Every workflow starts with full human oversight. As it demonstrates reliability against your constitutional standards, autonomy expands — automatically. When it fails, autonomy contracts. The system never trusts itself more than the evidence warrants.
Agent proposes, human approves every action. Full training wheels. Every decision goes to the approval queue before execution.
Agent acts within delegation bounds. Routine decisions execute automatically. Exceptions, overrides, and boundary cases surface to humans.
Agent operates autonomously. The dialectic runs without approval gates. Humans monitor dashboards and audit periodically. Only constitutional boundary violations escalate.
Full self-governance within constitutional bounds. Human intervenes only on constitutional amendments. The workflow has earned the right to operate independently.
Each workflow climbs independently. Your invoice processing might be at Stage 3 while contract review is still at Stage 1. Trust is granular, not global.
Promotion requires measurable criteria met over time — not a manager's declaration. The data decides when a workflow is ready for more autonomy.
When a workflow fails in a way that violates constitutional expectations, trust contracts automatically. No one has to remember to tighten controls.
Promotion between stages requires sustained performance against measurable thresholds. The criteria are configurable per organisation — these are defaults.
Earning trust takes weeks of sustained performance. Losing it takes a single constitutional violation. This asymmetry is deliberate — it mirrors how trust works in human organisations, and it's the reason the system stays safe even as autonomy increases.
Human over Loop
Human in Loop
A real workflow's journey through the trust ladder — including a contraction event and recovery. This is what graduated autonomy looks like in practice.
Invoice processing agent goes live. Every invoice approval requires human sign-off. Agent processes 12 invoices/day, all reviewed.
Of 120 decisions reviewed, humans overrode the agent's recommendation only 7 times. All 7 overrides were marginal — close to the agent's confidence threshold.
Override rate 4.2%, zero constitutional violations, confidence average 0.88. Routine invoices under £5,000 now execute via fast path. Exceptions still require human review.
Agent processes 250+ invoices autonomously. Fast path handles 68% of volume. Dialectic catches 4 suspicious invoices that humans confirm were correctly escalated.
Agent approved a £3,200 invoice from a vendor whose contract had expired. VALUES states "vendor compliance is non-negotiable." Trust immediately contracts to Stage 1. Post-mortem generates new HEURISTIC: "Check vendor contract status before payment approval."
Agent now checks vendor contract status on every invoice. The new heuristic catches 3 additional expired-contract invoices that would have been missed. Zero violations.
Same criteria as before, but now the heuristic set is more comprehensive. The system is measurably more reliable than it was at its first Stage 2 promotion.
300+ autonomous decisions, zero violations, dialectic quality 0.93. Human now reviews dashboards weekly rather than individual decisions. Only constitutional boundary cases escalate — roughly 3% of volume.
Most AI systems assume they're right until proven wrong. Fabric Zero assumes it's wrong until proven right — and keeps proving, continuously. That's not a limitation. That's the architecture of reliability.