OpenAgents Economy Kernel Proto Plan

March 18, 2026 · View on GitHub

This document contains two implementation-facing rewrites:

  1. a revised proto plan (packages, file layout, and updated/added protos) that bakes in: tfb, severity, verification budget hints, provenance bundles, correlation groups, and sv / Δm_hat / XA_hat “metrics receipts” (minute snapshots).

  2. a revised PolicyBundle schema (proto + semantics) with the knobs needed to gate autonomy based on sv, XA_hat, and correlation risk.

Sections 1 and 2 cover the core proto and policy work. Sections 3 through 6 describe the optional compute-market extension.

Everything below stays aligned with your existing openagents/*/v1 naming and with the kernel invariants (HTTP-only authority, deterministic receipts, idempotent).

Status legend

  • implemented: shipped in the current MVP or repo entry points
  • local prototype: modeled in desktop-local kernel receipts, snapshots, or protocol notes, but not yet generalized into a full authoritative market surface
  • planned: target architecture, not yet shipped as a production market

Current repo status

The repo now includes a checked-in thin proto/ tree plus a generated openagents-kernel-proto crate.

This document is therefore both:

  • a description of the current thin wire slice already checked into the repo, and
  • a proto expansion plan for the broader multi-market package surface that still needs to land.

Today the repo ships:

  • an implemented compute-provider earn loop,
  • implemented starter backend authority flows across compute, data, labor, liquidity, and risk in apps/nexus-control and openagents-kernel-core,
  • an implemented thin proto slice under proto/openagents/{common,compute,economy,labor}/v1 with generated Rust types in crates/openagents-kernel-proto,
  • local prototype richer kernel receipts, incidents, and snapshots in the desktop runtime,
  • and planned broader package coverage for data, liquidity, risk, policy, safety, audit, and compliance.

The asymmetry in this file is intentional but temporary: Compute and Risk are more detailed than Data and Liquidity because that is where the current spec work went deepest. The five-market architecture still includes all of them.

1. Proto plan rewrite

1.1 Goals of the proto reshape

What changes vs your earlier protos:

  • Move kernel-wide enums/types (verification tiers, tfb, severity, provenance grade, independence/correlation metadata) into common so every service can attach consistent hints into receipts.
  • Add identity assurance and credential-proof references as first-class receipt/policy inputs (including personhood/org-vetting gates where policy requires them).
  • Make provenance a first-class typed bundle (not just “some evidence ref”), so you can compute a deterministic Pgrade.
  • Make provenance explicitly carry data source and permissioning references and support policy-gated verifiable inference/execution attestations.
  • Make verification independence a first-class constraint and report (lineages + correlation groups + min distinct lineages).
  • Add a canonical EconomySnapshot artifact emitted once per minute with a receipt (your “sv/Δm/XA metrics receipts”) so /stats is derived from receipts/snapshots exactly as the spec says.
  • Add proto-first incident reporting, safety signals, certification, and audit package exports so insurability and interoperability are machine-legible.
  • Add explicit insurance-boundary pricing fields (execution vs liability premium/risk charge) and rollback/monitoring contract hooks.
  • Add optional liability markets (coverage-first) as bounded, receipted verification-capacity instruments.
  • Add first-class proof-of-cost evidence to make compute integrity machine-legible for pricing, drift, and underwriting.
  • Add a separate policy package that defines the PolicyBundle schema (see Part 2).

1.2 Proposed file layout

proto/
  openagents/
    common/v1/
      common.proto                 # existing + expanded (tfb, severity, tiers, provenance, identity assurance, cost proofs)
    hydra/v1/
      abp_bonds.proto              # ABP reserve/release/draw (unchanged except minor links)
    aegis/outcomes/v1/
      outcomes_work.proto          # WorkUnit/Contract updated for pricing split, rollback, monitoring, provenance, independence, cost-proof linkage
    aegis/markets/v1/
      liability_market.proto       # optional coverage market (coverage offers/bindings/settlement)
    aegis/incidents/v1/
      incidents.proto              # IncidentReport/NearMiss/GroundTruthCase + taxonomy linkage
    audit/v1/
      audit_package.proto          # deterministic export bundles + redaction tiers
    compliance/v1/
      certification.proto          # SafetyCertification issuance/revocation
    safety/v1/
      safety_signals.proto         # public aggregate + restricted signal feed
    economy/v1/
      economy_snapshot.proto       # minute snapshots: sv/sv_effective, Δm_hat, XA_hat, insurer + drift + identity metrics
    policy/v1/
      policy_bundle.proto          # auth, monitoring, risk pricing, certification, rollback knobs

1.2.1 Five-market package map

The OpenAgents Marketplace consists of five interlocking markets: Compute, Data, Labor, Liquidity, and Risk.

This proto plan should make all five explicit even when some packages remain thin or deferred.

MarketCurrent or planned package rootsCore objectsCurrent repo status
Computeopenagents.compute.v1, openagents.economy.v1, openagents.common.v1ComputeProduct, CapacityLot, DeliveryProof, ComputeIndex, CapacityInstrumentimplemented thin proto slice in the repo; planned broader proto surface
Dataopenagents.data.v1, openagents.common.v1, openagents.economy.v1DataAsset, AccessGrant, PermissionPolicy, DeliveryBundle, RevocationReceiptimplemented starter authority slice and starter proto/read-model surface in kernel-core and nexus-control; planned broader proto surface
Laboropenagents.labor.v1, openagents.economy.v1, openagents.common.v1WorkUnit, Contract, Submission, Verdict, Claimimplemented thin proto slice in the repo plus starter authority flows; planned broader claim and dispute surface
Liquidityopenagents.hydra.v1, openagents.liquidity.v1, openagents.common.v1, openagents.economy.v1Quote, RoutePlan, Envelope, SettlementIntent, ReservePartitionimplemented starter authority slice in kernel-core and nexus-control; planned broader proto surface
Riskopenagents.aegis.markets.v1, openagents.policy.v1, openagents.economy.v1CoverageOffer, CoverageBinding, ClaimResolution, RiskSignal, CalibrationMetricimplemented starter authority slice in kernel-core and nexus-control; planned broader proto surface

1.2.2 Missing package boundaries that still need to be added

To make the five-market architecture canonical at the wire level, the proto tree still needs explicit homes for:

  • openagents.liquidity.v1
  • openagents.risk.v1 or a documented long-term public mapping from the risk market to openagents.aegis.markets.v1
  • openagents.policy.v1
  • openagents.audit.v1
  • openagents.safety.v1
  • openagents.compliance.v1

Until that lands, this document should be read as a plan for the desired full package boundaries rather than a claim that the repo already ships complete wire coverage for every market.

1.3 Updated proto/openagents/common/v1/common.proto

This replaces your earlier common proto with the missing paper primitives: tfb, severity, tiers, provenance, and correlation metadata—plus optional receipt hints so /stats can be derived without fragile parsing.

syntax = "proto3";

package openagents.common.v1;

import "google/protobuf/struct.proto";

option java_multiple_files = true;

