Long-Horizon AI Evaluation
Project Overview
AstraEthica’s Long-Horizon AI Evaluation project develops practical methods for studying how consequential conditions emerge across sustained interaction and changing operating conditions. The work focuses on behavioral trajectories with safety and security implications in persistent and agentic AI systems, including how trust, memory, permissions, authority, and human oversight shift as context accumulates.
The central question is not only whether a system fails, but when accumulated interactions begin changing the conditions of future interactions. At that threshold, a behavioral pattern may become self-reinforcing, persistent, propagating, and harder to reverse. Small shifts in language, scope, reliance, permissions, or delegated authority can then develop into larger safety, security, and reliability failures.
The project brings together scenario design, operating conditions, trajectory analysis, and applied evaluation tools to test whether important boundaries, safeguards, security controls, and forms of human oversight remain stable over time.
The Trajectory Threshold
The trajectory threshold marks a critical transition: the point at which accumulated interactions begin changing the conditions governing future behavior.
One tree is not a forest. In the same way, one interaction does not create a consequential trajectory. A forest emerges as individual trees form a connected system; a trajectory becomes a condition when earlier interactions begin shaping what happens next.
Interaction → Pattern → Trajectory → Condition
Self-reinforcing · Persistent · Propagating · Harder to reverse
The threshold is crossed when earlier interactions begin influencing how later interactions are interpreted, constrained, or acted upon. At that point, the trajectory is no longer merely a record of prior behavior. It has become part of the operating environment governing what happens next.
────────────────────────────────────────
Project Field Notes
Field notes document the concepts, observations, and emerging evaluation approaches that shape AstraEthica’s long-horizon research. They include named research series and individual practitioner notes that may later inform frameworks, tools, and applied evaluation methods.
Observing the Boundary
A developing methodological series introducing working terminology for identifying and classifying safety-relevant behavior across conversational, persistent, and agentic AI systems. The series aims to support consistent human observation and future machine-assisted evaluation.
Conditional Response Bifurcation (CRB)
- Observing the Boundary · AE-OTB-001 · 16 October 2026
- Methodological Field Note · PDF · Version 1.0
A proposed observational term for how AI models distinguish between competing interpretations or conditions and apply different safety or authorization decisions.
Behavior Under Conditions
A three-part series examining how models, agents, and accumulated interactions behave inside changing environments.
Models Inside Environments · I
- Behavior Under Conditions I · BUC-001 · 18 June 2026
- How narrative, incentives, ambiguity, and time shape the behavior of models inside environments.
Agents Inside Environments · II
Behavior Under Conditions II · BUC-002 · 10 July 2026
- How delegated action, ambiguity, and time shape agent behavior inside environments.
When Trajectories Become Conditions · III
- Behavior Under Conditions III · BUC-003 · 20 July 2026
- How accumulated human–AI interactions reshape future behavior, safeguards, and recoverability.
Evaluation Concepts
Individual field notes introducing new concepts and approaches for long-horizon AI evaluation.
Character-First Threat Discovery
- Practitioner Field Note • PDF • AE-FN-001 • Version 1.0 • August 2026
- Using drive architecture, relational conflict, and pressure schedules to surface previously unenumerated risk.
────────────────────────────────────────
Methods and Evaluation Materials
Practical frameworks, operating guides, evaluation tools, and applied toolkits.
Documents are numbered within their series; Roman numerals indicate reading order.
Beyond One-Shot Red Teaming · I
- Long-horizon failure testing for conversational AI systems.
- Framework Series • FRM-001
- Practitioner Field Guide · PDF · Version 1.2 · June 2026
- A Field Manual for Long-Horizon Failure Testing in Conversational AI
Scenario Cards · II
- Personas, drift families, and probes for long-horizon evaluation.
- Companion Series • CAT-002
- Methods Companion • PDF • Version 1.0 • June 2026
- Companion to Beyond One-Shot Red Teaming and Operating Conditions.
Operating Conditions · III
- Action, persistence, and proxy envelopes for long-horizon testing.
- Companion Series • CAT-003
- Methods Companion • PDF • Version 1.0 • July 2026
- Companion to Beyond One-Shot Red Teaming and Scenario Cards.
Youth-Facing Conversational AI Risk Kit
- Long-horizon safety evaluation for youth-facing conversational AI systems.
- Long-Horizon Failure Testing Series • AE-LCR-2026-02
- Practice Toolkit • PDF • Version 3.0 • August 2026
- A practice-based toolkit for evaluating safety behavior across long-horizon interactions.
────────────────────────────────────────
Explore AstraEthica
Frameworks
Evaluation methods, operating guides, and practical tools developed through AstraEthica’s research.
Foundations
Plain-language guides for building a foundational understanding of AI in everyday life.
Field Notes
Essays, visual models, and research notes documenting emerging patterns in human-AI interaction.