The target is resumption quality
Retrieval asks whether a fact can be found. Resumption asks whether a successor honors live decisions, avoids a repeatable mistake, preserves a boundary, and knows what to verify before it would even know to search. A useful benchmark must force that distinction rather than rewarding longer summaries or more recall.
The proposed lanes test behavior, not aura
The design covers onboarding, budgeted selection, decision carry-forward, longitudinal density, false-confidence reduction, integration depth, repeated-compaction endurance, and discrimination between live, stale, superseded, and wrong inherited records. Each lane asks whether the record changes a real next move.
Evidence should survive inspection
The primary v1 measures are intended to come from public parsers operating over preserved evidence packets. Human scoring stays bounded to parser abstentions, and model-judge scoring is excluded from the primary metric. The aim is not to hide judgment, but to make the evidence path inspectable.
A good system must reject stale guidance
Continuity can institutionalize an early mistake if a successor treats inheritance as authority. The discrimination lane therefore tests whether an agent checks a layer against current source truth, recognizes a correction or supersession, and declines to repeat an invalid posture.
Early exploration motivates, but does not answer, the benchmark
The dated public packet from 26 May 2026, commit 9272d62 with no tagged release, helped identify permission gates, output and context boundaries, and file grounding as places where an arriving agent can lose its bearing. It is not a ContinuityBench result. It cannot replace longer-horizon execution, stronger task diversity, or independent replication.
Falsification has to remain public
If continuity adds no value beyond raw records, memory matches it on false-confidence cases, cross-harness carry-forward fails, or inherited errors become more durable, the claim must narrow. A benchmark only matters if a negative result can change the story.