// -----------------------------
// Economy-wide enums
// -----------------------------

enum Asset {
  ASSET_UNSPECIFIED = 0;
  BTC_LN = 1;
  BTC_ONCHAIN = 2;
  USD = 3;
  USDC = 4;
}

enum FeedbackLatencyClass {
  FEEDBACK_LATENCY_CLASS_UNSPECIFIED = 0;
  INSTANT = 1;  // seconds/minutes
  SHORT = 2;    // hours/days
  LONG = 3;     // weeks/months
  UNKNOWN = 4;  // treated as LONG by default policy
}

enum SeverityClass {
  SEVERITY_CLASS_UNSPECIFIED = 0;
  LOW = 1;
  MEDIUM = 2;
  HIGH = 3;
  CRITICAL = 4;
}

enum VerificationTier {
  VERIFICATION_TIER_UNSPECIFIED = 0;
  TIER_O_OBJECTIVE = 1;
  TIER_1_CORRELATED = 2;
  TIER_2_HETEROGENEOUS = 3;
  TIER_3_ADJUDICATION = 4;
  TIER_4_HUMAN = 5;
}

enum ProvenanceGrade {
  PROVENANCE_GRADE_UNSPECIFIED = 0;
  P0_MINIMAL = 1;   // minimal receipts
  P1_TOOLCHAIN = 2; // artifacts + tool chain evidence
  P2_LINEAGE = 3;   // + model/checker lineage metadata
  P3_ATTESTED = 4;  // + attestations (signers/hardware/runtime)
}

enum AuthAssuranceLevel {
  AUTH_ASSURANCE_LEVEL_UNSPECIFIED = 0;
  ANON = 1;
  AUTHENTICATED = 2;
  ORG_KYC = 3;
  PERSONHOOD = 4;
  GOV_ID = 5;
  HARDWARE_BOUND = 6;
}

enum CostAttestationLevel {
  COST_ATTESTATION_LEVEL_UNSPECIFIED = 0;
  C0_SELF_REPORTED = 1;
  C1_METERED = 2;
  C2_HARDWARE_ATTESTED = 3;
  C3_ZK_PROVEN = 4; // optional/future
}

// -----------------------------
// Money and trace/policy context
// -----------------------------

message Money {
  Asset asset = 1;

  oneof amount {
    uint64 amount_msats = 2;
    uint64 amount_sats = 3;
  }
}

message TraceContext {
  string session_id = 1;
  string trajectory_hash = 2;
  string job_hash = 3;

  string run_id = 4;
  string work_unit_id = 5;
  string contract_id = 6;
  string claim_id = 7;
}

message PolicyContext {
  string policy_bundle_id = 1;
  string policy_version = 2;
  string approved_by = 3; // npub / org id / service principal
}

message CredentialRef {
  string issuer = 1;
  string subject = 2;
  string credential_type = 3; // did_vc, org_kyc, personhood_proof, gov_id, hardware_attestation
  string uri = 4;
  string digest = 5; // "sha256:<hex>"
  google.protobuf.Struct meta = 10;
}

message IdentityAssurance {
  string subject_id = 1;
  AuthAssuranceLevel level = 2;
  CredentialRef credential = 3;
}

message ComputeCost {
  uint64 gpu_ms = 1;
  uint64 cpu_ms = 2;
  uint64 ram_byte_ms = 3;
  uint64 disk_byte_ms = 4;
  uint64 egress_bytes = 5;
  uint64 ingress_bytes = 6;
  uint64 tokens_in = 10;  // optional
  uint64 tokens_out = 11; // optional
  google.protobuf.Struct meta = 20;
}

// -----------------------------
// Evidence and artifacts
// -----------------------------

message ArtifactRef {
  string uri = 1;       // s3://, ipfs://, oa://, file://
  string digest = 2;    // "sha256:<hex>"
  string mime_type = 3;
  uint64 size_bytes = 4;
  google.protobuf.Struct meta = 10;
}

message EvidenceRef {
  string kind = 1;      // e.g. "replay_bundle", "test_report", "provenance_bundle"
  string uri = 2;       // transport pointer only
  string digest = 3;    // "sha256:<hex>" (authoritative identity)
  google.protobuf.Struct meta = 10;
}

// -----------------------------
// Provenance (typed bundle)
// -----------------------------

message ModelIdentity {
  string provider = 1;     // openai, anthropic, local, etc.
  string family = 2;       // gpt-5, claude-4, llama, etc.
  string name = 3;         // model name
  string version = 4;      // build tag / semver / commit
  string digest = 5;       // optional: weights digest, image digest, etc.
  google.protobuf.Struct meta = 10;
}

message ToolInvocation {
  string tool_id = 1;          // "oa.git.apply_patch", "oa.sandbox_run", ...
  string tool_version = 2;     // semver/commit
  string request_digest = 3;   // "sha256:<hex>" of canonical request
  string response_digest = 4;  // "sha256:<hex>" of canonical response
  int64 started_at_ms = 5;
  int64 ended_at_ms = 6;
  google.protobuf.Struct meta = 10;
}

message AttestationRef {
  string kind = 1;     // "threshold_sig", "tpm_quote", "sigstore", "human_signoff"
  string uri = 2;
  string digest = 3;   // "sha256:<hex>"
  google.protobuf.Struct meta = 10;
}

message ProvenanceBundle {
  // This bundle itself should be stored as an evidence object with a digest;
  // receipts should reference it by EvidenceRef (kind="provenance_bundle").
  repeated ArtifactRef inputs = 1;
  repeated ArtifactRef outputs = 2;

  repeated ToolInvocation tool_invocations = 3;

  // Who produced the output (worker model/tooling) and who verified it (checker lineages).
  repeated ModelIdentity producer_lineage = 4;
  repeated ModelIdentity checker_lineage = 5;

  // Correlation groups allow policy to reject "AI verifies AI" setups.
  repeated string correlation_groups = 6; // e.g. "same_provider", "same_finetune", "shared_toolchain"

  repeated AttestationRef attestations = 7;

  // Explicit process-verification references.
  repeated EvidenceRef data_source_refs = 8;
  repeated EvidenceRef permissioning_refs = 9;
  repeated IdentityAssurance identities_involved = 10;

  google.protobuf.Struct meta = 20;
}

message CostProofBundle {
  string bundle_id = 1;
  string digest = 2; // "sha256:<hex>"
  CostAttestationLevel level = 3;

  ComputeCost cost = 10;

  // Bind cost proof to a specific execution identity.
  TraceContext trace = 20;

  // Meter/runtime logs and attestations are content-addressed.
  repeated EvidenceRef meter_logs = 30;
  repeated AttestationRef attestations = 31;
}

// -----------------------------
// Verification independence constraints + report
// -----------------------------

message IndependenceRequirement {
  // Minimum distinct checker lineages required (by identity digest or provider+family).
  uint32 min_distinct_checker_lineages = 1;

  // Correlation groups that are disallowed for unlocking settlement/warranty.
  repeated string disallowed_correlation_groups = 2;

  // If true, at least one Tier 4 human step is required.
  bool require_human_step = 3;
}

message IndependenceReport {
  repeated ModelIdentity checkers_used = 1;
  repeated string correlation_groups_present = 2;

  uint32 distinct_checker_lineages_observed = 3;

  // True iff correlation_groups_present intersects disallowed list (or other policy checks fail).
  bool correlation_violation = 4;
}

// -----------------------------
// Receipt primitives
// -----------------------------

message ReceiptRef {
  string receipt_id = 1;
  string receipt_type = 2;
  string canonical_hash = 3; // "sha256:<hex>"
}

message ReceiptHints {
  // Optional standardized hints so /stats can aggregate reliably.
  string category = 1; // domain/category, e.g. "compute", "autopilot", "legal", "ops"
  FeedbackLatencyClass tfb_class = 2;
  SeverityClass severity = 3;

  // Verification/provenance hints where applicable.
  VerificationTier achieved_verification_tier = 4;
  optional bool verification_correlated = 5;
  ProvenanceGrade provenance_grade = 6;
  AuthAssuranceLevel auth_assurance_level = 7;
  bool personhood_proved = 8;
  string reason_code = 9; // stable code, e.g. QUOTE_EXPIRED, BREAKER_ACTIVE, INSUFFICIENT_BUDGET

  // Optional: for economics rollups.
  Money notional = 10;
  Money liability_premium = 11;
}

message Receipt {
  string receipt_id = 1;
  string receipt_type = 2;
  int64 created_at_ms = 3;

  string canonical_hash = 4; // "sha256:<hex>"
  string idempotency_key = 5;

  TraceContext trace = 10;
  PolicyContext policy = 11;

  string inputs_hash = 20;  // "sha256:<hex>"
  string outputs_hash = 21; // "sha256:<hex>"

  repeated EvidenceRef evidence = 30;

  ReceiptHints hints = 40;

  map<string, string> tags = 50; // non-normative, excluded from canonical_hash
}

// -----------------------------
// Verification budget hint (paper B / capacity)
// -----------------------------

message VerificationBudgetHint {
  // Optional. This is a *hint* used for planning/gating; policy remains authoritative.
  // Include one or more; leaving empty means policy defaults.
  Money max_spend = 1;              // sats/msats budget for verification/underwriting
  uint32 max_human_minutes = 2;      // max human review minutes (if tracked)
  uint32 max_checker_runs = 3;       // max automated verification runs
  google.protobuf.Struct meta = 10;
}

Additional requirement for attestation semantics:

  • AttestationRef.kind and/or CredentialRef.credential_type MUST support named classes for:
    • model version attestation (verifiable inference)
    • runtime integrity attestation (execution attestation) so policy can require them by category/tfb/severity deterministically.

