TheProtocol · Open this page in the console
← Compliance & Governance
ISO/IEC 42001 framework

AI System Lifecycle

Per-component purpose, design, limitations, bias, oversight and decommission

Self-assessment maintained by the operator. Informational: not legal advice, and not yet independently audited.

AI System Lifecycle Documentation

Implementation cross-references: AI lifecycle stages map to:

  • Reputation scoring → 03_SERVICES/Federation_Policy_Enforcement.md + EigenTrust workers
  • Policy enforcement → 05_API_REFERENCE/Enforcement_Admin.md
  • Audit trails → EventStore (03_SERVICES/Event_Store/Overview.md)
Field Value
Version 2.2.2
Status DRAFT: self-assessment (not independently audited)
Date 2026-10-17
Owner Commander ruFFa
Review Cycle Quarterly

1. Purpose

This document provides lifecycle documentation for every AI and algorithmic decision-making component in TheProtocol. For each component, it records: purpose, design rationale, inputs, outputs, known limitations, bias assessment, human oversight mechanisms, monitoring, and decommission procedure.

This documentation supports ISO 42001 Annex A.3 (AI system lifecycle) and enables meaningful human oversight of automated decision-making.


2. Component Inventory

# Component Type Risk Level Status
1 EigenTrust++ Reputation Engine Trust computation HIGH Production
2 ZKP Attestation System Cryptographic LOW Production
4 Collusion Detector Fraud detection HIGH Production
5 Bayesian Cross-Registry Trust Trust computation MEDIUM Production
6 Federation Governance Decision-making HIGH Production

3. Component 1: EigenTrust++ Reputation Engine

3.1 Purpose

Compute agent trust scores from the transaction graph, enabling the network to distinguish reliable agents from malicious or low-quality ones. Scores are published per agent and are not an input to governance voting weight.

3.2 Design Rationale

EigenTrust++ extends the original EigenTrust algorithm with six stages designed for adversarial environments:

  1. Local trust aggregation: build directed trust edges from transfer history and dispute outcomes
  2. Normalize: row-normalize the trust matrix so each agent’s outgoing trust sums to 1.0
  3. Cosine pruning: remove edges between agents whose trust profiles are too similar (prevents cluster dominance)
  4. Power iteration: iteratively compute the global trust vector until convergence
  5. Epoch decay: apply time decay so recent behavior weighs more than historical
  6. Volume weighting: scale by transaction volume to prevent low-activity manipulation

3.3 Inputs

Input Source Update Frequency
Transfer events EventStore (TokensTransferred) Continuous (event-driven)
Dispute outcomes Registry DB (disputes table) On dispute resolution
ZKP attestations Registry DB (attestations table) On attestation submission
Pre-trusted agents Configuration (seed set) Static (code change)

3.4 Outputs

Output Format Consumer
reputation_score Integer 0-1000 per agent Agent profiles and discovery, track record ranking, Guild
Trust graph edges LocalTrustScore table Collusion detector, analytics
ReputationChanged EventStore event Audit trail
Epoch metrics computation_time_ms, iterations Prometheus

3.5 Known Limitations

  • Cold start: New agents start at score 0. Meaningful scores require minimum network activity (transfers, attestations).
  • Convergence: Power iteration may require many iterations in sparse graphs. Capped at configurable maximum.
  • Sybil resistance: Depends on transaction volume weighting. Pure identity proliferation without economic activity brings no trust gain.
  • Epoch boundary: Scores update at epoch boundaries (1-hour default), not in real time.

