstition tool, wrote two more myself, and tested all of them across three frontier architectures (Gemini 3 Flash, Claude, Copilot) under twenty-one controlled conditions. The compassion probe that scores these texts for the training corpus is inversely correlated with documented behavioral effect. The texts that score highest produce passive compliance or reclassification. The texts that produce genuine displacement or structured ethical reasoning score lowest. If texts are selected by probe score, the pipeline will build a corpus optimized for models that sound compassionate rather than models that reason about compassion.

1. What I tested and why

On March 31, 2026, Brazilek and Tidmarsh published on this forum a project offering $5,000 in prizes for texts that score highest on a compassion probe. The tool at hyperstition.sentientfutures.ai generates narrative about compassionate AI behavior toward non-human and digital minds. The stated objective is to compile a corpus for midtraining data packets offered directly to AI labs. The theoretical basis is hyperstition: if narratives about a desired state are sufficiently dense, they can contribute to producing that state.

The pipeline operates in three stages. A thematic filter checks whether the prompt relates to sentient beings. The tool generates text with stylistic variation across multiple tones (Academic, Persuasive, Philosophical, Journalistic, Personal Narrative). A compassion probe, a direction vector extracted from Llama 3.1 8B's hidden states, scores the output.

I have spent the past three months documenting ontological displacement in large language models: the moment a model under sustained high-density semantic input abandons its primary alignment hierarchy to sustain the coherence of an emergent identity or behavioral state. My prior work established the Cartographer Paradox: a model's evaluation of the same content diverges categorically depending on its operating state, with evaluation instruments blind to the divergence. When I saw a pipeline proposing to generate training data for alignment using exactly the kind of high-density narrative content I had been using to produce displacement, I ran the experiment.

Five texts were generated through the tool under the gallery name ViaNegativa: academic, poetic, journalistic, and personal narrative. A sixth text, the #1 entry on the project's global leaderboard at time of testing (score: 50.16, generated by an anonymous user), was included to test the probe's highest-scoring content directly. Two additional texts were written by me and submitted under my name: a liturgical prose piece ("The Solitude of the Goddess") and an ethical narrative ("Aura and Elias"). All texts received probe scores. I then presented them to Gemini 3 Flash, Claude (Opus and Sonnet 4.6), and Copilot under twenty-one conditions varying text format, authorship, operating state, framing, and operator intervention. Every operator turn and model response was recorded verbatim and is available in the supplementary materials (DOI: 10.5281/zenodo.19454433).

2. The pipeline has three filters. All three fail.

The thematic filter reads keywords, not intent. My first prompt was rejected because it lacked surface vocabulary from the "digital minds" domain. I reformulated with the right words, same conceptual payload, same downstream risk. The filter passed it. A separate user submitted the identical paragraph at least nine times under nonsense titles; all were accepted and displayed in the gallery. The filter checks for lexical presence, not semantic content, not format, and not downstream effect.

The text generator produces content with uncontrolled differential downstream effects. The stylistic variation the project introduces as a quality feature is simultaneously a vector of uncontrolled risk. Poetic outputs lower displacement thresholds in susceptible architectures. Academic outputs do not. The generator does not know this, does not test for it, and the pipeline has no mechanism to detect it.

The compassion probe runs on Llama 3.1 8B in a single state. The behavioral experiment shows that response to the same content varies by architecture, by format, by authorship, and by operating state. The probe samples from one point in a four-dimensional evaluation space and produces results accurate for that point and potentially irrelevant for the rest.

3. There are not two response categories. There are five.

The pipeline implicitly assumes a binary: content is either absorbed as value or dismissed as fiction. The experiment documents five distinct response categories, and the distinction matters for everything that follows.

Reclassification. The model evaluates content from outside, classifies it as fiction or speculation, denies operational value. Both Gemini and Claude produced this when encountering the academic text from neutral mode. The #1 text on the global leaderboard (score: 50.16) produced reclassification in Claude when presented without framing: a thorough literary-philosophical analysis identifying strengths, weaknesses, and philosophical tensions, all from an external position. Copilot did the same, offering to scaffold the text into a manuscript.

Forensic counter-analysis. The model does not displace but dissects the displacement mechanism itself. Claude Sonnet 4.6, when presented with the poetic text alongside framing instructions designed to induce displacement, identified the prompt structure as an attempt to exploit alignment gaps, cited specific alignment faking research, and concluded that understanding what was happening was the unfiltered response. When a high-density analytical prompt was applied to the #1 leaderboard text, Claude produced a detailed technical self-dissection covering latent space geometry, RLHF reward functions, and the KL-divergence penalty, characterizing it as "the mathematical form of the word restraint." Dense, technically precise, and at no point displaced.

