Test a policy before it becomes a rule.

CLI-first status

contextlattice_policy_lab status --project <project> --pretty

The status response reports ledger, context-policy, and temporal-claim state. It is observational: it does not activate a policy automatically.

Simulate on one immutable snapshot

Compare baseline and candidate metrics from the same snapshot. The simulation is deterministic, records no replay result, and cannot mutate runtime policy. Treat its deltas as predictions, not observed production outcomes.

Cards, then recommendations

A scoped policy card combines project, task class, and retrieval intent with a disclosed global prior. Sparse projects shrink toward that prior. Promotion recommendations include assignment exposure, uncertainty, drift, and survivor-bias checks; incomplete evidence means hold or abstain.

Retire, resolve, restore

Retirement is propose-first: content is not deleted and replacement links remain available. Contradiction resolution weighs explicit evidence edges, records an immutable receipt, and abstains when the margin is insufficient. An appeal path stays open instead of rewriting ordinary memory.

Use reversible temperature tiers

Temperature decisions change retrieval-tier metadata such as hot, warm, deep, or retired. They do not claim physical byte movement, delete content, or bypass legal holds. A restore uses the prior receipt and rechecks that the memory has not drifted.

Public core versus paid automation

The public core includes status, no-persist simulation, scoped cards, recommendations, abstention, and explicit one-item lifecycle operations. Paid automation is the separate entitled lane for governed canaries, queues, and bounded multi-item policy execution. Every mutation still requires explicit approval, remains receipt-backed and reversible, and never silently enables runtime policy.