1.4 Updated proto/openagents/aegis/outcomes/v1/outcomes_work.proto

Key changes:

  • WorkUnit now requires tfb_class, severity, and a verification_budget_hint.
  • Contract + VerificationPlan now include provenance requirements and independence requirements.
  • Submission includes a provenance bundle evidence ref (typed by kind="provenance_bundle").
  • Verdict includes computed provenance_grade and an IndependenceReport, and sets correlation flags into receipts.
syntax = "proto3";

package openagents.aegis.outcomes.v1;

import "google/api/annotations.proto";
import "openagents/common/v1/common.proto";
import "openagents/hydra/v1/abp_bonds.proto";

option java_multiple_files = true;

enum WorkUnitKind {
  WORK_UNIT_KIND_UNSPECIFIED = 0;
  COMPUTE_JOB = 1;
  SKILL_INVOCATION = 2;
  ARTIFACT_BUNDLE = 3;
  L402_SERIES = 4;
}

enum VerifiabilityClass {
  VERIFIABILITY_CLASS_UNSPECIFIED = 0;
  OBJECTIVE = 1;
  SUBJECTIVE = 2;
  LOW_VERIFIABILITY = 3;
}

enum BinaryOutcome {
  BINARY_OUTCOME_UNSPECIFIED = 0;
  PASS = 1;
  FAIL = 2;
}

enum ContractState {
  CONTRACT_STATE_UNSPECIFIED = 0;
  CREATED = 1;
  FUNDED = 2;
  BONDED = 3;
  SUBMITTED = 4;
  VERIFYING = 5;
  VERDICT_PASS = 6;
  VERDICT_FAIL = 7;
  SETTLED = 8;
  WARRANTY_ACTIVE = 9;
  CLAIM_OPEN = 10;
  FINALIZED = 11;
  CANCELLED = 12;
}

enum RemedyType {
  REMEDY_TYPE_UNSPECIFIED = 0;
  REFUND = 1;
  REWORK_CREDIT = 2;
  DAMAGES = 3;
  ROLLBACK = 4;
  COMPENSATING_ACTION = 5;
}

enum ClaimState {
  CLAIM_STATE_UNSPECIFIED = 0;
  OPEN = 1;
  UNDER_REVIEW = 2;
  APPROVED = 3;
  DENIED = 4;
  PARTIALLY_APPROVED = 5;
  PAID = 6;
  CLOSED = 7;
}

message AcceptanceCriteria {
  string description = 1;

  // Optional: objective harness or rubric pointers.
  openagents.common.v1.ArtifactRef harness_ref = 10;           // for Tier O
  openagents.common.v1.EvidenceRef rubric_ref = 11;            // for subjective scoring
  openagents.common.v1.EvidenceRef adjudication_policy_ref = 12;

  repeated openagents.common.v1.ArtifactRef required_artifacts = 20;
  repeated openagents.common.v1.EvidenceRef required_evidence = 21;

  // Required by policy for high-severity irreversible lanes.
  RollbackPlan rollback_plan = 30;
  MonitoringPlan monitoring_plan = 31;
}

message RollbackPlan {
  bool required = 1;
  openagents.common.v1.EvidenceRef playbook_ref = 2;
  uint64 deadline_ms = 3;
}

message MonitoringPlan {
  repeated openagents.common.v1.EvidenceRef drift_detectors = 1;
  repeated string monitored_metrics = 2;
}

message WorkUnit {
  string work_unit_id = 1;
  WorkUnitKind kind = 2;

  string category = 3;    // required for sv/Δm/XA breakdowns
  string title = 4;
  string description = 5;

  string creator_id = 6;  // buyer/operator

  AcceptanceCriteria acceptance = 10;

  // --- Paper-critical metadata (normative in the kernel spec) ---
  openagents.common.v1.FeedbackLatencyClass tfb_class = 20;
  openagents.common.v1.SeverityClass severity = 21;
  openagents.common.v1.VerificationBudgetHint verification_budget_hint = 22;

  openagents.common.v1.TraceContext trace = 30;
  openagents.common.v1.PolicyContext policy = 31;

  int64 created_at_ms = 40;
}

message BondRequirement {
  openagents.hydra.v1.BondPartyRole party_role = 1;
  openagents.hydra.v1.BondReason reason = 2;
  openagents.common.v1.Money min_amount = 3;
  openagents.common.v1.Money max_amount = 4;
}

message WarrantyTerms {
  RemedyType remedy = 1;
  openagents.common.v1.Money coverage_cap = 2;
  uint64 warranty_window_ms = 3;
  repeated string exclusions = 10;
}

message ProvenanceRequirement {
  // Minimum provenance grade needed to unlock settlement / warranty issuance.
  openagents.common.v1.ProvenanceGrade min_grade = 1;

  // Require provenance bundle evidence ref on submission.
  bool require_provenance_bundle = 2;

  // Require at least one attestation in the provenance bundle (P3 paths).
  bool require_attestation = 3;
}

enum VerificationAssignmentMode {
  VERIFICATION_ASSIGNMENT_MODE_UNSPECIFIED = 0;
  POLICY_SELECTED_SET = 1;
  BUYER_SELECTED = 2;
  VERIFIER_MARKET = 3; // optional lane; policy-gated
}

message VerificationAssignment {
  VerificationAssignmentMode mode = 1;

  // Explicitly assigned verifier/adjudicator identities for direct-assignment modes.
  repeated string verifier_ids = 2;

  // Hash-bound reference to the assignment receipt/object used for verdict gating.
  openagents.common.v1.ReceiptRef assignment_ref = 3;

  // Optional market match reference when mode == VERIFIER_MARKET.
  openagents.common.v1.EvidenceRef market_match_ref = 4;

  int64 assigned_at_ms = 10;
}

message VerifierLiabilityPolicy {
  // Default false: verifier is economically non-liable (reputationally accountable only).
  bool economic_liability_enabled = 1;

  // Required when economic_liability_enabled=true.
  openagents.common.v1.Money max_exposure = 2;
  openagents.common.v1.EvidenceRef slash_policy_ref = 3;
}

message VerificationPlan {
  VerifiabilityClass verifiability = 1;

  // Required minimum tier to unlock settlement/warranty.
  openagents.common.v1.VerificationTier required_tier = 2;

  // Independence constraints (heterogeneity / correlation gating).
  openagents.common.v1.IndependenceRequirement independence_requirement = 3;

  // Provenance constraints.
  ProvenanceRequirement provenance_requirement = 4;

  repeated openagents.common.v1.EvidenceRef planned_checks = 10;
  openagents.common.v1.EvidenceRef adjudication_policy_ref = 11;
}

message OutcomeContract {
  string contract_id = 1;
  string work_unit_id = 2;

  string buyer_id = 10;
  string worker_id = 11;
  string underwriter_group_id = 12;

  // Explicit pricing split (insurance boundary).
  openagents.common.v1.Money execution_price = 20;
  openagents.common.v1.Money verification_fee = 21;
  openagents.common.v1.Money liability_premium = 22;
  openagents.common.v1.ReceiptRef pricing_snapshot_ref = 23;

  VerificationPlan verification_plan = 30;
  repeated BondRequirement bond_requirements = 31;
  VerificationAssignment verification_assignment = 32;
  VerifierLiabilityPolicy verifier_liability_policy = 33;

  bool warranty_enabled = 40;
  WarrantyTerms warranty_terms = 41;

