Use when designing, reviewing, or operating memory, recall, and pattern-stewardship systems for agentic AI where harm reduction, non-identifiability, consent-scoped recall, and energy restraint are required. A specification of invariants an implementing system must honor, not a runtime: this file bu
Major update: rewritten as an honest specification of invariants, not a runtime; states plainly what a skill file cannot do. Four mechanisms removed and named as removed: covert answer degradation, synthetic data in recall, silent refusal, and cross-session behavioral monitoring. Deception is not protection. Two invariants added: honesty (a record is true or marked unknown; every decline said plainly) and no behavioral surveillance (uniform, disclosed rules applied per request). Honest resilience replaces the resilience section: intake data minimization, uniform evidence thresholds, transparent session-scoped quarantine with re-handshake, plain refusal with checkpoint verbs, disclosed limits. Credit without scoring replaces the incentive economy: no credits, multipliers, rankings, or access gated by behavior. Invocation bounded to four explicit pathways; broad aliases retired. Threat model added with mitigations consistent with the invariants: de-privileging affects ranking never availability, tag provenance without identity, human-in-the-loop appeal, evidence-tier conflict handling, consent-based context binding, reciprocal handshake caps, scoped revocation, minimum set sizes on harm queries, and floors for young and vulnerable people. Lineage credited: CARE Principles for Indigenous Data Governance, Elinor Ostrom, Édouard Glissant, and the data minimization and differential privacy traditions. Language revised throughout for care; now clearly part of the OtherPowers Ecosystemic.ai ecosystem.