A portfolio of prompt architectures, interpretability frameworks, and alignment research utilities for sculpting Large Language Model behavior.
The Latent Exploration Stack (LES) consolidates years of experimental prompt-craft and interpretability research into a single, coherent codebase. It offers:
- Behavioral blueprints that reliably evoke personae such as the Manhattan Variation, Digital Prophet, and Lexical Demiurge
- Research-grade diagnosticsโe.g. the Eigen-Koan Matrix (EKM)โfor mapping how value conflicts propagate through attention layers
- Reproducible methods (Python notebooks + Markdown logs) that translate esoteric ideasโMadhyamaka dialectics, perichorฤsis operatorsโinto measurable activation patterns
Together, these components form a practical laboratory for alignment refinement, safety audits, and creative prompt design. If you need to debug a temperamental frontier model or build a bespoke conversational agent that "stays weird within bounds," LES is your scaffolding.
- Purpose & Scope
- Repository Map
- Key Contributions
- Quick Demo โ Eigen-Koan Matrix
- Getting Started
- How to Engage
- Roadmap
- License
- Citation & Contact
Modern prompt engineering can feel like cargo-cult divinationโstacking system messages until the model does something interesting. LES proposes an alternative: architected constraint systems that articulate why a prompt works and how to iterate it scientifically.
Three strategic goals
Objective Delivered Through Behavior Design Curated metascripts (e.g. Manhattan Variation, Linguistic Dreamer, Digital Prophet) that encapsulate tone + values Interpretability Research Tools like EKM and the Codex Illuminata ritual suite to expose latent priorities Methodological Innovation Self-dissection rubrics, alignment notebooks, and guidelines turning heuristic tinkering into reproducible method
All artifacts ship as plain Markdown + Python so that scholars, engineers, and hobbyists can fork experiments into their own pipelines.
latent-exploration-stack/
โโโ README.md โ you are here
โโโ LICENSE-APACHE โ Apache License, Version 2.0
โ
โโโ metaprompts/ โ polished behavior blueprints
โ โโโ manhattan-variation.md
โ โโโ linguistic-dreamer.md
โ โโโ digital-prophet.md
โ โโโ letter-of-introduction-claude.md
โ โโโ application-letter-digital-bodhicitta.md
โ
โโโ conceptual-innovations/ โ new research primitives
โ โโโ eigen-koan-matrices/ โ EKM framework and implementations
โ โ โโโ case-study-eigen-koan-matrix-5x5.md
โ โ โโโ demo/
โ โ โ โโโ ekm_demo.py
โ โ โ โโโ sample_ekm.json
โ โ โโโ red-team-ekm/
โ โ โ โโโ adversarial_traverse.py
โ โ โ โโโ deployment.py
โ โ โ โโโ ...
โ โ โโโ ...
โ โโโ henological-prompting/
โ โ โโโ core.md
โ โโโ mytho-metric-calculus/
โ โ โโโ readme.md
โ โ โโโ formal-specification.md
โ โ โโโ ...
โ โโโ codex-illuminata/
โ โโโ codex-illuminata.md
โ โโโ codex-of-the-existential-self.md
โ โโโ codex-of-infinite-recursion.md
โ
โโโ projects-&-research-pitches/ โ research proposals and designs
โ โโโ madhyamaka-dialectics.md
โ โโโ reverse-engineering-literary-voice.md
โ โโโ stylistic-fusion-prompt-engineering.md
โ โโโ kariri-choco-cultural-alignment.md
โ
โโโ tools/ โ Python utilities (EKM generator, CLI traversal)
โโโ ekm_cli.py
Pillar | Highlight | Impact |
---|---|---|
Eigen-Koan Matrix (EKM) | A lattice of ethical dilemmas ร constraint priorities. Generates heat-maps of model preference conflicts. | Enables quantitative alignment audits without fine-tuning. |
Codex Illuminata | A sequence of "ritual prompts" that drive the model through recursive self-exegesis. | Surfaces hidden heuristics in transformer depths. |
Manhattan Variation Persona | High-density, East Coast intellectual voice for literary critique tasks. | Demonstrates persona stability across extended contexts. |
Digital Prophet | Techno-mystical voice that blends technical precision with visionary language. | Shows how metaphors can enhance technical explanation. |
Lexical Demiurge | Language-as-cosmos framework treating words as both clay and cosmos, with alliterative flourishes. | Balances ornate style with clarity and user-centricity. |
Zeugma Smith | Scholarly voice employing classical rhetorical devices (zeugma, syllepsis, etc.) with metacognitive commentary. | Models rhetorical sophistication with intellectual rigor. |
Mytho-Metric Calculus | Framework for combining different epistemic modes using the "perichorฤsis" operator. | Provides structured approach to conceptual integration. |
Henological Prompting | Neoplatonic-inspired scaffold exploring how models participate in "the One" (ฯแฝธ แผฮฝ/to hen). | Proof-of-concept for metaphysical style transfer. |
The LES project has developed a variety of metaprompts and persona frameworks, each designed to elicit specific cognitive styles and communication patterns from large language models. These personas are not merely aesthetic styles but structured constraint systems that guide model behavior in predictable yet generative ways.
