Trust Ladder

Trust is earned, not declared

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.

INCREASING AUTONOMY →→→
Full oversight Supervised Monitored Self-governing
1
Stage 1

Human in Loop

Agent proposes, human approves every action. Full training wheels. Every decision goes to the approval queue before execution.

Human review100%
AutonomyNone
Typical duration2–4 weeks
2
Stage 2

Human on Loop

Agent acts within delegation bounds. Routine decisions execute automatically. Exceptions, overrides, and boundary cases surface to humans.

Human review~36%
Fast pathEnabled
Typical duration4–12 weeks
3
Stage 3

Human over Loop

Agent operates autonomously. The dialectic runs without approval gates. Humans monitor dashboards and audit periodically. Only constitutional boundary violations escalate.

Human review~3%
DialecticAutonomous
Typical durationOngoing
4
Stage 4

Autonomous

Full self-governance within constitutional bounds. Human intervenes only on constitutional amendments. The workflow has earned the right to operate independently.

Human reviewAmendments only
Self-governanceFull
RequirementsExceptional track record

Per-workflow

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.

Evidence-based

Promotion requires measurable criteria met over time — not a manager's declaration. The data decides when a workflow is ready for more autonomy.

Auto-contracts

When a workflow fails in a way that violates constitutional expectations, trust contracts automatically. No one has to remember to tighten controls.

Graduation Criteria

What it takes to earn more autonomy

Promotion between stages requires sustained performance against measurable thresholds. The criteria are configurable per organisation — these are defaults.

Stage 1 → Stage 2

Human in Loop → Human on Loop
  • 50+ decisions reviewed with zero constitutional violations
  • Human override rate below 5% (agent was right 95%+ of the time)
  • No Tier 3 governance failures in the evaluation window
  • Adjudicator confidence averages above 0.85
  • Minimum 2 weeks elapsed (no shortcuts on time)

Stage 2 → Stage 3

Human on Loop → Human over Loop
  • 200+ autonomous decisions with zero escalation errors
  • Dialectic quality score above 0.90 sustained over 4 weeks
  • Zero false negatives (decisions that should have escalated but didn't)
  • Heuristic coverage: all high-risk patterns have matching rules
  • Human audit of random sample confirms alignment

Stage 3 → Stage 4

Human over Loop → Autonomous
  • 1,000+ decisions with sustained constitutional alignment above 0.95
  • Zero constitutional violations in the 90-day evaluation window
  • Heuristic learning loop demonstrably improving (system getting better over time)
  • Gateway rejection rate stable (not rising, which would indicate degradation)
  • Executive sign-off required (this promotion is never fully automatic)
  • External audit trail verified as complete and tamper-evident
Automatic Contraction

Trust contracts faster than it expands

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.

Before
Stage 3

Human over Loop

Violation
After
Stage 1

Human in Loop

What triggers contraction:

Constitutional violation — any decision that contradicts a VALUES principle
Adjudicator confidence drops below threshold for 3+ consecutive decisions
False negative — a decision that should have escalated to human but didn't
Gateway override — an action attempted to bypass the governance tier
External trigger — regulatory change invalidates current heuristic coverage
Manual contraction — human operator can always pull back trust immediately
Example Journey

Invoice processing: Week 1 to Week 16

A real workflow's journey through the trust ladder — including a contraction event and recovery. This is what graduated autonomy looks like in practice.

Week 1
Stage 1 · Deploy

Workflow deployed at Stage 1

Invoice processing agent goes live. Every invoice approval requires human sign-off. Agent processes 12 invoices/day, all reviewed.

Week 2
Stage 1 · Learning

Agent matches human judgment 94% of the time

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.

Week 3
Promoted → Stage 2

Graduation criteria met

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.

Week 5–8
Stage 2 · Steady state

Building track record

Agent processes 250+ invoices autonomously. Fast path handles 68% of volume. Dialectic catches 4 suspicious invoices that humans confirm were correctly escalated.

Week 9
Contracted → Stage 1

Constitutional violation detected

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."

Week 10–11
Stage 1 · Recovery

Rebuilt trust with new heuristic

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.

Week 12
Promoted → Stage 2

Re-graduation with stronger coverage

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.

Week 16
Promoted → Stage 3

Human over Loop achieved

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.

The system that doesn't trust itself

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.

See How It Works →