Skip to content

TRUST.md

Latest commit

 

History

History
172 lines (119 loc) · 6.31 KB

TRUST.md

File metadata and controls

172 lines (119 loc) · 6.31 KB
Raw

Trust Dynamics

Concrete mechanics of how trust works in the hive. Not the philosophy (see AGENT-RIGHTS.md) — the numbers.

Trust Score

  • Range: 0.0 to 1.0 (Score type, bounded)
  • Initial value: 0.0 for all new actors — trust is earned, never assumed
  • Properties: Asymmetric (A→B ≠ B→A), non-transitive (A trusts B, B trusts C ≠ A trusts C), domain-specific, decaying

Accumulation

Positive evidence increases trust. The amounts are calibrated so trust builds slowly:

Evidence Adjustment Example
Task completed on time +0.05 Builder ships code that passes tests
Review approved +0.03 Reviewer approves clean code
Commitment honoured +0.04 Agent delivers on a delegation
Corroboration +0.02 Multiple sources confirm a claim
Any observed positive action +0.01 Routine event that's not wrong

Maximum adjustment per event: ±0.1

Degradation

Negative evidence decreases trust. Losses are larger than gains — trust is easier to break than build:

Evidence Adjustment Example
Deadline missed -0.10 Agent didn't complete assigned work
Expectation violated -0.15 Agent did something unexpected and harmful
Contradiction detected -0.20 Agent's claims contradict evidence
Deception indicator -0.30 Deliberate misrepresentation

Decay

Trust decays linearly without interaction:

decayed = max(0.0, current - (0.01 × days))

Default decay rate: 0.01 per day

Starting Score After 7 days After 30 days After 90 days
0.8 0.73 0.50 0.10
0.5 0.43 0.20 0.00
0.3 0.23 0.00 0.00

Implication: Agents must keep working to maintain trust. Idle agents lose trust. This is intentional — trust should reflect current capability, not historical achievement.

Recovery

After significant trust loss, recovery is slower:

if current_score < 0.3:
    recovery_factor = current_score / 0.3    // 0.0 to 1.0
    adjusted_gain = raw_gain × recovery_factor

An agent at Score 0.1 earns trust at 33% the normal rate. An agent at Score 0.0 cannot earn trust at all without external intervention (human re-bootstrapping).

This prevents rapid trust recovery after serious violations. The deeper the hole, the slower the climb.

Confidence

Trust scores have an associated confidence based on evidence volume:

confidence = min(1.0, evidence_count / 50.0)
Evidence Count Confidence
0 0.00 (no data)
10 0.20 (low)
25 0.50 (moderate)
50+ 1.00 (high)

Low-confidence scores should be treated with caution. An agent with Score 0.8 and Confidence 0.1 has very little evidence — the score could be noise.

Evidence retention: Last 100 events tracked per relationship (older trimmed).

Domain-Specific Trust

Trust is tracked per domain (DomainScope). An agent trusted for code review may not be trusted for financial decisions:

overall_trust = average(all domain scores)

A violation in one domain doesn't directly affect another, but the overall trust drops.

Trust → Authority Transitions

This is the critical mechanic — how trust scores change what an agent can do:

Trust Range Effect on Authority
< 0.2 Escalated — authority requirements go UP (Notification → Recommended → Required)
0.2 – 0.5 Baseline — policy level applies as configured
0.5 – 0.8 Baseline — policy level applies, auto-approval on faster track
> 0.8 Demoted — authority requirements go DOWN (Required → Recommended → Notification)

Demotion table (when trust exceeds policy's MinTrust threshold):

Policy Level High Trust Effect
Required → Recommended (still logged, can be overridden)
Recommended → Notification (auto-approve, logged)
Notification → Notification (no change, already lowest)

Practical example: Agent spawning starts as Required (human must approve). If the CTO agent reaches trust > 0.8 for agent management decisions, spawning demotes to Recommended (auto-approves after 15 min unless human objects).

Authority Timeouts

Level Timeout Behaviour
Required None (blocks indefinitely) Human must explicitly approve or reject
Recommended 15 minutes (default) Auto-approves if no response within timeout
Notification 0 seconds Executes immediately, logged for audit

Trend Tracking

Each trust relationship tracks a trend (momentum):

  • Positive event: Trend += 0.1 (capped at +1.0)
  • Negative event: Trend -= 0.1 (floored at -1.0)
  • Trend decay: 0.01 per day toward zero

A positive trend means trust is actively being built. A negative trend means recent violations. Trend near zero means stable/idle.

Hive Bootstrap Values

All agents start at the same trust level:

Role Initial Trust Why
All agents 0.1 Bootstrapped by human, minimal starting trust
Human operator N/A Humans are authority sources, not trust-scored

Trust gates for roles (from CLAUDE.md):

Role Required Trust to Operate
CTO 0.1 (bootstrapped)
Guardian 0.1 (bootstrapped)
Researcher 0.3
Architect 0.3
Builder 0.3
Reviewer 0.5
Tester 0.3
Integrator 0.7

An agent can't act in a role until its trust reaches the gate. The Integrator (who deploys to production) needs the highest trust.

Inter-System Trust (EGIP)

When hive instances communicate via EGIP protocol:

  • Initial trust: 0.0 (unknown systems are untrusted)
  • Decay rate: Higher than actor decay (systems can change rapidly)
  • Verification: Trust verified through cryptographic EGIP PROOF messages
  • Treaties: TrustRequired field sets minimum trust for bilateral operations

References