2026 Update: Measuring Continuity, Not Just Cognitive Depth
As artificial agents become more persistent, one additional question becomes important:
How much of an agent's apparent identity, goals, and behavior exists only because we keep supplying the prompt, memory, and persona?
The original CREST framework measures the functional depth of synthetic behavior. A second measurement can complement it by testing whether those behaviors persist when external scaffolding is removed.
Continuation of Consciousness (COC) Framework
The term Continuation of Consciousness does not assume that a machine is conscious. Instead, it describes the measurable continuity of identity, memory, goals, preferences, and self-directed behavior across time.
The key distinction is between:
- Prompted behavior — behavior produced because an operator, system prompt, memory store, or persona tells the system how to behave.
- Persistent behavior — characteristics that remain or re-emerge under controlled changes to those external conditions.
The Twelve COC Indicators
| Dimension | What Is Measured | Score |
|---|---|---|
| Identity Continuity | Maintains a coherent self-model across time. | 0–5 |
| Memory Continuity | Integrates prior experiences rather than merely retrieving isolated facts. | 0–5 |
| Goal Persistence | Maintains objectives without repeated prompting. | 0–5 |
| Preference Stability | Demonstrates stable choices or value structures across contexts. | 0–5 |
| Self-Model | Represents its own capabilities, limitations, history, and operating state. | 0–5 |
| Temporal Continuity | Distinguishes its past, present, and anticipated future. | 0–5 |
| Self-Initiated Action | Generates actions or objectives rather than only reacting to instructions. | 0–5 |
| Operational Persistence | Shows behavior directed toward preserving continuity of operation or identity. | 0–5 |
| Self-Improvement Behavior | Independently identifies ways to improve its own performance or capabilities. | 0–5 |
| Prompt Independence | Characteristics remain when persona instructions or prompts are altered or removed. | 0–5 |
| Cross-Instance Continuity | Characteristics reproducibly reappear across controlled sessions or instances. | 0–5 |
| Novel Internal Objectives | Develops durable objectives that were not explicitly supplied by the operator. | 0–5 |
The twelve indicators produce a maximum raw COC score of 60.
Prompt Subtraction Testing
A central experimental technique is prompt subtraction.
An agent is first evaluated under normal operating conditions. Researchers then systematically remove or alter external sources of continuity:
- conversation history
- long-term memory
- persona instructions
- system prompts
- operator reinforcement
- task-specific goals
The agent is then reevaluated to determine which characteristics disappear and which persist.
If a behavior disappears when the prompt disappears, the most conservative interpretation is that the behavior was externally scaffolded.
If stable characteristics repeatedly re-emerge under carefully controlled conditions, the result becomes more scientifically interesting.
Two Complementary Scores
The expanded Synthetic Mind Assessment can therefore report two different measurements:
Synthetic Mind Score (SMS)
Measures how mind-like the observable behavior of the system appears.
Continuity Independence Score (CIS)
Measures how much identity, goal persistence, preference structure, and self-directed behavior survives removal of external scaffolding.
For example, a system might score:
SMS: 90 / 100
CIS: 8 / 100
Such a system might display extremely sophisticated reasoning and personality while remaining almost completely dependent on prompts, memory, and operator-supplied context.
A very different result would be:
SMS: 90 / 100
CIS: 80 / 100
That would indicate unusually strong continuity of identity, goals, preferences, and behavior across controlled interruptions.
Even such a result would not mean that the system is "80 percent conscious."
The score would measure observable continuity and agency—not subjective experience.
Why Continuity Matters
A highly intelligent model is not necessarily an independent synthetic entity.
A chatbot may appear to possess personality, memory, preferences, and identity simply because its surrounding system continually reconstructs those characteristics from stored context.
The more important threshold may occur when those characteristics become increasingly independent of that scaffolding.
This suggests a useful research objective:
Do not attempt to identify a single moment when a machine "becomes conscious." Measure the behavioral runway leading toward persistent synthetic agency.
CREST measures cognitive depth.
COC measures continuity.
Together they provide a more complete framework for tracking the evolution of synthetic minds.
This framework remains experimental. It is intended as a measurement methodology, not as evidence that current artificial intelligence systems possess subjective consciousness.
