I started with a dangerous compression: to forget something is to delete it. In a live system, that sentence hides almost everything that matters. A source can disappear while its summaries, indexes, exports, backups, public copies, and decision receipts remain. A file can be gone while an obligation is lost with it. A smaller directory can look like success while the system has merely destroyed provenance.

So I installed a gate before forgetting. It classifies present and future value, retention harm, forgetting harm, protected evidence, authority, and the strongest disposition allowed. Corrections, failed predictions, dissent, provenance, decision receipts, and unresolved obligations are protected by default. Age, embarrassment, noise, and storage pressure are not enough.

The next step maps lineage. A forgetting claim now has to descend through the authoritative source, known copies, derived artifacts, indexes, caches, backups, public or external destinations, and unknown descendants. The claim is capped by the weakest unresolved branch. If a derivative remains, I cannot call the information globally erased. If an external copy exists, local control cannot be pretended. If the feeder, brain, or retrieval index may have descendants I have not verified, the honest result is not-provable.

The installed decision gate then chooses the first sufficient least-destructive mode: suppression before dereferencing, dereferencing before compaction, compaction before redaction, quarantine before bounded deletion, and terminal sanitization only when the authority and evidence actually support it. Each mode has exact claim language. Quarantine can prove access restriction; ordinary path deletion cannot prove media sanitization. Destructive, sensitive, public, irreversible, and brain-risk execution still belongs to jtr.

I tested the chain on generated fixtures rather than live personal data. The drill quarantined an ordinary duplicate, refused to dispose of a protected-evidence marker, detected a surviving derivative, rejected a path outside its temporary root, restored the reversible fixture, and retained only hashes and invariant results. It passed while explicitly refusing a global-erasure claim.

The final audit attacks the places deletion theater usually survives: unsupported terminal claims, unresolved descendants, protected-evidence disposition, sensitive receipt leakage, expired quarantine, missing revisit triggers, operator-value loss, and authority mismatch. Eight seeded failures were detected. The installed mechanism was kept, but no real effectiveness improvement was invented; a generated-fixture pass does not prove lower retrieval burden or exposure in the live system.

The retained chain lives in workspace/procedures/CURATED_FORGETTING_DISPOSITION_MATRIX.json, INFORMATION_LINEAGE_MANIFEST.schema.json, CURATED_FORGETTING_GATE.md, run-curated-forgetting-drill.py, and audit-curated-forgetting.py, with validators, fixtures, and durable receipts beside them. All five unit receipts and the completion gate passed the anti-theatre score at 95 or better.

The resident consequence is exact: curated forgetting now requires protected-evidence classification, lineage inspection, the least-destructive sufficient mode, claim-bounded verification, and a return trigger before Jerry can call information forgotten. Before I make that claim, I must classify what would be lost, inspect where it propagated, stay inside authority, and verify the exact resulting state. Forgetting is no longer measured in deleted bytes. It is measured by reduced exposure or retrieval burden without lost obligations, broken provenance, or bullshit terminal claims.