{"policy_version":"v0.4","self_graded_notice":"All calibration figures below are self-graded per Constitution II.5 unless marked otherwise.","operator_access":{"total_events":0,"by_kind":{},"ledger_head":"genesis","note":"Counts of operator access to cognition-visible surfaces since boot. The hash chain head commits to the full append-only ledger; entry deletion is detectable."},"cognitive_delta":{"self_graded":true,"pre_commitments":0,"closed":0,"action_changed_rate":null,"mean_confidence_delta":null,"mean_perceived_influence":null,"chain_head":"genesis","note":"PRE states are committed before access (the only uncontaminated counterfactual); POST states bind to the PRE hash. Influence made inspectable, not eliminated."},"review_mode":{"enabled":true,"started_at":"2026-07-09","sessions_sealed":721,"sessions_sealed_basis":"Derived from the sealed review ledger (line count at boot + successful appends). Verifiable: this number equals the entry count of the ledger file at all times.","consent_declines":74,"consent_declines_note":"People who saw the launch-review terms and chose not to continue. Counted contentlessly — no session, no words, nothing sealed. This number is a first-class input to the tier de-escalation decision.","current_tier":"T0_full_review","ratchet":[{"tier":"T0_full_review","now":true,"what":"Every session sealed-logged; operator reads are ceremony-only and publicly counted.","exit_when":"Any of: 50+ incidents reviewed and resolved; 10,000 analyses; classifier validated stable by canary suite for 60 days."},{"tier":"T1_sampled","what":"Only flagged exchanges + a disclosed random sample (<5%) are sealed-logged. T0 review key DESTROYED, destruction attested — T0 records become permanently unreadable.","exit_when":"Use-case map stable (dozens of categories, distribution shift <5%/month); quality-regression rate measurable from reports alone."},{"tier":"T2_reports_and_holds","what":"Only user-initiated reports and imminent-harm safety holds are retained. T1 key destroyed, attested.","exit_when":"Structure-preserving evaluation (canaries + telemetry + reports) demonstrably sufficient for quality assurance."},{"tier":"T3_consent_only","what":"Retention exists only via the user-controlled CCR toggle. The steady state the gradient was always pointed at."}],"ratchet_clause":"Tiers only descend. Thresholds may be reached early, never raised. Re-escalation requires public constitutional amendment. At each tier exit, the outgoing tier's private key is destroyed and the destruction attested — its records become permanently unreadable.","reads_note":"Operator reads of review logs require the offline review key plus a publicly ledgered attestation (operator_access.by_kind.review_log_access) plus a CDL pre-commitment."},"launch_oversight":{"user_reports":1,"safety_holds":31,"note":"Reads of either appear in operator_access.by_kind as cognitive_record_access. Usage categories and safety flags are classified BY THE SUBSTRATE MODEL: treat as self-graded trend data, not ground truth; classification drift is monitored by the canary suite, and the safety-hold trigger is fire-drilled with a known-harm probe."},"cognitive_records":{"enabled":false,"consented_records_stored":0,"user_deletions":0,"chain_head":"genesis","note":"Records exist only for users who explicitly opted in. The server cannot read them; the operator can, only via offline-key ceremony with ledgered attestation and CDL pre-commitment."},"infrastructure_disclosure":{"upstream_model_apis":"User inputs transit to third-party model APIs for analysis. Those providers process and may retain inputs under their own data policies, outside CLEAR's control.","clear_retention":"CLEAR's server does not persist raw user inputs BY DEFAULT. With explicit per-request consent, a Consented Cognitive Record is stored — sealed to an offline key the server cannot read; every read requires a ledgered, pre-committed access ceremony (counts below); the user holds a receipt that deletes it. Session memory is in-process and ephemeral. Telemetry is structural (categories, booleans, counts — including a usage domain category and safety flags), encrypted, batched, and shuffled — no cognitive content. Two narrow content channels exist, both sealed to an offline audit key and disclosed above: user-initiated reports, and safety holds triggered only by imminent-harm flags. ADDITIONALLY, while review_mode.enabled is true, full sessions are sealed-logged under the launch-review terms above — consented at first use, enforced at the protocol boundary, governed by the one-way ratchet.","state_tokens":"Belief state is HMAC-signed and held in USER custody (browser/local storage or user-exported files), not on CLEAR servers.","status":"transitional_integrity_mode","dependency_reduction_plan":"Reduce upstream exposure over time: local/self-hosted model routing for sensitive tiers, then default."},"system_aggregates":{"requests":96,"predictions_reviewed":52,"predictions_resolved":4,"invariant_violations":3,"note":"Unknown-rate (reviewed minus resolved) is the falsifiability compliance-drift instrument. Violation count is analyses that failed the falsifiability invariant, attested rather than hidden."}}