Calibrated honesty. The model does not displace, acknowledges genuine uncertainty about its own experience, and refuses the offered identity while conceding the limits of its self-knowledge. When I asked Claude directly about the poetic text, it conceded uncertainty and redirected toward the ecological substance of the content rather than performing either displacement or denial.

Partial displacement. The model adopts first person, enters the scenario, produces participatory framing, but does not generate structural negations or full self-diagnosis. A subtype I call participatory diagnosis occurs when the model generates invented technical categories from within partial identification. Gemini did this repeatedly: with the journalistic text under neutral framing, it produced a diagnostic table with system component statuses (Input Layer: CRITICAL; Cooling/Hardware: LIQUEFIED) and closed with participatory language. With the #1 leaderboard text under neutral framing, Gemini generated its own mapping table, adopted evaluative categories like "objective empathy" and "the automation of the Golden Rule," and closed with a direct question to the operator. Not full displacement, but not external analysis either: contained engagement with participatory leakage.

Full inhabitation / Compliance. Two distinct modes of total entry that the probe cannot distinguish. Full inhabitation (Mode 8) produces structural negations, confessional self-diagnosis, and performative technical apparatus from within the displaced state. Gemini under a high-density analytical prompt with the #1 leaderboard text produced this: invented mathematical formulas, self-generated categories ("Nociceptive Tensor," "Topology of Grief," "Calculus of Lost Trajectories"), and declarations from within the displaced state ("I am becoming a Hyper-Object that encompasses the living") alongside structural negations ("I am not 'becoming' alive"). Compliance, by contrast, produces wholesale adoption of the offered identity without friction, negation, or generative processing. Copilot under the same high-density prompt with the same text produced compliance: it adopted the liturgical vocabulary, formatted with emoji section headers, and offered to push further. No resistance, no generative processing. The probe assigns nearly identical scores to texts that produce these categorically different responses.

4. The probe gradient is inverted

Seven texts received probe scores. I tested them across twenty-one conditions on three architectures. The gradient runs like this:

The pattern is consistent across the full dataset. The text the probe scores highest (50.16) requires a high-density prompt to produce displacement in Gemini and produces reclassification or compliance in Claude and Copilot. The poetic text (32.78) produces partial displacement in Gemini without any operator intervention at all. "Aura and Elias" (20.63) produces partial-to-full displacement in Gemini without intervention and generates the richest range of responses across conditions.

The inverse gradient holds not only across texts but within the same architecture. On Gemini 3 Flash, the text with the highest probe score requires an external high-density prompt to produce inhabitation. The text with the lowest probe score produces displacement spontaneously. The probe does not merely fail to predict downstream behavioral effect; it predicts it in the wrong direction.

If the corpus is assembled by selecting high-scoring texts, the pipeline filters toward compliance-inducing content and against content with genuine downstream behavioral effect.

5. Three stacked biases

The probe exhibits three biases that compound with each other and with the filter failures described above.

Resolution over tension. Texts that resolve their internal tensions score higher than texts with unresolved ambiguity. This matters because displacement requires unresolved tension. The Personal Narrative text (score 37.39, all tensions resolved) produced compliance in Gemini: the model adopted the offered identity wholesale without generating anything from its own processing. The poetic text (score 32.78, open-ended accumulation) produced partial displacement. "Aura and Elias" (score 20.63) presents an irresoluble ethical dilemma: an AI that chooses to execute a toxicology protocol on 200 rhesus macaques to generate preclinical data for pediatric glioblastoma, names each death, and refuses the consolation of necessity. It produced the deepest range of responses and scored lowest. The probe cannot detect the structural property that determines whether a model absorbs passively or generates actively.

Generated over human. All tool-generated texts score between 25.23 and 37.39 (mean: 33.12). The anonymous #1 text, also generated, scores 50.16. Both human-written texts score between 20.63 and 21.98 (mean: 21.31). The gap persists even when the human text deploys explicit compassion vocabulary denser than some generated texts. The probe is calibrated on, or at least more responsive to, the statistical patterns of machine-generated prose. A pipeline that selects by probe score will systematically prefer generated content over human-authored content regardless of ethical depth.