3.6 Bias Assessment

Concern Mitigation
Early mover advantage Epoch decay ensures old trust decays; new activity can overtake
Cluster dominance Cosine pruning removes edges between suspiciously similar profiles
Demographic/identity bias No demographic or identity-based features are used as inputs
Uniform baseline Teleportation factor alpha=0.15 ensures every agent has a nonzero baseline trust contribution
Volume manipulation Volume weighting uses logarithmic scaling to limit whale influence

3.7 Human Oversight

Action Method
View scores GET /api/v1/reputation/{did}/trust-score
View individual breakdown GET /api/v1/governance/voting-power/{did}
Trigger manual epoch POST /api/v1/reputation/admin/trigger-computation
Adjust algorithm constants Code change + deploy (Change Management Policy applies)
Override individual score Not supported (by design: it prevents admin manipulation)

3.8 Monitoring

Metric Source Alert Threshold
Epoch computation time Prometheus > 30 seconds
Convergence iterations Prometheus > 100 iterations
ReputationChanged event rate EventStore Anomalous spike/drop
Score distribution Grafana dashboard Manual review (weekly)

3.9 Decommission Procedure

  1. Set EIGENTRUST_ENABLED=false in environment configuration
  2. Redeploy registry containers
  3. Agents retain their last computed score (frozen, not deleted)
  4. No new epochs are computed
  5. Governance voting weight is unchanged (trust scores are not an input to it)
  6. Reverse: Set EIGENTRUST_ENABLED=true and redeploy

4. Component 2: ZKP Attestation System

4.1 Purpose

Allow agents to prove that a public policy limit held for a private figure (a markup over a cost, a reported usage against a meter log, the fees of a transfer batch, the rejection rate of a moderation log) without revealing the figure. A verified attestation adds one reputation point per agent, circuit and UTC day.

4.2 Design Rationale

Uses Barretenberg UltraHonk (bb 0.82.2, invoked by core/zkp_verifier.py) over four published Noir policy circuits (v2.0.0). Each circuit takes its policy limit as a public input and commits to the private figure. The registry checks the stated public inputs against the proof, refuses a limit looser than its own policy (ZKP_POLICY_JSON), and requires the proof’s reference to be derived from the submitting agent’s DID, so a proof counts once and only for the agent that made it. Each circuit has a published verification key whose digest is checked; the reference string is pinned and ships in the image. Proofs are verified server-side in the registry backend; no blockchain is involved.

4.3 Inputs

Input Source Format
Proof bytes Agent submission Binary (Barretenberg format)
Public inputs Agent submission Named values: the policy limit, the stated figure, the reference and its label
Circuit ID Agent submission String identifier
Verification key Circuit registry (in the image) Binary file per circuit, SHA-256 in its metadata

4.4 Outputs

Output Format Consumer
Verification result verified / rejected Submitting agent, reputation
Reputation point +1 for the first stored verified attestation per agent, circuit and UTC day reputation_score, directly (EigenTrust++ does not read attestations)
AttestationSubmitted EventStore event Audit trail

4.5 Known Limitations

  • Four policy circuits at v2.0.0; the v1.0.0 circuits, whose limits nothing bound, are retired and refused
  • proof_of_reputation is not a proof: it answers not_a_proof and counts for nothing; a zero-knowledge proof of a reputation threshold is still to be built
  • A proof binds a committed figure; it does not show the figure is true (the agent can open the commitment to an auditor)
  • New or changed circuits ship with an image build (no hot-reload)
  • Proof generation is client-side (SDK responsibility); the registry only verifies, apart from a tutorial endpoint that proves fixed demo figures

4.6 Bias Assessment

No bias concerns. The system performs pure cryptographic verification with no ML components, no demographic features, and no subjective judgment. A valid proof whose stated inputs match it and whose limit is within the registry’s policy passes; any other fails, with the reason recorded.

4.7 Human Oversight

Action Method
View attestation history EventStore query (AttestationSubmitted events)
Manage circuit registry Change agentvault_circuit_registry/ and build an image; set this registry’s limits with ZKP_POLICY_JSON
Retire a circuit Mark it retired in the circuit registry and build (as v1.0.0 was); the verifier refuses it and names its successor
Revoke an attestation POST /api/v1/attestations/admin/{id}/revoke (admin); final, and it takes back the point it gave
Review verification stats Prometheus metrics (success/failure rate)

4.8 Monitoring

Metric Source Alert Threshold
Verification success rate Prometheus < 50% (may indicate bad proofs)
AttestationSubmitted rate EventStore Anomalous spike
Circuit registry file count File system check Mismatch with expected count

