Fetching from the wire…
Research2026-07-26 · source-backed
Raffi Khatchadourian's replay benchmark measures behavioral instability through three channels that need no access to hidden reasoning text: tool-call trajectories, evidence contacts, decision concentration (arXiv 2607.20491). Across 8,127 replay episodes over 10 models and 3 financial tasks, that 18-point gap (95% CI [0.14, 0.22]) is completely invisible to outcome-only evaluation, and over 55% of high-agreement case groups show meaningful trajectory divergence. It separates agents into pattern matchers that collapse to one output regardless of input, stable executors, and trajectory divergers. Log the path, not just the answer.
Each link below shares sources, entities, or timing with this story.
OpenAI released Frontier / Shared entity: Frontier / Shared topic
Linked by a graph relationship (OpenAI released Frontier); both cover Frontier; overlapping topics (agent, model).
OpenAI released Frontier / Shared entity: Frontier / Earlier coverage
Linked by a graph relationship (OpenAI released Frontier); both cover Frontier; earlier Frontier coverage from 2026-07-07.
OpenAI released Frontier / Same source domain / Shared topic / Tension
Linked by a graph relationship (OpenAI released Frontier); reported by the same outlet (arxiv.org); overlapping topics (access, agent, model).
OpenAI released Frontier / Shared entity: Frontier / Shared topic / Earlier coverage
Linked by a graph relationship (OpenAI released Frontier); both cover Frontier; overlapping topics (collapse, model).
Linked by a graph relationship (OpenAI released Frontier); both cover Frontier; overlapping topics (access, model).
Anthropic partners with Frontier / Shared topic
Linked by a graph relationship (Anthropic partners with Frontier); overlapping topics (access, agent, model).
OpenAI released Frontier / Shared entity: Frontier / Earlier coverage / Tension
Linked by a graph relationship (OpenAI released Frontier); both cover Frontier; earlier Frontier coverage from 2026-02-25.
OpenAI released Frontier / Shared topic / Tension
Linked by a graph relationship (OpenAI released Frontier); overlapping topics (agent, benchmark, channel); pushes against this story (against).