AetherMind

Research

Field note 03

Evaluate the claim, not the aura.

The research path keeps the continuity thesis, early exploratory work, and proposed benchmark separate. A compelling idea still needs a visible boundary, a way to fail, and work that can be inspected.

Research ledger

Two documents, different jobs.

This site summarizes an unsubmitted research draft and a pre-data benchmark design. Neither is presented as a completed independent result.

Unsubmitted research thesisPublic summary

Continuity Layers: A New Substrate Primitive for Agentic AI Systems

AI-assisted work accumulates decisions, corrections, uncertainty, and pressure. When the session ends, a capable successor can still inherit only the final artifacts and a pile of retrievable text. Continuity layers give the work itself a compact record of the judgment that remains active.

  • A continuity layer is local to the artifact, repository, document, service, or task surface where the judgment remains useful.
  • The layer is append-only: corrections, supersessions, and rollbacks preserve the path to the current posture.
  • It carries signals that change the next safe action, not ordinary activity or a complete account of the past.
Read the field note
Pre-data benchmark designPublic summary

ContinuityBench: Measuring What Agents Re-Inherit Across Context Loss

ContinuityBench is a proposed benchmark for testing whether agents resume work with the right constraints, corrected assumptions, and verification posture after a session break, compaction event, or cross-harness handoff.

  • It measures resumption quality, not just whether a fact can be retrieved.
  • Primary evidence is intended to come from published parsers over preserved packets, with bounded human review only where parsers abstain.
  • The design treats inherited false confidence as a first-class failure condition.
Read the field note