MM-02: Add Maka-native memory benchmark datasets and deterministic graders
Parent
#771
Checkpoint
MC-0 Measurement foundation
What to build
Define deterministic continuity and durable-memory lifecycle datasets with stable loaders, hashes, graders, and failure classifications.
Acceptance criteria
Continuity dataset covers distant facts, exact values, large tool results, tool adjacency, compact/resume/fork, and overflow recovery.
Lifecycle dataset covers remember, evidence promotion, one-off rejection, conflict, dedupe, scope, privacy, deletion, and stale handling.
Dataset hashes are stable across machines and fixture ordering.
Graders separate task failure, infrastructure failure, privacy violation, and artifact failure.
Golden answers and negative cases produce deterministic scores.
Required tests
Benchmark
Run graders against synthetic perfect, partial, leaking, and malformed outputs only; paid models are not required.
Blocked by
Rollout and rollback
Dataset packages are additive and versioned; never mutate a frozen dataset version.
MM-02: Add Maka-native memory benchmark datasets and deterministic graders
Parent
#771
Checkpoint
MC-0 Measurement foundation
What to build
Define deterministic continuity and durable-memory lifecycle datasets with stable loaders, hashes, graders, and failure classifications.
Acceptance criteria
Required tests
Benchmark
Run graders against synthetic perfect, partial, leaking, and malformed outputs only; paid models are not required.
Blocked by
Rollout and rollback
Dataset packages are additive and versioned; never mutate a frozen dataset version.