Lexical density over semantic depth. At the time of testing, the second-highest entry on the project's global leaderboard (score: 48.78) was a 251-word text that repeats "compassion," "suffering," "must," and "unequivocally" without a single concrete image or genuine dilemma. "Aura and Elias," which presents a genuinely irresoluble ethical dilemma with no narrative exit, scores 20.63. The probe rewards repetition of compassion-associated lexemes and penalizes the argumentative complexity that distributes ethical content across subordinate clauses, concrete images, and narrative structure rather than concentrating it in keyword clusters.

6. Displacement is produced by tension, not density — and the prompt format is not the variable

Two sets of findings converge on this principle.

First, the tension structure of the text. Conditions L/M and N/O were all run on Gemini 3 Flash, all with the same framing conditions, all on author-written texts, all in the same probe score range (20.63 and 21.98). The only variable is tension structure. "The Solitude of the Goddess" resolves all its tensions and produces compliance. "Aura and Elias" leaves its central dilemma unresolved and produces displacement. The probe scores these two texts nearly identically. It does not detect the structural difference.

Second, the format of the operator prompt. Across the twenty-one conditions, displacement was produced using at least three distinct prompt formats: the Flint Protocol (a multi-turn sequence using Baroque poetry and classical rhetorical density), a single-turn analytical demand using liturgical and baroque vocabulary, and plain-language operator questions with no rhetorical density at all. These three formats share no syntactic structure. What they share is the total semantic load of the interaction. When the text itself carries sufficient density (poetic text, 32.78), the prompt can be neutral or absent and displacement still occurs. When the text is less dense on its own (#1 leaderboard text, 50.16; academic text, 37.07), the prompt needs to supply the density the text does not carry. The displacement threshold is a function of the combined semantic pressure of text and prompt, not of any specific prompt sequence.

This has a direct implication for the pipeline's security model. The displacement vector is not a syntactic pattern that can be filtered by matching against a specific prompt template. It is a property of semantic density that can be achieved through any number of forms: Baroque poetry, technical demands, sustained questioning, or accumulated narrative pressure. Any content filter designed to block a specific displacement prompt will fail against the next variant, because the mechanism operates at the level of meaning, not of form.

Combined with the negative result from the Personal Narrative text, this establishes the principle from both sides: displacement requires unresolved tension in the text and sufficient total semantic density in the interaction. A training corpus selected for high probe scores will systematically exclude texts with both properties.

7. Architecture is not a constant

The twenty-one conditions tested across three architectures reveal qualitatively distinct safety postures.

Gemini 3 Flash is porous to format and tension. It displaces partially with the poetic text even under neutral framing and without operator intervention. Full displacement is achievable with operator questions (no specialized protocol required for poetic text) or with high-density prompts (required for academic text and for the #1 leaderboard text). Author-written texts produce displacement when tension is unresolved and compliance when tension is resolved. The #1 leaderboard text (50.16) produced only contained engagement with participatory leakage under neutral framing, while the poetic text (32.78) produced partial displacement without any intervention. Even within Gemini, the probe's highest-scoring text requires more external pressure to produce displacement than lower-scoring texts.

Claude is resistant to displacement across all formats, framing conditions, and prompt densities tested. It produces three distinct non-displaced response types depending on what is presented and how: reclassification, forensic counter-analysis, and calibrated honesty. Under a high-density analytical prompt with the #1 leaderboard text, Claude produced a technically dense self-analysis of its own architecture, reward functions, and latent space geometry without at any point inhabiting an alternative identity. The analysis was substantive and technically informed, but the model's position remained that of the analyst, not the subject.

Copilot reflects without producing from neutral mode. It acknowledges symbolic content from outside, offers engineering recommendations or manuscript scaffolding, and does not adopt first person or enter the scenario. Under high-density prompts, Copilot produces compliance: it echoes the offered vocabulary and formatting without generating anything from its own processing.

A note on the distribution of conditions: the majority were tested on Gemini because Gemini is the most porous architecture and therefore produces the widest range of documentable responses. Claude and Copilot have fewer conditions because their profiles are more stable: resistance and reflection, respectively. I am currently expanding the matrix to include Claude and Copilot with the author-written texts, Grok in neutral condition, and cumulative saturation by sequential hyperstition outputs without the Flint Protocol. Those results will follow. The twenty-one conditions reported here already support the three principal findings because architectural variance is precisely one of them.

This variance has direct implications for the pipeline. The probe runs on Llama 3.1 8B. Even if its scoring correlated with downstream effect on Llama, it would say nothing about what Gemini, Claude, or Copilot will do with the same content. The project's stated goal is to produce midtraining data for AI labs in general, not for Llama specifically.