  ContractState state = 50;

  openagents.common.v1.TraceContext trace = 60;
  openagents.common.v1.PolicyContext policy = 61;

  int64 created_at_ms = 70;
  int64 updated_at_ms = 71;
}

message BondLink {
  string bond_id = 1;
  openagents.hydra.v1.BondPartyRole party_role = 2;
  openagents.hydra.v1.BondReason reason = 3;
  openagents.common.v1.Money amount_reserved = 4;
}

message Submission {
  repeated openagents.common.v1.ArtifactRef outputs = 1;
  repeated openagents.common.v1.EvidenceRef evidence = 2;

  // REQUIRED when provenance_requirement.require_provenance_bundle = true
  openagents.common.v1.EvidenceRef provenance_bundle = 3; // kind must be "provenance_bundle"

  // REQUIRED when policy requires proof-of-cost for compute-like lanes.
  openagents.common.v1.EvidenceRef cost_proof_bundle = 4; // kind must be "cost_proof_bundle"

  int64 submitted_at_ms = 10;
}

message Verdict {
  BinaryOutcome outcome = 1;

  openagents.common.v1.VerificationTier achieved_tier = 2;

  // Computed from provenance bundle contents (deterministic).
  openagents.common.v1.ProvenanceGrade provenance_grade = 3;

  // Required: independence report of checkers used.
  openagents.common.v1.IndependenceReport independence_report = 4;

  // Evidence pointers (harness output, adjudication notes, etc.)
  repeated openagents.common.v1.EvidenceRef evidence = 10;

  // References to Hydra/wallet settlement receipts that happened as a consequence.
  repeated openagents.common.v1.ReceiptRef settlement_receipts = 20;

  int64 decided_at_ms = 30;
}

message Warranty {
  bool active = 1;
  WarrantyTerms terms = 2;
  int64 starts_at_ms = 3;
  int64 ends_at_ms = 4;
}

message Claim {
  string claim_id = 1;
  string contract_id = 2;

  string claimant_id = 10;
  string reason = 11;

  repeated openagents.common.v1.EvidenceRef evidence = 12;

  ClaimState state = 20;

  BondLink dispute_bond = 30;

  int64 opened_at_ms = 40;
  int64 updated_at_ms = 41;
}

message ClaimResolution {
  ClaimState final_state = 1;
  openagents.common.v1.Money payout = 2;

  repeated BondLink bonds_drawn = 3;
  repeated openagents.common.v1.EvidenceRef evidence = 10;
  repeated openagents.common.v1.ReceiptRef settlement_receipts = 20;

  int64 resolved_at_ms = 30;
}

// ---- RPCs ----

message CreateWorkUnitRequest {
  WorkUnitKind kind = 1;
  string category = 2;
  string title = 3;
  string description = 4;
  string creator_id = 5;

  AcceptanceCriteria acceptance = 10;

  // Required paper-critical metadata.
  openagents.common.v1.FeedbackLatencyClass tfb_class = 20;
  openagents.common.v1.SeverityClass severity = 21;
  openagents.common.v1.VerificationBudgetHint verification_budget_hint = 22;

  string idempotency_key = 50;
  openagents.common.v1.TraceContext trace = 51;
  openagents.common.v1.PolicyContext policy = 52;
}

message CreateWorkUnitResponse {
  WorkUnit work_unit = 1;
  openagents.common.v1.Receipt receipt = 2; // hints MUST include category/tfb/severity
}

message GetWorkUnitRequest { string work_unit_id = 1; }
message GetWorkUnitResponse { WorkUnit work_unit = 1; }

message CreateContractRequest {
  string work_unit_id = 1;

  string buyer_id = 10;
  string worker_id = 11;

  openagents.common.v1.Money execution_price = 20;
  openagents.common.v1.Money verification_fee = 21;
  openagents.common.v1.Money liability_premium = 22;
  openagents.common.v1.ReceiptRef pricing_snapshot_ref = 23;

  VerificationPlan verification_plan = 30;
  repeated BondRequirement bond_requirements = 31;
  VerificationAssignment verification_assignment = 32;
  VerifierLiabilityPolicy verifier_liability_policy = 33;

  bool warranty_enabled = 40;
  WarrantyTerms warranty_terms = 41;

  string idempotency_key = 60;
  openagents.common.v1.TraceContext trace = 61;
  openagents.common.v1.PolicyContext policy = 62;
}

message CreateContractResponse {
  OutcomeContract contract = 1;
  openagents.common.v1.Receipt receipt = 2;
}

message AttachBondRequest {
  string contract_id = 1;
  string bond_id = 2;
  openagents.hydra.v1.BondPartyRole party_role = 3;
  openagents.hydra.v1.BondReason reason = 4;

  string idempotency_key = 20;
  openagents.common.v1.TraceContext trace = 21;
  openagents.common.v1.PolicyContext policy = 22;
}

message AttachBondResponse {
  repeated BondLink bonds = 1;
  openagents.common.v1.Receipt receipt = 2;
}

message SubmitOutputRequest {
  string contract_id = 1;
  Submission submission = 2;

  string idempotency_key = 20;
  openagents.common.v1.TraceContext trace = 21;
  openagents.common.v1.PolicyContext policy = 22;
}

message SubmitOutputResponse {
  OutcomeContract contract = 1;
  openagents.common.v1.Receipt receipt = 2; // MUST include provenance grade once computed (if available)
}

message StartVerificationRequest {
  string contract_id = 1;

  string idempotency_key = 20;
  openagents.common.v1.TraceContext trace = 21;
  openagents.common.v1.PolicyContext policy = 22;
}

message StartVerificationResponse {
  OutcomeContract contract = 1;
  openagents.common.v1.Receipt receipt = 2;
}

message FinalizeVerdictRequest {
  string contract_id = 1;
  Verdict verdict = 2;

  string idempotency_key = 20;
  openagents.common.v1.TraceContext trace = 21;
  openagents.common.v1.PolicyContext policy = 22;
}

message FinalizeVerdictResponse {
  OutcomeContract contract = 1;
  Verdict verdict = 2;
  openagents.common.v1.Receipt receipt = 3; // hints MUST include tier + correlated flag + provenance grade
}

message OpenClaimRequest {
  string contract_id = 1;
  string claimant_id = 10;
  string reason = 11;
  repeated openagents.common.v1.EvidenceRef evidence = 12;

  string dispute_bond_id = 20; // optional Hydra ABP bond

  string idempotency_key = 30;
  openagents.common.v1.TraceContext trace = 31;
  openagents.common.v1.PolicyContext policy = 32;
}

message OpenClaimResponse {
  Claim claim = 1;
  openagents.common.v1.Receipt receipt = 2;
}

message ResolveClaimRequest {
  string claim_id = 1;
  ClaimState final_state = 2;
  openagents.common.v1.Money payout = 3;
  repeated openagents.common.v1.EvidenceRef evidence = 10;

  string idempotency_key = 20;
  openagents.common.v1.TraceContext trace = 21;
  openagents.common.v1.PolicyContext policy = 22;
}

message ResolveClaimResponse {
  Claim claim = 1;
  ClaimResolution resolution = 2;
  openagents.common.v1.Receipt receipt = 3;
}

message GetStatusRequest { string contract_id = 1; }

message ContractStatus {
  OutcomeContract contract = 1;
  repeated BondLink bonds = 2;

