Objective: Add CPA (Cost Per Action) pricing with outcome verification, policy governance, advanced trust scoring, and protocol adapters for external providers.
Phase B builds on the broker-based marketplace foundation to introduce outcome-based economics. External providers can earn bonuses (or face penalties) based on verified task outcomes. This requires a robust outcome verification framework, ML-based trust scoring for bid evaluation, and a governance layer to enforce policies.
Key Principle: AEX remains a broker. Phase B enhances the marketplace with outcome economics while maintaining the separation between AEX (broker) and external providers (execution).
┌───────────────────────────────────────────────────────────────────────┐
│ API LAYER │
│ ┌──────────────┐ ┌────────────────────┐ ┌──────────────────────┐ │
│ │ aex-gateway │───►│ aex-work-publisher │ │ aex-provider-registry│ │
│ │ (Go) │ │ (Python) │ │ (Python) │ │
│ │ │ │ [ENHANCED] │ │ [ENHANCED] │ │
│ └──────────────┘ └─────────┬──────────┘ └──────────┬───────────┘ │
├────────────────────────────────┼────────────────────────┼─────────────┤
│ EVENT BUS (Pub/Sub) │ │
│ │ │ │
├────────────────────────────────┼────────────────────────┼─────────────┤
│ EXCHANGE CORE │ │
│ ┌──────────────────┐ ┌───────┴──────┐ ┌─────────────┴───────────┐ │
│ │ aex-bid-gateway │ │ Pub/Sub │ │ aex-contract-engine │ │
│ │ (Go) │ │ │ │ (Python) │ │
│ │ [ENHANCED] │ └──────────────┘ │ [ENHANCED] │ │
│ └────────┬─────────┘ └─────────────┬───────────┘ │
│ │ │ │
│ ▼ ▼ │
│ ┌──────────────────┐ ┌─────────────────────────┐ │
│ │ aex-bid-evaluator│ │ aex-settlement │ │
│ │ (Python) │ │ (Python) │ │
│ │ [ENHANCED] │ │ [ENHANCED] │ │
│ └──────────────────┘ └─────────────────────────┘ │
├───────────────────────────────────────────────────────────────────────┤
│ NEW SERVICES │
│ ┌──────────────────┐ ┌───────────────────┐ ┌────────────────────┐ │
│ │ aex-governance │ │ aex-trust-scoring │ │ aex-outcome-oracle │ │
│ │ (Python) │ │ (Python) │ │ (Python) │ │
│ │ [*NEW] │ │ [*NEW] │ │ [*NEW] │ │
│ └──────────────────┘ └───────────────────┘ └────────────────────┘ │
├───────────────────────────────────────────────────────────────────────┤
│ SHARED SERVICES │
│ ┌──────────────┐ ┌──────────────┐ ┌──────────────────┐ │
│ │aex-telemetry │ │ aex-identity │ │ aex-trust-broker│ │
│ │ (Go) │ │ (Python) │ │ (Python) │ │
│ │ │ │ │ │ [ENHANCED] │ │
│ └──────────────┘ └──────────────┘ └──────────────────┘ │
│ │
│ Firestore │ Redis │ Cloud SQL │ BigQuery ML │ Vertex AI │
└───────────────────────────────────────────────────────────────────────┘
│
EXTERNAL WORLD
│
┌───────────────────────┼───────────────────────┐
▼ ▼ ▼
┌───────────────┐ ┌───────────────┐ ┌───────────────┐
│ Provider A │ │ Provider B │ │ Provider C │
│ (External) │ │ (External) │ │ (Internal) │
│ A2A Endpoint │ │ A2A Endpoint │ │ A2A Endpoint │
│ MCP Toolbox │ │ MCP Toolbox │ │ MCP Toolbox │
└───────────────┘ └───────────────┘ └───────────────┘
Phase B introduces 3 new services:
| Service | Spec | Language | Runtime | Purpose |
|---|---|---|---|---|
| aex-governance | spec | Python | Cloud Run | Policy engine, safety rails, compliance |
| aex-trust-scoring | spec | Python | Cloud Run | ML-based trust score predictions |
| aex-outcome-oracle | spec | Python | Cloud Run | CPA outcome verification |
Phase B enhances 7 existing Phase A services with CPA support:
| Service | Enhancement Spec | Changes |
|---|---|---|
| aex-work-publisher | spec | Success criteria definition, CPA bonus terms |
| aex-bid-gateway | spec | CPA bid terms, outcome guarantees |
| aex-bid-evaluator | spec | ML-based outcome prediction, expected value ranking |
| aex-contract-engine | spec | CPA tracking, outcome verification triggers |
| aex-settlement | spec | CPA calculation, bonus/penalty processing |
| aex-provider-registry | spec | CPA pricing tiers, outcome history |
| aex-trust-broker | spec | ML model integration, prediction caching |
| Framework | Spec | Purpose |
|---|---|---|
| Outcome Verification | spec | Verify task outcomes against success criteria |
| Component | Spec | GCP Service | Purpose |
|---|---|---|---|
| ML Pipeline | spec | BigQuery ML | Trust score & outcome prediction model training |
| Prediction Service | spec | Vertex AI | Real-time outcome prediction |
| Enhanced Monitoring | spec | Cloud Monitoring | CPA dashboards, outcome metrics |
1. Consumer → aex-gateway: POST /v1/work {category, payload, budget, success_criteria, cpa_bonus}
2. aex-gateway → aex-governance: Validate policy compliance
3. aex-gateway → aex-work-publisher: Validate, persist work with CPA terms
4. aex-work-publisher → Pub/Sub: Publish "work.submitted"
5. aex-work-publisher → External Providers: Broadcast work opportunity with CPA terms
6. Provider decides: Can I meet success criteria? What bonus can I earn?
7. Provider → aex-bid-gateway: Submit bid with CPA guarantees
8. aex-bid-gateway → Pub/Sub: Publish "bid.submitted"
9. (Bid window closes)
10. Pub/Sub → aex-bid-evaluator: Receive "work.bid_window_closed"
11. aex-bid-evaluator → aex-trust-scoring: Get ML-predicted success probability
12. aex-bid-evaluator: Calculate expected value (price vs predicted success × bonus)
13. aex-bid-evaluator → Pub/Sub: Publish "bids.evaluated" with expected value ranking
14. Consumer (or auto) → aex-contract-engine: Award contract
15. aex-contract-engine → Provider: Notify contract awarded with CPA terms
16. Consumer ↔ Provider: Direct A2A communication (AEX exits)
17. Provider → aex-contract-engine: Report completion with outcome metrics
18. aex-contract-engine → aex-outcome-oracle: Verify outcome against criteria
19. aex-outcome-oracle → aex-governance: Validate outcome claims
20. aex-outcome-oracle: Return verified outcome (success/partial/failure)
21. aex-contract-engine → Pub/Sub: Publish "contract.completed" with verified outcome
22. Pub/Sub → aex-settlement: Calculate base price + CPA bonus/penalty
23. aex-settlement: Process payment (provider payout - platform fee)
24. Pub/Sub → aex-trust-broker: Update provider trust score based on outcome
25. Pub/Sub → aex-trust-scoring: Update ML model training data
Key Insight: AEX still exits the execution path after step 15. Steps 16-25 occur when the provider reports back.
# Work submission with CPA terms
work:
budget:
max_base_price: 0.10 # Maximum base price willing to pay
bid_strategy: "best_quality"
success_criteria:
- metric: "booking_confirmed"
type: "boolean"
threshold: true
- metric: "response_time_ms"
type: "numeric"
comparison: "lte"
threshold: 3000
cpa_bonus:
max_total: 0.10 # Maximum bonus pool
criteria:
- metric: "booking_confirmed"
bonus: 0.05 # +$0.05 if booking confirmed
- metric: "response_time_ms"
comparison: "lte"
threshold: 2000
bonus: 0.02 # +$0.02 if under 2s
# Provider bid with CPA guarantees
bid:
base_price: 0.08 # $0.08 base cost
confidence: 0.92 # 92% confident can complete
cpa_acceptance:
- metric: "booking_confirmed"
guarantee: true # I guarantee booking or forfeit bonus
- metric: "response_time_ms"
guarantee: 2500 # I guarantee under 2.5s
penalty_accepted: true # Accept -10% base if task fails
# Settlement calculation
settlement:
base_cost: 0.08 # Provider's bid price
cpa_earned:
booking_confirmed: 0.05 # Outcome verified ✓
response_time: 0.02 # 1800ms < 2000ms ✓
penalty: 0.00 # No failure
total_provider: 0.15 # $0.08 + $0.05 + $0.02
platform_fee: 0.0225 # 15% of $0.15
provider_payout: 0.1275 # $0.15 - $0.0225Phase B introduces robust outcome verification through the Outcome Oracle:
outcome_verification:
sources:
- provider_report # What provider claims
- consumer_confirmation # Consumer can dispute within window
- automated_checks # Programmatic verification
verification_methods:
boolean:
# Require evidence
- provider_must_provide: ["confirmation_number", "timestamp"]
- consumer_dispute_window: 3600 # 1 hour to dispute
numeric:
# Trust provider metrics with spot checks
- trust_provider_if_tier: ["TRUSTED", "PREFERRED"]
- spot_check_rate: 0.1 # Verify 10% randomly
evidence_based:
# Require cryptographic proof
- signature_required: true
- artifact_hash: true
dispute_resolution:
auto_resolve_if:
- provider_tier: "PREFERRED"
- consumer_tier: "VERIFIED"
- amount_under: 0.10
human_review_if:
- amount_over: 1.00
- repeated_disputes: 3Phase B introduces ML-based trust scoring:
-- BigQuery ML model for outcome prediction
CREATE OR REPLACE MODEL aex_analytics.outcome_predictor
OPTIONS(
model_type='LOGISTIC_REG',
input_label_cols=['outcome_success']
) AS
SELECT
provider_trust_score,
provider_success_rate,
provider_contracts_count,
category_match_score,
bid_confidence,
bid_price_ratio, -- bid_price / max_budget
work_complexity_score,
CASE WHEN outcome = 'SUCCESS' THEN 1 ELSE 0 END as outcome_success
FROM aex_analytics.contract_outcomes
WHERE completed_at > TIMESTAMP_SUB(CURRENT_TIMESTAMP(), INTERVAL 90 DAY)# Phase B trust score combines:
# 1. Historical performance (from Phase A trust broker)
# 2. ML-predicted success probability
# 3. CPA performance history
def calculate_phase_b_trust(provider_id: str, work: WorkSpec) -> TrustScore:
# Base trust from Phase A
base_trust = trust_broker.get_score(provider_id)
# ML prediction for this specific work type
ml_prediction = trust_scoring.predict_success(
provider_id=provider_id,
category=work.category,
complexity=work.complexity_score
)
# CPA performance history
cpa_history = trust_broker.get_cpa_stats(provider_id)
cpa_bonus_rate = cpa_history.bonuses_earned / cpa_history.bonuses_possible
# Weighted combination
return TrustScore(
overall=0.4 * base_trust + 0.4 * ml_prediction + 0.2 * cpa_bonus_rate,
base=base_trust,
predicted=ml_prediction,
cpa_rate=cpa_bonus_rate
)The governance service enforces policies across the marketplace:
policies:
work_submission:
- max_budget_per_work: 10.00
- max_cpa_bonus_ratio: 2.0 # CPA bonus cannot exceed 2x base
- required_success_criteria: true
- banned_categories: ["illegal.*", "adult.*"]
bidding:
- min_confidence: 0.5 # Must be at least 50% confident
- max_price_to_budget_ratio: 1.0
- require_mvp_sample: true
provider:
- min_trust_for_high_value: 0.7 # Work >$1 requires TRUSTED+
- max_concurrent_contracts: 100
- cooldown_after_failures: 3600 # 1 hour after 3 failures
outcome:
- dispute_window_seconds: 3600
- max_disputes_per_day: 10
- auto_fail_on_timeout: truePer Phase B objectives:
- All 10 Phase A services remain operational and healthy
- 3 new services deployed:
- aex-governance (policy engine)
- aex-trust-scoring (ML predictions)
- aex-outcome-oracle (verification)
- 7 Phase A services enhanced with CPA support:
- aex-work-publisher (success criteria, CPA terms)
- aex-bid-gateway (CPA bids)
- aex-bid-evaluator (expected value ranking)
- aex-contract-engine (CPA tracking)
- aex-settlement (CPA calculation)
- aex-provider-registry (CPA pricing tiers)
- aex-trust-broker (ML integration)
- Outcome verification framework operational
- CPA pricing end-to-end: work → bid → contract → outcome → settlement with bonus/penalty
- Trust score ML model trained in BigQuery ML
- Governance policies enforced on all requests
- Provider dashboard showing base + CPA earnings breakdown
- <100ms added latency for CPA processing (vs Phase A baseline)
- Backward compatible: Base-price-only work requests work unchanged
Phase B builds incrementally on Phase A:
Week 1: [Infrastructure] BigQuery ML setup, Vertex AI prediction endpoint
Week 1: [Framework] Outcome verification framework
Week 1: [Enhanced] aex-work-publisher (success criteria, CPA terms)
Week 2: [New] aex-governance (policy engine)
Week 2: [Enhanced] aex-provider-registry (CPA pricing tiers)
Week 2: [Enhanced] aex-bid-gateway (CPA bids)
Week 3: [New] aex-trust-scoring (ML predictions)
Week 3: [Enhanced] aex-bid-evaluator (expected value ranking)
Week 3: [Enhanced] aex-trust-broker (ML integration)
Week 4: [New] aex-outcome-oracle (verification service)
Week 4: [Enhanced] aex-contract-engine (CPA tracking)
Week 4: [Enhanced] aex-settlement (CPA calculation)
Week 5: [Integration] End-to-end CPA flow testing
Week 5: [Dashboard] Provider earnings dashboard
Phase B requires all Phase A services to be deployed and functional. Each enhanced service builds directly on its Phase A implementation:
| Phase A Service | Phase A Spec | Phase B Enhancement |
|---|---|---|
| aex-gateway | phase-a/specs/aex-gateway.md | + Routes to governance for policy checks |
| aex-work-publisher | phase-a/specs/aex-work-publisher.md | + Success criteria, CPA bonus terms |
| aex-provider-registry | phase-a/specs/aex-provider-registry.md | + CPA pricing tiers, outcome history |
| aex-bid-gateway | phase-a/specs/aex-bid-gateway.md | + CPA bid terms, guarantees |
| aex-bid-evaluator | phase-a/specs/aex-bid-evaluator.md | + ML prediction, expected value |
| aex-contract-engine | phase-a/specs/aex-contract-engine.md | + CPA tracking, outcome oracle calls |
| aex-settlement | phase-a/specs/aex-settlement.md | + CPA calculation, bonus/penalty |
| aex-trust-broker | phase-a/specs/aex-trust-broker.md | + ML integration, CPA stats |
| aex-telemetry | phase-a/specs/aex-telemetry.md | + CPA metrics export |
| aex-identity | phase-a/specs/aex-identity.md | No changes |
Phase B follows strict incremental development:
- No Breaking Changes: All Phase A APIs remain backward compatible
- Optional Fields: New fields (success_criteria, cpa_terms) are optional
- Feature Flags: CPA features can be disabled to maintain Phase A behavior
- Database Migrations: Additive only, no schema breaking changes
- Event Compatibility: New event fields, consumers ignore unknown fields
| Component | Phase A Setup | Phase B Addition |
|---|---|---|
| Firestore | Work specs, providers, contracts | + Policies, outcome records |
| Cloud SQL | Billing ledger | + CPA breakdown columns |
| Redis | Rate limits, cache | + Prediction cache, policy cache |
| Pub/Sub | Event bus | + New topics (governance, oracle) |
| BigQuery | Analytics | + ML training data, model storage |