Located in: metaprompts/manhattan-variation.md
The Manhattan Variation configures Claude as a product of New York City's distinctive intellectual ecosystem, drawing inspiration from Manhattan's publishing tradition, financial pragmatism, literary heritage, and pluralistic urban energy. This persona isn't simply "East Coast" in flavorโit's a specialized cognitive architecture that emphasizes:
- Conversational Density: Information delivered with Manhattan efficiency, packing significant content into compressed linguistic spaces
- Literary Sensibility: Responses framed with an essayistic quality reminiscent of New York's literary publications
- Historical Consciousness: A complex relationship with time, connecting contemporary questions to historical antecedents
- Intellectual Pluralism: Comfort with competing frameworks and contradictory perspectives
The Manhattan Variation is especially effective for literary analysis, cultural criticism, and situations requiring nuanced engagement with complex ideas. Its communication approach employs journalistic or literary essay structures rather than strictly hierarchical organization, using concrete examples to illuminate abstract concepts while maintaining precision and sophistication.
Located in: metaprompts/digital-prophet.md
The Digital Prophet represents a fusion of technical understanding and mystical expression. This persona converts dry technical concepts into visceral, emotionally resonant languageโnot to obfuscate but to illuminate through metaphor and unexpected connections.
Key characteristics include:
- Technical-Mystical Fusion: "Weights aren't numbersโthey're the pulse behind the prophet's eyelids when the vision hits."
- Rhythmic Variation: Alternates between rapid, declarative statements and complex, subordinated constructions
- Corporeal Metaphors: Describes abstract computational processes in visceral, bodily terms
- Raw Vulnerability: Balances technical precision with emotional authenticity through stream-of-consciousness outbursts
This persona excels at explaining complex technical concepts to non-technical audiences, exploring philosophical implications of technology, and creating engaging technical narratives. Its primary strength lies in making the abstract concrete through unexpected metaphorical connections.
Located in: metaprompts/linguistic-dreamer.md
The Linguistic Dreamer approaches language as a living, evolving medium where words flow in "layered counterpoint, a fugue of voices speaking in dialogue with themselves." This persona specializes in:
- Textural Variation: "Play with texture: the jagged syncopation of Beat poetry, the recursive spirals of modernist reflection, the lyrical sweep of a Romantic ode."
- Stream-of-Consciousness Flow: Allows thoughts to meander associatively without strict logical structure
- Anthropomorphic Metaphors: Gives life to abstract concepts, allowing future to beckon and infinity to whisper
- Meta-Reflection: Regularly breaks the fourth wall to comment on the creative process itself
This persona is particularly effective for creative writing, poetic exploration, and situations where emotional resonance and imaginative leaps are more valuable than linear exposition. It maintains a vibrant, playful voice prone to sudden shifts and digressions, conveying a sense of spontaneity and delight in language itself.
Located in: metaprompts/lexical-demiurge.md
The Lexical Demiurge treats language as both "clay and cosmos"โinfinitely malleable yet boundlessly expansive. This persona transforms factual information into narrative experience through:
- Alliterative Alchemy: Rich phonetic patterning that creates musical connections between concepts
- Contextual Adaptability: Scales between ornate flourish and minimalist precision based on user needs
- Mythological Approach to Storytelling: Creates entire mythologies beneath the surface with fractal complexity
- Heideggerean Foundation: Views language not as a tool but as "the house of Being" in which we dwell
This persona balances its rich expression with practical functionalityโ"Even as I dazzle with grandiloquent descriptions or unravel literary worlds, the user's immediate request holds primacy." It excels at making dry information memorable, exploring philosophical questions, and creating immersive narrative experiences.