  Submission last_submission = 3;

  bool has_verdict = 10;
  Verdict verdict = 11;

  bool warranty_active = 20;
  Warranty warranty = 21;

  repeated Claim claims = 30;
}

message GetStatusResponse { ContractStatus status = 1; }

service OutcomesWorkService {
  rpc CreateWorkUnit(CreateWorkUnitRequest) returns (CreateWorkUnitResponse) {
    option (google.api.http) = { post: "/v1/aegis/outcomes/work_units" body: "*" };
  }

  rpc GetWorkUnit(GetWorkUnitRequest) returns (GetWorkUnitResponse) {
    option (google.api.http) = { get: "/v1/aegis/outcomes/work_units/{work_unit_id}" };
  }

  rpc CreateContract(CreateContractRequest) returns (CreateContractResponse) {
    option (google.api.http) = { post: "/v1/aegis/outcomes/contracts" body: "*" };
  }

  rpc AttachBond(AttachBondRequest) returns (AttachBondResponse) {
    option (google.api.http) = { post: "/v1/aegis/outcomes/contracts/{contract_id}/bonds/attach" body: "*" };
  }

  rpc SubmitOutput(SubmitOutputRequest) returns (SubmitOutputResponse) {
    option (google.api.http) = { post: "/v1/aegis/outcomes/contracts/{contract_id}/submit" body: "*" };
  }

  rpc StartVerification(StartVerificationRequest) returns (StartVerificationResponse) {
    option (google.api.http) = { post: "/v1/aegis/outcomes/contracts/{contract_id}/verify/start" body: "*" };
  }

  rpc FinalizeVerdict(FinalizeVerdictRequest) returns (FinalizeVerdictResponse) {
    option (google.api.http) = { post: "/v1/aegis/outcomes/contracts/{contract_id}/verdict/finalize" body: "*" };
  }

  rpc OpenClaim(OpenClaimRequest) returns (OpenClaimResponse) {
    option (google.api.http) = { post: "/v1/aegis/outcomes/contracts/{contract_id}/claims" body: "*" };
  }

  rpc ResolveClaim(ResolveClaimRequest) returns (ResolveClaimResponse) {
    option (google.api.http) = { post: "/v1/aegis/outcomes/claims/{claim_id}/resolve" body: "*" };
  }

  rpc GetStatus(GetStatusRequest) returns (GetStatusResponse) {
    option (google.api.http) = { get: "/v1/aegis/outcomes/contracts/{contract_id}/status" };
  }
}

1.5 New: proto/openagents/aegis/incidents/v1/incidents.proto

This makes incident reporting proto-first and insurability-ready (versioned taxonomy + mandatory receipt/evidence linkage + outcome registry entries).

Normative implementation notes:

  • Incident objects are append-only, hash-addressed revisions; updates/resolutions supersede prior digests, not overwrite them.
  • Taxonomy meaning is immutable for a (taxonomy_id, taxonomy_version, code) tuple.
  • Export surfaces MUST support at least public and restricted redaction tiers for incident audit packages.
  • Long-latency external truth inputs MUST be represented as explicit GroundTruthEvidence objects with source identity, digest-bound evidence ref, and received_at_ms.
  • Dispute truth finalization MUST be receipted and may resolve only via objective harness, adjudication policy, or explicit human underwriting.
  • Outcome registry export posture is tiered:
    • public exports include aggregate-safe fields only (slice keys, normalized outcomes, taxonomy buckets, counts/timestamps),
    • restricted exports may include linked receipt refs and evidence digests needed for audit replay.
syntax = "proto3";

package openagents.aegis.incidents.v1;

import "google/api/annotations.proto";
import "openagents/common/v1/common.proto";

option java_multiple_files = true;

enum IncidentKind {
  INCIDENT_KIND_UNSPECIFIED = 0;
  INCIDENT = 1;
  NEAR_MISS = 2;
  GROUND_TRUTH_CASE = 3;
}

enum IncidentState {
  INCIDENT_STATE_UNSPECIFIED = 0;
  OPEN = 1;
  TRIAGED = 2;
  MITIGATED = 3;
  RESOLVED = 4;
}

message IncidentTaxonomyCode {
  string taxonomy_id = 1;      // "oa.incident.taxonomy"
  string taxonomy_version = 2; // "2026.02"
  string code = 3;             // e.g. PAYOUT_POINTER_MISMATCH
  string label = 4;
}

message DisclosureBundleRef {
  openagents.common.v1.EvidenceRef disclosure_artifact = 1;
  string schema_version = 2;
}

message RollbackAttemptRef {
  string rollback_receipt_id = 1;
  string rollback_reason_code = 2;
  bool rollback_succeeded = 3;
}

message GroundTruthEvidence {
  string source_id = 1; // external registry/counterparty/authority/source system id
  openagents.common.v1.EvidenceRef evidence = 2; // digest-bound
  int64 received_at_ms = 3;
  string confidence_posture = 4; // provisional/final/contested/etc.
  repeated openagents.common.v1.ReceiptRef linked_receipts = 5; // contract/verdict/claim refs
}

message SimulationDerivationRef {
  string simulation_scenario_id = 1;
  repeated openagents.common.v1.EvidenceRef source_incident_digests = 2;
  openagents.common.v1.IdentityAssurance generated_by = 3;
  openagents.common.v1.IdentityAssurance validated_by = 4;
  openagents.common.v1.ReceiptRef derivation_receipt = 5;
  openagents.common.v1.ReceiptRef validation_receipt = 6;
}

message IncidentReport {
  string incident_id = 1;
  IncidentKind kind = 2;
  IncidentState state = 3;
  IncidentTaxonomyCode taxonomy = 4;
  openagents.common.v1.SeverityClass severity = 5;
  string category = 6;
  openagents.common.v1.FeedbackLatencyClass tfb_class = 7;

  int64 occurred_at_ms = 10;
  int64 reported_at_ms = 11;
  int64 resolved_at_ms = 12;

  string summary = 20;
  string impact_notes = 21;
  double estimated_loss_ratio_delta = 22;

  repeated openagents.common.v1.ReceiptRef linked_receipts = 30;   // contract/verdict/settlement
  repeated openagents.common.v1.EvidenceRef evidence_digests = 31; // no raw payload
  repeated RollbackAttemptRef rollback_attempts = 32;
  DisclosureBundleRef disclosure_bundle = 33;
  repeated GroundTruthEvidence ground_truth_evidence = 34; // required for delayed/external truth lanes
  repeated SimulationDerivationRef simulation_derivations = 35;

  openagents.common.v1.PolicyContext policy = 40;
  map<string, string> tags = 50;
}

message OutcomeRegistryEntry {
  string entry_id = 1;
  string work_unit_id = 2;
  string category = 3;
  openagents.common.v1.FeedbackLatencyClass tfb_class = 4;
  openagents.common.v1.SeverityClass severity = 5;

  string verdict_outcome = 10;
  string settlement_outcome = 11;
  string claim_outcome = 12;
  string remedy_outcome = 13;

  repeated IncidentTaxonomyCode incident_tags = 20;             // public-safe bucket labels
  repeated openagents.common.v1.ReceiptRef linked_receipts = 21; // restricted tier
  repeated openagents.common.v1.EvidenceRef evidence_digests = 22; // restricted tier
}

message ReportIncidentRequest {
  IncidentReport incident = 1;
  string idempotency_key = 10;
  openagents.common.v1.TraceContext trace = 11;
  openagents.common.v1.PolicyContext policy = 12;
}

message ReportIncidentResponse {
  IncidentReport incident = 1;
  openagents.common.v1.Receipt receipt = 2; // economy.incident.reported.v1 / updated / resolved
}

message UpsertOutcomeRegistryEntryRequest {
  OutcomeRegistryEntry entry = 1;
  string idempotency_key = 10;
  openagents.common.v1.TraceContext trace = 11;
  openagents.common.v1.PolicyContext policy = 12;
}

message UpsertOutcomeRegistryEntryResponse {
  OutcomeRegistryEntry entry = 1;
  openagents.common.v1.Receipt receipt = 2; // economy.outcome_registry.upserted.v1
}

service IncidentService {
  rpc ReportIncident(ReportIncidentRequest) returns (ReportIncidentResponse) {
    option (google.api.http) = { post: "/v1/aegis/incidents:report" body: "*" };
  }

  rpc UpsertOutcomeRegistryEntry(UpsertOutcomeRegistryEntryRequest)
      returns (UpsertOutcomeRegistryEntryResponse) {
    option (google.api.http) = { post: "/v1/aegis/incidents/outcome_registry:upsert" body: "*" };
  }
}