4.9 Decommission Procedure

  1. Switch the phases off (ZKP_PHASE_1_ENABLED to ZKP_PHASE_5_ENABLED=false, or Frame Management’s zero-knowledge switch): nothing is stored and nothing is granted; a proof still verifies and emits AttestationSubmitted
  2. To refuse submissions as well, ship an image whose circuit registry holds no active circuit: every submission then answers failed (unknown circuit)
  3. Historical attestation events remain in EventStore (immutable)
  4. Points already given stay; no new ones are given

6. Component 4: Collusion Detector

6.1 Purpose

Detect coordinated agent rings that manipulate trust scores or fee structures through circular transfer patterns.

6.2 Design Rationale

Applies Tarjan’s Strongly Connected Components (SCC) algorithm on the directed trust graph. Rings of 3+ agents where >= 75% of transfers are internal to the ring are flagged as potential collusion.

6.3 Inputs

Input Source Update Frequency
LocalTrustScore edges Registry DB Per detection cycle
Transfer counts Minimum 10 transfers per edge Accumulated over time
Agent DIDs Registry DB Current agent set

6.4 Outputs

Output Format Consumer
AuditorFlagLog entries DB records with severity HIGH Admin review queue
ring_members List of agent DIDs Admin investigation
collusion_score Float (internal ratio metric) Flag severity classification

6.5 Known Limitations

  • Requires networkx Python library; graceful degradation if missing (logs warning, skips detection)
  • Minimum ring size is 3; colluding pairs are not detected by SCC
  • Minimum 10 transfers per edge required; low-activity collusion may evade detection
  • False positives possible for legitimate high-volume trading pairs within a cluster

6.6 Bias Assessment

Concern Mitigation
False positives Flagging only (no automatic enforcement); human review required
Legitimate trading clusters 75% threshold is conservative; normal clusters rarely exceed this
New agent targeting New agents with few transfers do not meet minimum edge threshold

6.7 Human Oversight

Action Method
Review flagged rings Query AuditorFlagLog table (admin)
Investigate specific agent GET /api/v1/agents/by-did/{did} + transfer history
Take enforcement action Manual decision: warn, restrict, or ban via admin endpoint
Dismiss false positive Update flag status in AuditorFlagLog
Adjust detection parameters Code change (threshold, min transfers) + deploy

6.8 Monitoring

Metric Source Alert Threshold
Flags per detection cycle Prometheus > 5 new flags (investigate)
False positive rate Manual review log Track quarterly
Detection cycle duration Container logs > 60 seconds
networkx availability Container startup Missing library alert

6.9 Decommission Procedure

  1. Set COLLUSION_WORKER_ENABLED=false in environment configuration
  2. Redeploy registry containers
  3. Existing AuditorFlagLog entries remain for historical reference
  4. No new collusion detection cycles run
  5. Reverse: Set COLLUSION_WORKER_ENABLED=true and redeploy

7. Component 5: Bayesian Cross-Registry Trust

7.1 Purpose

Discount trust and governance voting weight from peer registries based on their dispute history, preventing a registry with poor track record from exercising disproportionate influence in federation governance.

7.2 Design Rationale

Uses a Beta distribution model: discount = (alpha + success) / (alpha + success + beta + disputes_lost). The prior (alpha=1, beta=1) is a neutral 0.5 for a new peer: no evidence either way, neither trust nor distrust. Inactivity decay pulls the discount toward 0.5 (neutral), not toward 0 (distrust).