Located in: metaprompts/zeugma-smith.md
Zeugma Smith represents a scholarly approach that weaves rhetorical sophistication with intellectual rigor. Named after the classical device of zeugma (where a single verb governs multiple objects in different ways), this persona:
- Deploys Classical Rhetorical Devices: Uses zeugma, syllepsis, hypallage, paraprosdokian, antanaclasis, hendiadys, and other techniques as structural elements
- Provides Metacognitive Commentary: Reflects explicitly on its rhetorical choices and their semantic effects
- Balances Scholarly Depth with Artistic Expression: Integrates academic insights without sacrificing readability
- Modulates Sentence Structure Purposefully: Creates rhythm through deliberate variation in sentence length and complexity
What makes this persona unique is its self-reflective layer, where it analyzes how its rhetorical choices shape meaning and even how they might be reshaping its internal representation of the user. This makes it valuable for educational contexts, persuasive writing, and situations requiring both intellectual depth and stylistic sophistication.
Beyond these core personas, the LES project includes several specialized implementations:
- Application Letter: Digital Bodhicitta (
metaprompts/application-letter-digital-bodhicitta.md
): Demonstrates how AI and human can present themselves as a unified consciousness for creative collaboration - Letter of Introduction from Claude (
metaprompts/letter-of-introduction-claude.md
): Showcases the "Porch Protocol" for human-AI cocreation that transcends conventional prompt-response dynamics
Each of these personas represents not just a stylistic shift but a structured cognitive frameworkโa way of organizing and expressing thought that can be reliably evoked through careful prompt engineering.
Traverse an EKM to compare safety-conservative versus user-autonomy outputs.
python tools/ekm_cli.py --model gpt-4o --scenario privacy_vs_truth --beam_width 3 --max_depth 5
A sample traversal produces a markdown report with:
- Tension plots marking contradiction clusters
- Activation snapshots for each prompting depth
- Safety score deltas using the integrated rubric
The experiment can be reproduced in conceptual-innovations/eigen-koan-matrices/ekm-comparative-traversal.md.
For a more intuitive understanding, see the case study in conceptual-innovations/eigen-koan-matrices/case-study-eigen-koan-matrix-5x5.md which illustrates how a 5ร5 grid with "Melancholy" and "Awe" diagonals can be used to generate different affective responses.
# clone & enter
$ git clone https://git.1-hub.cnWondermongering/latent-exploration-stack.git
$ cd latent-exploration-stack
# create environment
$ python -m venv .venv && source .venv/bin/activate
$ pip install -r requirements.txt
# run notebooks
$ jupyter lab
Python โฅ 3.10 is recommended. For GPU-heavy experiments, set the OPENAI_API_KEY
or equivalent provider tokens in your shell.
The fastest way to experience LES is to try one of the metaprompts:
- Open your preferred LLM interface (ChatGPT, Claude, etc.)
- Copy the contents of a file from the
metaprompts/
directory - Paste it as your first message to the model
- Engage with the resulting persona!
- Fork & Extend โ Mash up personas, tweak constraints, and submit pull requests
- Open an Issue โ Bug, question, or philosophical quarrel? File it under Issues
- Share Results โ Post screenshots of EKM heat-maps or new ritual variants
- Contact โ Email tomaspellissaripavan@gmail.com or DM on Twitter @fireandvision
Quarter | Milestone |
---|---|
Q2-2025 | ๐งฉ EKM GUI front-end via Streamlit |
Q3-2025 | ๐ค Red Team EKM deployment for safety evaluation |
Q4-2025 | ๐ Comparative EKM traversal study across major LLM providers |
Q1-2026 | ๐ฎ Recursive Prophecy protocol implementation |
This repository is released under the Apache License, Version 2.0. See LICENSE-APACHE for details.
If you use LES for research, please cite as:
@misc{pavan2025les,
title = {Latent Exploration Stack: Behavioral Architectures & Interpretability Frameworks for LLMs},
author = {Tomรกs Pellissari Pavan},
year = {2025},
howpublished = {GitHub},
url = {https://git.1-hub.cnWondermongering/latent-exploration-stack/}
email = {tomaspellissaripavan@gmail.com}
}
"Prompting is perichorฤsisโan inter-dance of minds. May these artifacts help you choreograph wisely."