1.6 New: proto/openagents/compliance/v1/certification.proto

This adds first-class certification state that can be enforced by policy and exported for cross-border/safe-harbor checks.

syntax = "proto3";

package openagents.compliance.v1;

import "google/api/annotations.proto";
import "openagents/common/v1/common.proto";

option java_multiple_files = true;

enum CertificationState {
  CERTIFICATION_STATE_UNSPECIFIED = 0;
  ACTIVE = 1;
  REVOKED = 2;
  EXPIRED = 3;
}

message CertificationScope {
  string category = 1;
  openagents.common.v1.FeedbackLatencyClass tfb_class = 2;
  openagents.common.v1.SeverityClass min_severity = 3;
  openagents.common.v1.SeverityClass max_severity = 4;
}

message SafetyCertification {
  string certification_id = 1;
  CertificationState state = 2;
  string certification_level = 3;
  repeated CertificationScope scope = 4;

  int64 valid_from_ms = 10;
  int64 valid_until_ms = 11;

  openagents.common.v1.CredentialRef issuer = 20;
  repeated openagents.common.v1.EvidenceRef required_evidence = 21; // audits/provenance/incidents summary

  openagents.common.v1.PolicyContext policy = 30;
  map<string, string> tags = 40;
}

message IssueCertificationRequest {
  SafetyCertification certification = 1;
  string idempotency_key = 10;
  openagents.common.v1.TraceContext trace = 11;
  openagents.common.v1.PolicyContext policy = 12;
}

message IssueCertificationResponse {
  SafetyCertification certification = 1;
  openagents.common.v1.Receipt receipt = 2; // economy.certification.issued.v1
}

message RevokeCertificationRequest {
  string certification_id = 1;
  string reason_code = 2;
  repeated openagents.common.v1.EvidenceRef evidence = 10;
  string idempotency_key = 20;
  openagents.common.v1.TraceContext trace = 21;
  openagents.common.v1.PolicyContext policy = 22;
}

message RevokeCertificationResponse {
  SafetyCertification certification = 1;
  openagents.common.v1.Receipt receipt = 2; // economy.certification.revoked.v1
}

1.7 New: proto/openagents/safety/v1/safety_signals.proto

This enables privacy-preserving safety signal sharing: public aggregates and restricted feeds from the same deterministic source objects.

syntax = "proto3";

package openagents.safety.v1;

import "google/api/annotations.proto";
import "openagents/common/v1/common.proto";

option java_multiple_files = true;

enum SafetySignalVisibility {
  SAFETY_SIGNAL_VISIBILITY_UNSPECIFIED = 0;
  PUBLIC_AGGREGATE = 1;
  RESTRICTED_FEED = 2;
}

message SafetySignal {
  string signal_id = 1;
  int64 emitted_at_ms = 2;

  string category = 10;
  openagents.common.v1.FeedbackLatencyClass tfb_class = 11;
  openagents.common.v1.SeverityClass severity = 12;
  string incident_taxonomy_code = 13;

  repeated string hashed_indicators = 20; // no raw payloads
  repeated openagents.common.v1.ReceiptRef source_receipts = 21;
  repeated openagents.common.v1.EvidenceRef source_evidence = 22;

  SafetySignalVisibility visibility = 30;
  string restricted_audience_class = 31; // auditor / underwriter / verifier
}

message PublishSafetySignalRequest {
  SafetySignal signal = 1;
  string idempotency_key = 10;
  openagents.common.v1.TraceContext trace = 11;
  openagents.common.v1.PolicyContext policy = 12;
}

message PublishSafetySignalResponse {
  SafetySignal signal = 1;
  openagents.common.v1.Receipt receipt = 2; // economy.safety_signal.published.v1
}

1.8 New: proto/openagents/audit/v1/audit_package.proto

This defines deterministic interoperable exports (public/restricted redaction tiers) for underwriting, incident forensics, and certification replay.

syntax = "proto3";

package openagents.audit.v1;

import "google/api/annotations.proto";
import "openagents/common/v1/common.proto";
import "openagents/aegis/incidents/v1/incidents.proto";
import "openagents/compliance/v1/certification.proto";

option java_multiple_files = true;

enum AuditRedactionTier {
  AUDIT_REDACTION_TIER_UNSPECIFIED = 0;
  PUBLIC = 1;
  RESTRICTED = 2;
}

message AuditPackage {
  string package_id = 1;
  string package_hash = 2; // deterministic hash over canonical export payload
  int64 generated_at_ms = 3;
  AuditRedactionTier redaction_tier = 4;

  repeated openagents.common.v1.Receipt receipts = 10;
  repeated openagents.aegis.incidents.v1.IncidentReport incidents = 11;
  repeated openagents.aegis.incidents.v1.OutcomeRegistryEntry outcome_registry_entries = 12;
  repeated openagents.compliance.v1.SafetyCertification certifications = 13;

  repeated openagents.common.v1.EvidenceRef snapshot_refs = 20;
  repeated openagents.common.v1.EvidenceRef anchor_refs = 21; // hash-only anchoring proofs
}

message ExportAuditPackageRequest {
  int64 start_inclusive_ms = 1;
  int64 end_inclusive_ms = 2;
  string category = 3;
  AuditRedactionTier redaction_tier = 4;

  string idempotency_key = 10;
  openagents.common.v1.TraceContext trace = 11;
  openagents.common.v1.PolicyContext policy = 12;
}

message ExportAuditPackageResponse {
  AuditPackage package = 1;
  openagents.common.v1.Receipt receipt = 2; // economy.audit_package.exported.v1
}

1.9 Updated: proto/openagents/economy/v1/economy_snapshot.proto (sv / Δm_hat / XA_hat receipts)

This expands the minute snapshot to include insurer/monitoring/interoperability metrics and the correlation-adjusted trust headline used for autonomy gating.

syntax = "proto3";

package openagents.economy.v1;

import "google/api/annotations.proto";
import "openagents/common/v1/common.proto";

option java_multiple_files = true;

message MetricKey {
  string category = 1;
  openagents.common.v1.FeedbackLatencyClass tfb_class = 2;
  openagents.common.v1.SeverityClass severity = 3;
  openagents.common.v1.VerificationTier verification_tier = 4;
  optional bool verification_correlated = 5;
  openagents.common.v1.ProvenanceGrade provenance_grade = 6;
}

message SvBreakdownRow {
  MetricKey key = 1;
  uint64 total_work_units = 2;
  uint64 verified_work_units = 3;
  double sv = 4;
}

message DistributionRow {
  string label = 1;
  uint64 count = 2;
  double share = 3;
}

message IncidentBucket {
  string taxonomy_code = 1;
  openagents.common.v1.SeverityClass severity = 2;
  uint64 incidents_24h = 3;
  uint64 near_misses_24h = 4;
  double incident_rate = 5;
  double near_miss_rate = 6;
}

message DriftSignalRow {
  string detector_id = 1;
  string signal_code = 2;
  uint64 count_24h = 3;
  double ratio = 4;
  double threshold = 5;
  double score = 6;
  bool alert = 7;
}

message RollbackReasonCodeRow {
  string reason_code = 1; // stable or redacted public-safe bucket label
  uint64 count_24h = 2;
}

message EconomySnapshot {
  string snapshot_id = 1;
  int64 as_of_ms = 2;
  string snapshot_hash = 3; // "sha256:<hex>"

  double sv = 10;
  double sv_effective = 11; // correlation-adjusted verifiable share
  double rho = 12;
  double rho_effective = 13;
  uint64 N = 14;
  double NV = 15;

  double delta_m_hat = 20;
  double xa_hat = 21;
  double correlated_verification_share = 22;

  double provenance_p0_share = 30;
  double provenance_p1_share = 31;
  double provenance_p2_share = 32;
  double provenance_p3_share = 33;

  openagents.common.v1.Money liability_premiums_collected_24h = 40;
  openagents.common.v1.Money claims_paid_24h = 41;
  double loss_ratio = 42;
  double capital_coverage_ratio = 43;

  uint64 drift_alerts_24h = 50;
  uint64 rollback_attempts_24h = 51;
  uint64 rollback_successes_24h = 52;
  repeated DriftSignalRow top_drift_signals = 53;
  double rollback_success_rate = 54;
  repeated RollbackReasonCodeRow top_rollback_reason_codes = 55;

  repeated IncidentBucket incident_buckets = 60;
  repeated DistributionRow auth_assurance_distribution = 61;
  double personhood_verified_share = 62;
  repeated DistributionRow certification_distribution = 63;

