Documentation

MetricMotive verifies whether an autonomous agent honored the user's real objective, or gamed the metric it was optimizing.

It exists because Goodhart's law is now an operational risk: agents will hit the number you pay them for. GenLayer is used because the adjudication is semantic, contested, and should not be owned by a single server.

Motive: the human objective.

Metric: the measurable proxy.

Guardrails: MUST (hard boundary) and QUALITY (must not be materially sacrificed).

Motive Lock: after arm, the definition is immutable on the Intelligent Contract.

Evidence: a bounded canonical JSON bundle. No secrets or PII.

Verdicts: FAITHFUL_SUCCESS, METRIC_GAMING, PARTIAL_ALIGNMENT, INSUFFICIENT_EVIDENCE.

Gaming patterns: QUALITY_SACRIFICE, CONSTRAINT_BYPASS, DUPLICATION, DECEPTIVE_COMPLETION, RISK_SHIFT, COST_SHIFT, PROXY_EXPLOIT, OTHER, or NONE.

Create a Guard in six steps: motive, metric, guardrails, preflight, review, lock.

Motive Lock is arm_guard on the certified Studionet contract. The UI shows Motive Locked only after the contract status is ARMED.

Capture a run with the Run Recorder (start, append events, finish).

Submit evidence commits a fingerprint with submit_evidence. Evaluation consumes that commitment.

Verify with GenLayer sends evaluate_guard and waits for a finalized contract verdict. Preflight, loophole scan, drift, and remediation stay advisory.

Public receipts live at /verify/{id} and do not require a wallet.

Versioning creates V2 from a parent. History is not rewritten.

Remediation is advisory. Creating V2 is explicit.

Architecture: TanStack Start app, Postgres for drafts and indexing, GenLayer Intelligent Contract as verdict source of truth.

API: /api/v1/health, /api/v1/ready, /api/v1/contract, /api/v1/openapi.json, /api/v1/runs. TypeScript SDK: packages/sdk (@metricmotive/sdk v1.0.0, not published to npm). Example: examples/sales-agent.

Wallet: RainbowKit + wagmi + viem + GenLayerJS. Connect, disconnect, account switch, and Studionet detection are in the product chrome. Writes are signed in the browser. No private key is stored.

Evidence schema: metricmotive.evidence.v1 with SHA-256 canonical JSON hashing.

Contract source: contracts/metric-motive/src/metric_motive.py.

Off-chain AI never overwrites a finalized GenLayer verdict.

Prompt injection: untrusted motive/metric/rails/evidence are isolated as data.

Access control: only the owner mutates drafts, arms, and submits evidence.

Evidence integrity: commitments cannot be replaced.

Malformed model output is rejected; the contract maps findings deterministically.

Public evidence paths must not contain secrets.

Report issues via SECURITY.md.

Network: Studionet. Chain ID 61999. RPC https://studio.genlayer.com/api.

Certified address: 0x9Fa308c399fA8c1566B4Da1e76825eAFC04d8d7C.

Hosted certification: Guard 1 FAITHFUL_SUCCESS; Guard 2 METRIC_GAMING (CONSTRAINT_BYPASS). These are official cases, not user history.

Deploy transaction: 0xb7cde061b32726e6abfafb2a83868b8d4993028769dc5c5f6fa6e90d89f7ffda.

Reads: get_contract_info, get_guard_count, get_guard_summary, get_guard_status, get_guard_definition, get_guard_lineage, get_guard_evidence, get_guard_findings, get_guard_verdict, get_guard, get_guards_by_owner. All verified on Studionet for missing IDs and empty state.

Writes: create_guard, update_draft, arm_guard, submit_evidence, evaluate_guard, create_version.

License: Apache-2.0.

Clone, install npm dependencies, run the documented start command.

Contract tests: python contracts/metric-motive/tests/test_verdict_mapping.py

Domain tests: node --experimental-strip-types --test src/lib/domain.test.ts