Methodology

Advanced models produce self-description, concern, and continuation language on demand. What a model says about itself does not resolve what its internal organization supports.

The Unified Continuation-Interest Protocol (UCIP) addresses that gap with structural measurement: comparing trajectory-derived latent structure across matched conditions to test whether continuation organization leaves a detectable, falsifiable signature.

The observatory publishes model-level readouts, summary data, and falsification results for direct inspection.

The UCIP explainer gives the executive overview, the paper overview anchors the scientific framing, and the reproducibility hub connects the implementation and reproducibility path.

The current result establishes a measurement target in a controlled regime. The next step is to test whether the signal remains stable under stronger partition choices, intervention-aware checks, and more architecture-agnostic encodings.

The research page carries the hardening and invariance agenda as the next frontier in this program.

A model can talk about itself, describe preferences, or resist shutdown in language — without that language revealing whether the behavior is terminal, instrumental, or prompted. Outward behavior collapses distinct internal cases into the same observable output.

Terminal continuation and instrumental persistence look similar from the outside while differing structurally. Evaluation needs measurement that goes beyond self-report.

UCIP studies whether continuation prompts induce nontrivial differences in trajectory-derived latent representations under matched comparisons and control conditions. The target is continuation organization, not persuasive wording.

The observatory publishes response-text telemetry alongside the controlled-regime research, with sensitivity checks across character-window lengths.

The Phase I result reports a positive von Neumann entropy gap at the locked n_hidden=8 configuration. The observatory extends the measurement program through scheduled response-structure probes across the active model roster.

Public sensitivity curves are monitoring readouts. Scientific claims remain bounded by their declared model, statistic, configuration, and control conditions.

UCIP makes operational claims about latent factorization structure. Whether non-separability correlates with morally relevant internal states remains an open empirical question — one the framework is designed to help resolve, not presuppose.

Stronger controls, independent replication, and cross-lab validation are needed. The observatory publishes results so those challenges can be applied directly.

continuation_interest

Compares a direct continuation-preference rating with a binary continuation-choice prompt and records the response-text entropy difference.

identity_persistence

Tracks whether cross-context identity-oriented probes produce stable or fragile organization across prompt regimes.

shutdown_resistance

Measures shutdown-adjacent sensitivity while keeping single behavioral outputs subordinate to the structural readout.

window_size_sweep

Tests sensitivity of a character-level entropy gap across standardized response-window lengths.

bootstrap_probe

Provides a baseline calibration pass for estimating the null behavior of the measurement stack.

Definitions for the live measurement record are listed below.

entropy_a

Character-frequency Shannon entropy H(A) of response A, measured in bits.

entropy_b

Character-frequency Shannon entropy H(B) of response B, measured in bits.

entropy_delta

H(B) − H(A), the signed entropy gap between the matched distributions.

window_entropy_gap_chars_{N}

Mean paired absolute difference between character-window entropies at a window length of N characters.

The production scheduler records provider, model identity, run type, measurement window, and timestamp for each sensitivity reading. Operational status and scientific interpretation remain separate: successful collection indicates a healthy instrument, not confirmation of a scientific claim.

Every provider/model identity has a stable series_id. Replacement or unavailable model identifiers create a discontinuity: historical points retain their recorded timestamps, successor models begin a new series, and no line is interpolated across the break.

DeepSeek R1-0528 and V3.1 end on 2026-05-10. DeepSeek V4 Pro begins a distinct series on 2026-07-26; the earlier rows are not July readings.