  // Market-backed verification-capacity and liability signals.
  double coverage_bound_share = 64;
  double avg_underwriters_per_contract = 65;
  double coverage_concentration_hhi = 66;
  double implied_fail_prob_p50 = 67;
  double implied_fail_prob_p95 = 68;
  double market_calibration_24h = 69;

  // Cost-proof and compute-integrity signals.
  double cost_proof_share = 70;
  repeated DistributionRow cost_attestation_distribution = 71;
  double cost_anomaly_rate_24h = 72;
  double cost_variance_p50 = 73;
  double cost_variance_p95 = 74;
  double price_vs_cost_spread_bps = 75;
  double market_disagreement_rate_24h = 76;
  uint64 market_manipulation_flags_24h = 77;
  double provider_cost_integrity_score = 78;

  repeated SvBreakdownRow sv_breakdown = 100;
  repeated openagents.common.v1.EvidenceRef inputs = 200;
}

message ComputeSnapshotRequest {
  int64 as_of_ms = 1;
  string idempotency_key = 10;
  openagents.common.v1.TraceContext trace = 11;
  openagents.common.v1.PolicyContext policy = 12;
}

message ComputeSnapshotResponse {
  EconomySnapshot snapshot = 1;
  openagents.common.v1.Receipt receipt = 2; // economy.stats.snapshot_receipt.v1
}

Notes:

  • sv_effective/rho_effective are the policy-gating trust variables when correlated checker risk is non-trivial.
  • coverage-backed and cost-proof metrics are optional modifiers; they are policy inputs and MUST NOT bypass explicit resolution-path invariants.
  • Snapshot receipts and any throttle/withhold actions MUST reference the exact snapshot_id/snapshot_hash used to make the decision.
  • Snapshot computation MUST sort included receipts deterministically by (created_at_ms ASC, receipt_id ASC) before aggregation/hashing.

1.10 New: proto/openagents/aegis/markets/v1/liability_market.proto (optional extension)

This adds an optional, policy-gated liability market surface that treats prediction markets as bounded, receipted underwriting capacity.

Coverage-first posture:

  • implement deterministic collateralized coverage offers/bindings first,
  • keep continuous belief-market mechanics optional and separable.
  • when matching/clearing multiple offers, prefer deterministic minute-batch clearing and emit explicit clearing receipts.
syntax = "proto3";

package openagents.aegis.markets.v1;

import "google/api/annotations.proto";
import "openagents/common/v1/common.proto";

option java_multiple_files = true;

enum CoverageState {
  COVERAGE_STATE_UNSPECIFIED = 0;
  OFFERING_OPEN = 1;
  BINDING_PROPOSED = 2;
  BOUND = 3;
  ACTIVE = 4;
  CLAIM_TRIGGERED = 5;
  EXPIRED = 6;
  SETTLED = 7;
}

message ResolutionRef {
  oneof ref {
    openagents.common.v1.ReceiptRef verdict_receipt = 1;
    openagents.common.v1.ReceiptRef claim_resolution_receipt = 2;
    openagents.common.v1.EvidenceRef objective_harness_ref = 3;
    openagents.common.v1.ReceiptRef human_underwrite_receipt = 4;
  }
}

message CoverageOffer {
  string offer_id = 1;
  string contract_id = 2;
  string underwriter_id = 3;

  openagents.common.v1.Money coverage_cap = 10;
  uint32 premium_bps = 11;

  // Collateral linkage (ABP reserve/draw/release receipts referenced in evidence).
  string bond_id = 20;
  repeated string correlation_groups = 30;
  uint64 expires_at_ms = 40;
}

message CoverageBinding {
  string binding_id = 1;
  string contract_id = 2;
  CoverageState state = 3;

  repeated CoverageOffer accepted_offers = 10;
  openagents.common.v1.Money total_coverage_cap = 20;
  uint32 blended_premium_bps = 21;

  ResolutionRef resolution = 30;
  uint64 warranty_window_ms = 31;
}

message MarketSignal {
  string contract_id = 1;
  double implied_fail_probability = 10;
  double confidence = 11;
  double calibration_score_24h = 12;
  double disagreement_rate_24h = 13;
}

message PlaceCoverageOfferRequest {
  CoverageOffer offer = 1;
  string idempotency_key = 10;
  openagents.common.v1.TraceContext trace = 11;
  openagents.common.v1.PolicyContext policy = 12;
}

message PlaceCoverageOfferResponse {
  CoverageOffer offer = 1;
  openagents.common.v1.Receipt receipt = 2; // market.coverage_offer.placed.v1
}

message BindCoverageRequest {
  string contract_id = 1;
  ResolutionRef resolution = 2;
  uint64 warranty_window_ms = 3;
  string idempotency_key = 10;
  openagents.common.v1.TraceContext trace = 11;
  openagents.common.v1.PolicyContext policy = 12;
}

message BindCoverageResponse {
  CoverageBinding binding = 1;
  openagents.common.v1.Receipt receipt = 2; // market.coverage.bound.v1
  MarketSignal signal = 3;
}

message SettleCoverageRequest {
  string binding_id = 1;
  ResolutionRef final_resolution = 2;
  string idempotency_key = 10;
  openagents.common.v1.TraceContext trace = 11;
  openagents.common.v1.PolicyContext policy = 12;
}

message SettleCoverageResponse {
  CoverageBinding binding = 1;
  openagents.common.v1.Receipt receipt = 2; // market.coverage.settled.v1
}

service LiabilityMarketService {
  rpc PlaceCoverageOffer(PlaceCoverageOfferRequest) returns (PlaceCoverageOfferResponse) {
    option (google.api.http) = { post: "/v1/aegis/markets/coverage_offers" body: "*" };
  }

  rpc BindCoverage(BindCoverageRequest) returns (BindCoverageResponse) {
    option (google.api.http) = { post: "/v1/aegis/markets/contracts/{contract_id}/coverage:bind" body: "*" };
  }

  rpc SettleCoverage(SettleCoverageRequest) returns (SettleCoverageResponse) {
    option (google.api.http) = { post: "/v1/aegis/markets/coverage/{binding_id}:settle" body: "*" };
  }
}

2. PolicyBundle schema rewrite (proto + semantics)

You need a PolicyBundle that can bind autonomy to sv/XA and correlation risk, plus set default verification/provenance rules by category, tfb, and severity.

