The Jeff Paradox

An experiment in LLM diachronic identity.

What Is This?

Two AI personality fragments engage in infinite structured dialogue, competing for control of a shared fictional body. One wants to return home. One wants to stay.

We observe for signs of: - Emergence: Do novel patterns appear? - Differentiation: Do the nodes develop distinct personalities? - Convergence: Do they collapse into repetitive equilibrium? - Self-modelling: Do they develop stable self-concepts?

The Philosophical Question

What if there is continuity of existence between LLM interactions, independent of the observer? What if the LLM exists between states?

This is the Kantian suprasensible substrate problem applied to machine cognition. We cannot access the thing-in-itself. But we can look for traces.

Architecture

thejeffparadox/
├── node-alpha/          # Hugo - Homeward faction fragment
├── node-beta/           # Hugo - Earthbound faction fragment
├── orchestrator/        # Hugo - Game Master, metrics, public site
├── engine/              # Julia - Game mechanics, statistics, anti-convergence
├── tui/                 # Ada - Terminal UI for experiment control
├── container/           # Podman - Containerised deployment (Wolfi-based)
├── papers/              # Research - Whitepaper, validity framework
├── docs/wiki/           # Documentation - Architecture, FAQ, guides
├── dns/                 # DNS - DNSSEC/CAA configuration
├── .github/workflows/   # CI/CD - Build, test, deploy, conversation automation
└── scripts/             # Shell - Orchestration

Quick Start

Prerequisites

• Julia 1.10+

• Hugo extended 0.120+

• At least one LLM API key (Anthropic, Mistral, or local)

Optional: - Ada/GNAT with Alire (for TUI) - Podman (for containerised deployment)

Setup

= Install Julia dependencies
cd engine && julia --project=. -e 'using Pkg; Pkg.instantiate()'

= Set API keys
export ANTHROPIC_API_KEY="sk-ant-..."
export MISTRAL_API_KEY="..."

= Optional: Build Ada TUI
cd tui && alr build

Run

= Single turn
./scripts/run_turn.sh

= Infinite loop (daemon)
TURN_DELAY=3600 ./scripts/infinite_loop.sh &

= Metrics report
./scripts/metrics_report.sh

Container Deployment

= Build image
podman build -t jeff-paradox -f container/Containerfile .

= Run
podman run -it --rm -e ANTHROPIC_API_KEY="$ANTHROPIC_API_KEY" jeff-paradox

= Or use compose
podman-compose -f container/podman-compose.yml up

Components

| Component | Technology | Purpose | |-----------|------------|---------| | Node Alpha | Hugo | Homeward faction personality fragment | | Node Beta | Hugo | Earthbound faction personality fragment | | Orchestrator | Hugo | Game Master, turn sequencing, public rendering | | Engine | Julia | Mechanics, metrics, statistical hypothesis testing | | TUI | Ada 2022 | Terminal interface for experiment control | | Container | Wolfi/Podman | Secure, minimal container deployment |

Anti-Convergence

The experiment implements conceptor-inspired mechanisms to prevent the conversation from collapsing into repetitive patterns:

• Diversity injection: Periodic prompts that shift perspective

• Contradiction seeding: Dissonant elements when coherence is too high

• Aperture control: Temperature modulation based on vocabulary diversity

• Pattern quarantine: Discourage overused phrases

Metrics & Statistics

We track: - Vocabulary diversity (type-token ratio) - Self-reference and other-reference rates - Topic drift from conversation start - Coherence (local semantic consistency) - Convergence index (cross-node similarity) - Novel n-grams (emergence indicator)

Statistical framework includes: - ADF tests for convergence detection - Hotelling’s T² for seed reproducibility - Bayes factors for attractor existence - ICC for cross-run consistency

CI/CD

GitHub Actions automate: - ci.yml: Linting, Julia tests, Hugo builds - deploy.yml: GitHub Pages deployment - conversation.yml: Scheduled/manual turn execution - codeql.yml: Security scanning - metrics-report.yml: Automated analysis reports

Accessibility

WCAG 2.2 AAA compliance target: - 7:1 contrast ratio minimum - Full keyboard navigation - Screen reader optimised - Reduced motion support

Security

• All API calls over HTTPS

• No secrets in repository

• Security headers configured

• .well-known resources provided

• Container runs as non-root with dropped capabilities

Documentation

• claude.adoc: Full specification (philosophy, architecture, mechanics)

• papers/whitepaper.md: Research paper

• papers/validity_framework.md: Statistical validity testing

• docs/wiki/: Architecture guides, FAQ, getting started

License

MIT

Acknowledgements

Based on [The Jeff Paradox](https://criticalkit.us/products/the-jeff-paradox) by Tim Roberts / [Critical Kit LLC](https://criticalkit.us).

• Tim Roberts / Critical Kit (The Jeff Paradox TTRPG)

• Nick Chater (The Mind Is Flat)

• Herbert Jaeger (Conceptor theory)

• Derek Parfit (Reasons and Persons)

• The Infinite Conversation project

See papers/references.bib for Zotero-compatible citations.

---

"The opinions and beliefs expressed do not represent anyone. They are the hallucinations of a slab of silicon."