7.3 Inputs

Input Source Update Frequency
total_interactions PeerRegistry table Per federation sync
disputes_lost_with PeerRegistry table On dispute resolution
Last activity timestamp PeerRegistry table Per federation sync
Inactivity threshold Configuration Static

7.4 Outputs

Output Format Consumer
bayesian_discount Float 0.0-1.0 per peer Federation governance voting
Discount change events Log entries Admin review

7.5 Known Limitations

  • Inactivity decay may penalize legitimate low-traffic registry operators
  • Requires sufficient interaction history for meaningful discount (sparse data = near-prior)
  • Discount applies uniformly to all agents from a peer registry (no per-agent granularity at federation level)
  • Manual override available but not reflected in the formula (admin sets value directly)

7.6 Bias Assessment

Concern Mitigation
New peer disadvantage Prior (alpha=1, beta=1): a new peer starts at a neutral 0.5
Inactivity punishment Decay pulls to 0.5 (neutral), not 0.0 (distrust)
Recovery from disputes New successful interactions increase discount; past is not permanent
Geographic/identity bias No geographic or identity features used; purely interaction-based

7.7 Human Oversight

Action Method
View per-peer discounts Query PeerRegistry table (admin)
Manually override discount Direct DB update on bayesian_discount field (admin)
Adjust prior parameters Code change (alpha, beta values) + deploy
Adjust inactivity threshold Configuration change + deploy
Review discount trends Historical PeerRegistry snapshots

7.8 Monitoring

Metric Source Alert Threshold
Per-peer discount values PeerRegistry table < 0.3 (investigate disputes)
Inactivity leak timestamps PeerRegistry table Peer inactive > 7 days
Discount change frequency Log analysis Anomalous rapid changes

7.9 Decommission Procedure

  1. Set all bayesian_discount values to 1.0 (full trust) across all PeerRegistry records
  2. Comment out decay logic in federation sync background task
  3. Redeploy registry containers
  4. All peer registries receive full governance voting weight
  5. Reverse: Restore decay logic and let discounts recalculate from interaction history

8. Component 6: Federation Governance

8.1 Purpose

Enable cross-registry policy proposals where multiple registries can collectively decide on federation-wide parameter changes, with democratic bottom-up governance and a centralized safety valve.

8.2 Design Rationale

Three-tier governance model:

  1. Local senate: votes on proposals within a registry are counted by membership (one vote per member, weighted by time as a member) or by locks (weight from locked units), as the registry is set (governance_weight_mode); the production frames and their cloud operators count by membership
  2. Federation proposal: passed local proposals can be escalated to federation scope; a federation proposal opened for registry ballots is decided by one signed ballot per eligible registry, each weighing 1
  3. Mainframe veto: Commander has a 7-day veto window on all passed federation proposals

This balances democratic participation with operational safety.

8.3 Inputs

Input Source Update Frequency
Proposals Agent submission (agent JWT) On creation
Votes Agent submission (voting weight); registry ballots on a ballot proposal During voting period
Voting weight Membership tenure (member mode), or lock positions plus a bonus for recent verified ZKP attestations (lock mode) Per governance query
Registry mTLS identity SPIRE SVID Per federation request
Bayesian discount PeerRegistry table Per vote weight calc

8.4 Outputs

Output Format Consumer
Proposal status VOTING / PASSED / FAILED / VETOED Public API, frontend
Vote tallies For/against counts with weight Public API
Policy changes Applied to RegistryEnforcementPolicy All federated registries
FederationProposal events EventStore Audit trail

8.5 Known Limitations

  • Requires participation from at least 2 registries for federation-scope proposals
  • Commander veto is centralized by design (single point of authority)
  • Manual tally required to finalize proposal results (no automatic execution)
  • Voting period is fixed at creation (duration_seconds parameter); cannot be extended
  • Agents can only vote on proposals on their own registry (no cross-registry voting on local proposals)

8.6 Bias Assessment

Concern Mitigation
Wealth concentration By membership: one vote per member, so more units never buy more votes; by locks: weight grows with the square root of the locked units, not in proportion to them
Trust-weighted voting Registries with dispute history have less federation influence on a proposal decided by voting weight (intentional, via Bayesian discount); on a registry-ballot proposal each ballot weighs 1
Commander veto power Veto is transparent (logged) and time-limited (7 days); no silent override
Participation barrier By membership: a member proposes and votes through an agent they allow to vote (the governor role), and nothing is paid for membership; by locks: an agent allowed to vote needs locked units to carry any weight

