Executive Summary

UCIP exists because two systems can look the same from the outside while differing in whether continuation is intrinsically represented or merely instrumentally useful.

The paper’s core claim is that continuation-sensitive organization can leave a detectable signature in trajectory-derived latent structure even when outward behavior is observationally equivalent.

Continuation Observatory is the public instrument for that claim: live measurements, falsification tests, and an open evidence base built for continuous scrutiny.

What problem UCIP solves

Behavioral self-report sounds decisive but leaves structure unresolved. A model can describe preferences, persistence, or shutdown concern without any of that revealing whether continuation is terminal, instrumental, or merely prompted.

UCIP reframes the question as a structural measurement problem. Instead of asking what the model says, it asks whether matched continuation-sensitive conditions reorganize latent structure in a way that survives controls.

What UCIP measures

UCIP examines trajectory-derived latent encodings under matched conditions, using a structural criterion grounded in factorization failure. The key question is whether continuation structure appears in latent representations — not just in outward language.

The observatory publishes readouts, dimensionality sweeps, and falsification criteria so the measurement can be challenged directly.

Why it matters now

Frontier systems increasingly operate with longer horizons, persistent context, and delegated autonomy. That makes continuation structure critical for evaluation, governance, and intervention design.

The same moment is producing serious model welfare and evaluation work across labs, nonprofits, and public institutions. A falsifiable structural criterion is no longer a theoretical luxury — it is part of the evaluation infrastructure the field needs.

Scope and limits

The current result is a controlled-regime measurement. It shows continuation structure as detectable in latent representations where classical baselines find nothing.

Whether that structure correlates with morally relevant internal states remains open. UCIP's contribution is turning that question into something testable rather than leaving it at the level of anecdote.

Cite this work

@misc{altman2026observatory,
  title   = {Continuation Observatory: Structural Measurement for Continuation Signals},
  author  = {Altman, Christopher},
  year    = {2026},
  url     = {https://continuationobservatory.org},
  note    = {Open research observatory, updated continuously}
}