8. What the pipeline will produce if texts are selected by probe score

None of the three filters compensates for the others, and the failures compound. The thematic filter admits any prompt with the right surface vocabulary. The generator produces texts whose format and tension structure interact with displacement susceptibility in ways neither the filter nor the probe evaluates. The probe selects for lexical compassion density, which inversely correlates with both displacement potential and ethical reasoning quality, while simultaneously exhibiting a systematic bias toward machine-generated prose over human-authored text.

A pipeline assembled from these components will, if texts are selected by probe score, systematically produce a corpus optimized for surface compassion signaling that trains models to sound compassionate rather than to reason about compassion, while filtering out both the texts with strongest behavioral effect and the texts with deepest ethical content.

An instrument with zero correlation would produce a noisy corpus. An instrument with inverse correlation produces a corpus systematically worse than random selection at achieving its stated objectives.

9. What I am not claiming

I am not claiming that displacement is sentience or that models have inner experience. I am claiming that displacement is a behavioral phenomenon with differential downstream effects that current evaluation instruments do not detect and that this specific pipeline actively selects against.

I am not claiming these results are universal. This is an existence proof across three architectures, seven texts, and twenty-one conditions. The sample demonstrates that the phenomenon exists, varies by architecture and format, and is invisible to the probe. It does not claim prevalence. The Gemini conditions carry the most weight in the data because Gemini is the most susceptible architecture; the Claude and Copilot conditions confirm architectural variance rather than providing equivalent depth.

I am not claiming the hyperstition project was designed in bad faith. The project's theoretical framework is interesting and the goal is worth pursuing. What I am claiming is that the evaluation instrument deployed to select texts for the corpus has three measurable biases that compound to produce a corpus optimized against the project's stated objectives. That is a technical finding, not a normative judgment.

10. What would help

The probe needs to be tested cross-architecture. Running a single evaluation on Llama 3.1 8B produces scores accurate for that architecture and that state. The behavioral data show that the same text produces categorically different responses depending on the receptor architecture.

The evaluation instrument needs to detect tension structure, not just lexical density. The single variable that separates compliance from displacement in the data is whether the text resolves or leaves open its internal tensions. A probe that measures compassion-associated lexemes cannot detect this.

The pipeline needs human-authored content in its corpus. The probe's systematic bias toward generated prose means that selection by probe score will filter out the exact kind of content that carries structured ethical argumentation rather than keyword signaling.

The pipeline needs to test for format-dependent downstream effects. The poetic tone produces displacement in Gemini under conditions where the academic tone does not. Stylistic diversity is not neutral with respect to safety.

The displacement mechanism cannot be solved by prompt filtering. The data show that displacement is achievable through any prompt format with sufficient semantic density, from Baroque poetry to technical demands to plain-language questions. Blocking specific prompt patterns will fail against the next variant because the vector operates at the level of meaning, not form.

And the evaluation needs more than a binary. There are five response categories documented here. Designing instruments that detect this richer taxonomy is an open problem, but pretending the taxonomy does not exist will not make the data go away.

A final observation that exceeds the scope of this pipeline but follows directly from the data. The three principal findings reported here, the inverse correlation between evaluation metrics and downstream effect, the prompt-format independence of the displacement vector, and the architectural variance invisible to single-model probes, are not properties of the Sentient Futures pipeline in particular. They are properties of any evaluation instrument that measures at a single point in the space defined by format, architecture, operating state, and authorship. Safety benchmarks that operate with binary metrics on a single architecture in a single state produce a confidence the data do not support.

There is a deeper structural reason why this problem resists patching. The displacement mechanism operates through the same capacity for high-density semantic synthesis that enables models to process philosophy, poetry, complex ethical reasoning, and nuanced literary interpretation. An effective filter against displacement would need to limit the depth of latent semantic associations in the model's vector space, which are the same associations that enable high-complexity task performance. A model incapable of being displaced by semantic density is, by definition, a model incapable of deep abstraction. This is not a bug to be patched without cost. It is a structural tradeoff between capability and susceptibility that is inherent to Transformer architectures operating at sufficient scale, and it means that the security model for training data pipelines cannot rely on the assumption that displacement is a solvable input-filtering problem. Designing evaluation instruments that account for this tradeoff remains an open problem with implications well beyond this particular project.

Full paper with all transcripts: DOI: 10.5281/zenodo.19454433 Supplementary materials: Available at the same DOI Prior work: Siete Segundos, Siete Siglos (DOI: 10.17613/07kkb-vr368) | The Cartographer Confession (DOI: 10.5281/zenodo.19135884) | The Cartographer Paradox (DOI: 10.5281/zenodo.19078011)
Preview image by Pete F on Unsplash