Members view and control what the AI remembers. Independent consent for triage memory, OCR memory, and summarisation memory.
+Four-phase verification system using mathematical similarity rather than generative checking. Confidence badges, claim-level source analysis, and security transparency — all tenant-scoped.
++ The standard approach to AI safety verification — using additional AI models to check AI output — shares a structural flaw with the systems it checks. When both layers are probabilistic, both hallucinate, and both reward confident outputs, they share failure modes. This is common-mode failure: the checker confirms the error because it reasons the same way as the system it checks. +
++ Guardian Agents resolve this by operating in a fundamentally different epistemic domain from the generation layer. The verification mechanism is embedding cosine similarity — a mathematical measurement of how closely an AI response aligns with source material. This is measurement, not interpretation. The watcher is not another speaker. The watcher is a measuring instrument. +
+Every AI response measured against source material via embedding similarity. Score-derived confidence tiers (verified, partially verified, unverified) presented to the member — not binary safe/unsafe labels.
+Individual claims mapped to sources or marked as unmatched. The system does not say “this claim is wrong” — it says “we could not find this in your community’s records.” Absence of evidence is not evidence of absence.
+Tenant-scoped baselines detect deviations in AI behaviour. What counts as anomalous in a parish archive differs from a neighbourhood coordination group — these are different values, not different calibrations.
+Moderator decisions feed back into threshold tuning. Evidence burden is deliberately asymmetric: loosening safety thresholds requires 85% confidence, tightening requires 60%. A regression monitor watches every approved change.
++ The architectural choices in Guardian Agents are not engineering decisions that happen to align with philosophical positions — they are philosophical commitments that demanded specific engineering responses. +
+Verification and generation must operate in different epistemic domains. The sayable (measurement) verifies what inevitably touches the unsayable (generation).
+No objective function resolves values conflicts. Tenant-scoped governance prevents hidden value hierarchies. Asymmetric evidence burdens make trade-offs visible.
+Polycentric governance with genuinely independent verification centres. Moderators, regression monitors, and audit trails — no single authority is root.
+Sovereign processing implements rangatiratanga — the community governs what happens to its own data. The platform exercises kaitiakitanga (guardianship), not ownership.
++ Full analysis: Guardian Agents and the Philosophy of AI Accountability traces the complete philosophical genealogy — from early twentieth-century Vienna to contemporary Aotearoa New Zealand — and examines why these traditions converge on the same architectural requirements. +
Open questions, collaboration opportunities, and data access
+ +How Wittgenstein, Berlin, Ostrom, and Te Ao Māori converge in a production governance architecture
+