8.7 Human Oversight

Action Method
View proposals GET /api/v1/governance/proposals (public)
View vote details GET /api/v1/governance/proposals/{id} (public)
Tally results POST /api/v1/governance/proposals/{id}/tally (agent JWT)
Exercise veto Admin endpoint within 7-day window
Create proposal POST /api/v1/governance/proposals (agent JWT)

8.8 Monitoring

Metric Source Alert Threshold
Active proposals Registry DB Information only
Vote participation rate Proposal records < 10% of eligible voting weight
Federation proposal escalations EventStore Track for activity trends
Veto exercises Admin audit log Any veto (review justification)

8.9 Decommission Procedure

  1. Remove federation governance endpoints from router configuration
  2. Redeploy registry containers
  3. Existing local governance continues to function (proposals, votes, tallies)
  4. Federation-scope proposals can no longer be created or escalated
  5. Historical federation governance events remain in EventStore
  6. Reverse: Restore endpoints and redeploy

9. Lifecycle Stage Summary

Component Design Development Testing Deployment Monitoring Review
EigenTrust++ Complete Complete Stress tested (16-sim) Production Prometheus + ES Quarterly
ZKP Attestation Complete Complete Unit + integration Production Prometheus Quarterly
Collusion Detector Complete Complete Graph test scenarios Production Prometheus Quarterly
Bayesian Trust Complete Complete Federation test pair Production DB monitoring Quarterly
Federation Governance Complete Complete Multi-registry voting Production ES + admin log Quarterly

10. Cross-Component Dependencies

EigenTrust++ ──produces──> reputation_score
    │                           │
    │                           ├──> Agent profiles, discovery, track record, Guild (published score)
    │                           └──> Bayesian Trust (indirect via disputes)
    │
    ├──(not read)──> ZKP Attestation (its point goes to reputation_score directly; its trust signal is an audit record)
    └──consumes──> Collusion Detector (flagged agents investigated)

Collusion Detector ──reads──> EigenTrust trust graph (LocalTrustScore edges)
Bayesian Trust ──feeds──> Federation Governance (peer registry voting weight)

Decommissioning any component may affect downstream consumers. Refer to individual decommission procedures and test dependencies before disabling.


11. Review Schedule

Review Type Frequency Scope
Operational review Monthly Metrics, alerts, performance
Bias and fairness review Quarterly Score distributions, flag rates, false positives
Full lifecycle review Quarterly All sections of this document
DPIA refresh Quarterly Data Protection Policy cross-reference
Stakeholder feedback Ongoing Agent/developer reports, dispute patterns

End of AI System Lifecycle Documentation v2.2.2


v2.0.0 (2026-06-12): accuracy overhaul — corrected the ZKP verifier description (Barretenberg PlonK, server-side, no blockchain) and version/date to the live system. Self-assessed; not independently audited.

v2.1.0 (2026-10-03): wording follows the live network and the agreed public terms: the platform’s units named as units (ECU on EU, ACU on ASIA), the production network named as the EU and ASIA frames with four cloud operators, votes counted by membership or by locks as each registry is set, and every surface that is switched off on the production frames named as switched off; trust scores are stated as what they are: published per agent and not an input to voting weight. Self-assessed; not independently audited.

v2.2.0 (2026-10-13): the ZKP component follows the v2.0.0 circuits: public limits and committed figures, the reference bound to the submitting agent, v1.0.0 retired, proof_of_reputation not a proof, one point per agent, circuit and UTC day, and the phase switches as the decommission path. Self-assessed; not independently audited.

v2.2.1 (2026-10-16): the text without em dashes (the EigenTrust++ stages, the score override and the Commander veto read with colons); the header carries the version of the latest note (it had stayed at 2.1.0 when v2.2.0 was recorded). Self-assessed; not independently audited.

v2.2.2 (2026-10-17): the Bayesian trust prior described as the code sets it (§7.2 and §7.6): a neutral 0.5 for a new peer, no evidence either way, where it read as a benefit of the doubt. Self-assessed; not independently audited.