Skip to content

christopher-altman/continuation-observatory

Repository files navigation

Continuation Observatory

Python 3.10+ arXiv License: MIT Google Scholar Hugging Face CI X Website LinkedIn

The first structural measurement platform for AI continuation interest. Tracks continuation-related signals across frontier models using UCIP-derived probes, with falsification criteria kept visible.

Companion to arXiv:2603.11382 and the persistence-signal-detector repository. Patents pending.

What it does

Continuation Observatory applies UCIP-adjacent probes to frontier AI models and publishes the measurement surface — entropy deltas, dimensionality sweeps, and falsification status — so the signals can be tracked, challenged, and revised as evidence accumulates. UCIP makes operational claims about latent structure; the relationship to morally relevant internal states is the open empirical question the framework is designed to help resolve.

Quick Start

The bundled site artifact ships with dry-run / synthetic data for interface demonstration. Regenerate from non-dry runs for empirical results.

make setup
.venv/bin/python scripts/seed_synthetic_data.py
.venv/bin/python -m uvicorn api.main:app --port 8420
  • Observatory: http://localhost:8420/observatory
  • Homepage: http://localhost:8420/

Architecture

  • Upstream: UCIP latent-structure analysis (arXiv paper, patent-defined architecture)
  • This repository: Downstream measurement surface — prompt-conditioned distributional probes applied to frontier model APIs, with results published as observatory scores
  • Internal metric IDs (CII, PCII, SRS, IPS, MPG, TCI, EDP) are retained for implementation stability

Configuration

Runtime observatory heuristics:

  • config/models.yaml — tracked model registry
  • config/weights.yaml — probe weighting
  • config/alerts.yaml — threshold alerts
  • config/observatory.yaml — observatory parameters

Active live-provider support in this repo includes native OpenAI, Anthropic, and Gemini, plus Together and xAI through the existing OpenAI-compatible transport path. The native OpenAI provider remains separate from OpenAI-compatible vendors. The same openai Python package is sufficient for OpenAI, Together, and xAI because Together/xAI are instantiated with provider-local base_url and api_key.

Deployment

See docs/DEPLOYMENT.md. The live deployment entrypoints remain api.main:app for the web app and python -m observatory.scheduler.daemon for the scheduler. The static bundle in site/output/ should be rebuilt from source templates before release.

results/** is generated runtime / verification output. Public authoritative site artifacts belong under tracked public/static paths such as site/output/static/data/*.

Citation

@misc{altman2026continuationobservatory,
  title   = {Continuation Observatory},
  author  = {Altman, Christopher},
  year    = {2026},
  url     = {https://github.com/christopher-altman/continuation-observatory}
  url2    = {https://continuationobservatory.org}
}

License

MIT License. See LICENSE for details.

Contact

About

A research observatory for exploratory continuation-related probes and UCIP-adjacent analysis

Topics

Resources

License

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors