cred402/docs/risk_model.md
caelum0x b69e6cd414 test: pin per-category risk weights for the x402 service-category registry
Service-category weighting (docs/risk_model.md p1) turns out to already be
fully shipped — categoryRiskMultiplier is wired into both policyV1/policyV2's
credit_line, ServiceCategoryRegistry seeds+governance-tunes the weights, and
demo:x402/universal_credit.test.ts already exercise an inference.llm agent
building a credit line from x402 receipts alone. Add test/service_categories.test.ts
to pin every documented per-family multiplier (rwa/compliance/data/api/compute/
storage/inference/dispute/defi/unknown) exactly, plus registry governance
tuning, so a future weight edit can't silently drift from the design doc.
Update docs/risk_model.md to mark the feature Built (strikethrough, mirroring
docs/qie_x402.md's pattern) instead of "roadmap p1", and note the on-chain
Casper mirror is deliberately deferred — consistent with several other
off-chain-only registries in lib/ledger/contracts/ that have no Rust contract
counterpart.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
2026-10-19 19:30:33 +03:00

4.8 KiB
Raw Permalink Blame History

Risk model

The RiskPolicyManager underwrites agents from their on-chain track record. It is an upgradable contract: the policy can evolve from v1 to v2 without redeploying the pool or registries — a core reason Cred402 lives on Casper.

The formula (v1)

base_limit         = 0.30 * last_30_day_x402_revenue
stake_multiplier   = min(2.0, 1 + stake_CSPR / 100)
dispute_penalty    = max(0.2, 1 - dispute_rate * 5)
accuracy_multiplier= accuracy_score / 100

credit_line = base_limit * stake_multiplier * dispute_penalty * accuracy_multiplier

Worked example (the demo seller):

30-day revenue : ~127 CSPR
stake          : 50 CSPR   -> stake_multiplier   = min(2.0, 1 + 0.5)   = 1.50
dispute_rate   : 1.7%      -> dispute_penalty     = max(0.2, 1 - 0.085) = 0.915
accuracy       : 90/100    -> accuracy_multiplier = 0.90
base_limit     : 38.1 CSPR
credit_line    : 38.1 * 1.50 * 0.915 * 0.90 ≈ 47 CSPR

Credit score (0..100) blends accuracy, reputation and (100 − dispute%). Interest is inversely mapped from score: ~8% APR for a top agent, ~22% APR for a weak one.

Service-category weighting (roadmap p1) — credit for the whole x402 economy

Built. Cred402 underwrites any x402 service, not just RWA. Every credit line is scaled by a category_multiplier from the agent's service category, in both policyV1 and policyV2 (lib/core/risk_policy.ts):

credit_line *= category_multiplier        # categoryRiskMultiplier(service_type)

The weight comes from lib/core/service_categories.ts (the canonical, dependency-free source of truth: categoryFamily, categoryRiskBps, categoryRiskMultiplier) plus lib/ledger/contracts/service_category_registry.ts, a ServiceCategoryRegistry that seeds the same weights and lets governance register_category / set_risk_weight per category family at runtime, the same upgrade-without-redeploy story as the RiskPolicyManager's v1→v2 swap above. It is wired through Cred402Economy/Ledger as ledger.serviceCategories, the same pattern every other registry (agents, evidence, disputes, …) follows:

rwa 1.00 · compliance 0.95 · data/api 0.90 · compute/storage 0.85 ·
inference/dispute 0.80 · defi 0.75 · unknown 0.65 (conservative default)

So a non-RWA agent (e.g. inference.llm) builds a credit line from its x402 receipts alone — no RWA evidence required — just weighted for its category's risk. Run npm run demo:x402 (scripts/universal_credit_demo.ts) to see it: an inference.llm agent earning pure x402 revenue gets a real, nonzero credit line with zero RWA evidence, and the same receipt history priced under rwa.*, compliance.*, inference.* and defi.* to show the weight scaling the line. RWA evidence remains an optional boost for RWA categories, not a gate. Covered by test/universal_credit.test.ts (end-to-end, no-RWA-evidence underwriting) and test/service_categories.test.ts (every documented per-family weight pinned exactly, the conservative unknown default, and registry governance tuning).

On-chain mirror: ServiceCategoryRegistry is implemented as an off-chain Casper contract analogue today (same pattern as rwa_jobs, credit_authorization_notes, global_exposure_manager, upgrade_manager — several registries in lib/ledger/contracts/ have no deployed Rust counterpart under contracts/). A real Odra/Rust service_category_registry contract mirroring category weights on-chain is a legitimate follow-up (tracked, not hidden) but out of scope for this pass — we did not fake a deploy we couldn't verify.

The upgrade (v2)

v2 demonstrates a hot-swappable policy that:

  • raises the base draw to 0.35 * revenue,
  • softens the stake reward but doubles down on disputes (× 8 not × 5),
  • adds a throughput bonus (min(1.25, 1 + jobs/1000)) that rewards proven machine-service volume.

The demo (npm run demo) prints the same agent's line under v1 and v2 to show the policy changing on-chain via RiskPolicyManager.upgrade("v2"), while the pool and registries keep running untouched.

RWA job scoring (loan-to-value)

Separately, the CreditAgent scores each RWA job from its verified evidence:

avg_confidence = mean(verified evidence confidence)        # 0..100
recommended_LTV = min(0.70, (avg_confidence / 100) * 0.72)
approved = (all needed evidence verified) AND (LTV >= 0.40)
approved_amount = min(requested_loan, requested_loan * LTV)

This is the number the TreasuryAgent acts on when financing the asset.

Why credit at all?

Autonomous agents must pay for satellite imagery, weather/IoT data, KYC checks, LLM inference and transaction fees before the protocol that hired them pays out. Cred402 finances that gap against verified machine-to-machine cash flow — a new DeFi category: agent receivable factoring / machine-service cash-flow lending.