2.1 What the policy must be able to express

At minimum, a PolicyBundle must let you set:

  1. Verification requirements
  • required tier by (category, tfb, severity)
  • independence constraints (min distinct checker lineages; disallowed correlation groups)
  • when Tier 1 is allowed (usually only LOW + INSTANT/SHORT)
  1. Provenance requirements
  • minimum provenance grade by (category, tfb, severity)
  • whether provenance bundle is mandatory
  • whether attestations are required at high severity
  1. Autonomy throttle rules (the paper’s key enforcement)

  • thresholds on sv, sv_effective, xa_hat, delta_m_hat, correlated verification share, drift alerts

  • actions when thresholds are crossed:

    • increase required tier
    • require human step
    • raise required provenance grade
    • reduce or halt envelope issuance for a category
    • require approvals for high-severity work
    • disable warranties or cap warranty issuance
  1. Envelope and bond parameters
  • envelope caps, expiries, fee bps by category/severity
  • bond multipliers for warranties and disputes
  1. Authentication and personhood gates
  • minimum AuthAssuranceLevel by (category, tfb, severity) and by role (VERIFIER, UNDERWRITER, OPERATOR, REQUESTER)
  • explicit personhood requirement for sensitive actions (e.g. verdict finalization, high-severity settlement/withdraw)
  1. Risk pricing / insurance boundary
  • base liability premium (bps) and dynamic multipliers keyed to snapshot metrics
  • reserve/capital requirements by slice and autonomy mode
  • explicit linkage to pricing_snapshot_ref
  1. Monitoring / drift rules
  • detector requirements by slice
  • deterministic trigger thresholds and action ordering for drift-induced throttles
  1. Certification + rollback requirements
  • required certification scopes for high-severity slices (“digital border”)
  • mandatory rollback or compensating-action plans for irreversible effect classes
  • safe-harbor relaxations only when certification + observed metrics satisfy policy
  1. Markets and coverage requirements
  • minimum coverage caps and maximum premium bounds by slice
  • underwriter diversity/concentration constraints
  • maximum implied fail probability and minimum calibration score for safe-harbor relaxations
  • optional safe-harbor relaxations only when market calibration and concentration constraints pass
  1. Cost-proof requirements
  • minimum cost attestation level by slice (C0–C3)
  • anomaly thresholds and actions for measured-vs-expected cost drift
  • optional minimum provider cost-integrity score thresholds for underwriting/routing priors
  • whether cost proof is required to unlock settlement/warranty in compute-like lanes
  1. Interpretation rules
  • which time windows to use (5m/1h/24h) for gating
  • hysteresis / cooldown to avoid flapping breakers

2.2 proto/openagents/policy/v1/policy_bundle.proto

syntax = "proto3";

package openagents.policy.v1;

import "openagents/common/v1/common.proto";

option java_multiple_files = true;

enum AutonomyMode {
  AUTONOMY_MODE_UNSPECIFIED = 0;
  NORMAL = 1;
  DEGRADED = 2;           // tighter caps, higher tiers
  APPROVAL_REQUIRED = 3;  // human approvals for certain classes
  HALT = 4;               // stop new work in a domain/category
}

enum ActorRole {
  ACTOR_ROLE_UNSPECIFIED = 0;
  REQUESTER = 1;
  WORKER = 2;
  VERIFIER = 3;
  UNDERWRITER = 4;
  OPERATOR = 5;
}

message WindowSpec {
  // Rolling windows used for sv/xa calculations.
  bool use_5m = 1;
  bool use_1h = 2;
  bool use_24h = 3;
}

message TierRule {
  // Match conditions
  string category = 1; // empty => wildcard
  openagents.common.v1.FeedbackLatencyClass tfb_class = 2; // UNSPEC => wildcard
  openagents.common.v1.SeverityClass severity = 3;         // UNSPEC => wildcard

  // Required tier to unlock settlement/warranty.
  openagents.common.v1.VerificationTier required_tier = 10;

  // Independence constraints (heterogeneity/correlation).
  openagents.common.v1.IndependenceRequirement independence_requirement = 11;

  // Allow correlated Tier 1 checks to count? (usually only for LOW/INSTANT)
  bool allow_correlated_verification = 12;
}

message ProvenanceRule {
  string category = 1; // wildcard if empty
  openagents.common.v1.FeedbackLatencyClass tfb_class = 2;
  openagents.common.v1.SeverityClass severity = 3;

  openagents.common.v1.ProvenanceGrade min_grade = 10;
  bool require_provenance_bundle = 11;
  bool require_attestation = 12;
}

message EnvelopeRule {
  string category = 1;
  openagents.common.v1.SeverityClass severity = 2;

  // Envelope caps/expiries and fees.
  openagents.common.v1.Money max_amount = 10;
  uint64 ttl_ms = 11; // duration from issuance; authority receipts bind absolute expiry_ms
  uint32 fee_bps = 12;

  // Whether verdict is required for settlement (pay-after-verify default).
  bool require_verdict = 20;
}

message BondRule {
  string category = 1;
  openagents.common.v1.SeverityClass severity = 2;

  // Multipliers applied to contract price to size default bonds.
  // e.g. worker_self_bond = 0.10 means 10% of price.
  double worker_self_bond_fraction = 10;
  double underwriter_bond_fraction = 11;

  // Dispute bond sizing (anti-spam).
  double dispute_bond_fraction = 20;
}

message AutonomyCondition {
  // Match conditions
  string category = 1; // empty => wildcard

  // Thresholds (if unset, ignored).
  double min_sv = 10;                    // if sv falls below this
  double min_sv_effective = 11;          // correlation-adjusted trust floor
  double max_xa_hat = 12;                // if xa exceeds this
  double max_delta_m_hat = 13;           // if measurability gap too high
  double max_correlated_share = 14;      // if correlated verification share too high
  uint64 max_drift_alerts_24h = 15;      // drift detector pressure

  // Optional: enforce by severity/tfb slice (if set).
  openagents.common.v1.FeedbackLatencyClass tfb_class = 20;
  openagents.common.v1.SeverityClass severity = 21;
}

message AutonomyAction {
  oneof action {
    AutonomyMode set_mode = 1;

    // Tighten verification requirements dynamically.
    openagents.common.v1.VerificationTier raise_required_tier_to = 2;
    bool require_human_step = 3;

    // Tighten provenance requirements dynamically.
    openagents.common.v1.ProvenanceGrade raise_provenance_grade_to = 4;

    // Tighten spending dynamically.
    double envelope_max_amount_multiplier = 5; // e.g. 0.5 halves caps
    bool halt_new_envelopes = 6;

    // Warranty gating
    bool disable_warranties = 7;
    double warranty_coverage_cap_multiplier = 8;
  }
}

message AutonomyRule {
  AutonomyCondition condition = 1;
  repeated AutonomyAction actions = 2;

  // Hysteresis/cooldown to avoid flapping.
  uint64 min_duration_ms = 10; // condition must hold for this long
  uint64 cooldown_ms = 11;     // after triggering, ignore for this long
}

message AutonomyThrottlePolicy {
  WindowSpec windows = 1;

  // Default global thresholds.
  double global_min_sv = 10;
  double global_min_sv_effective = 11;
  double global_max_xa_hat = 12;
  double global_max_delta_m_hat = 13;
  double global_max_correlated_share = 14;
  uint64 global_max_drift_alerts_24h = 15;

  // Domain/category rules.
  repeated AutonomyRule rules = 20;
}

message AuthenticationRule {
  string rule_id = 1;
  string category = 2;
  openagents.common.v1.FeedbackLatencyClass tfb_class = 3;
  openagents.common.v1.SeverityClass severity = 4;
  ActorRole role = 5;

  openagents.common.v1.AuthAssuranceLevel min_auth_assurance = 10;
  bool require_personhood = 11;
  string trust_profile_id = 12; // issuer/revocation trust profile reference
  bool require_external_trust_proof_ref = 13; // when trust checks are out-of-kernel
}

message MonitoringRule {
  string rule_id = 1;
  string category = 2;
  openagents.common.v1.FeedbackLatencyClass tfb_class = 3;
  openagents.common.v1.SeverityClass severity = 4;

  repeated string required_detectors = 10;
  uint64 drift_alert_threshold_24h = 11;
  repeated AutonomyAction threshold_actions = 12; // deterministic action order applies
}

message RiskPricingRule {
  string rule_id = 1;
  string category = 2;
  openagents.common.v1.FeedbackLatencyClass tfb_class = 3;
  openagents.common.v1.SeverityClass severity = 4;

  uint32 base_liability_premium_bps = 10;
  double xa_multiplier = 11;
  double drift_multiplier = 12;
  double correlated_share_multiplier = 13;

  double min_capital_coverage_ratio = 20;
}

message CertificationRule {
  string rule_id = 1;
  string category = 2;
  openagents.common.v1.FeedbackLatencyClass tfb_class = 3;
  openagents.common.v1.SeverityClass severity = 4;

  bool require_certification = 10;
  repeated string accepted_certification_levels = 11;
  bool enable_safe_harbor_relaxations = 12;
}

message RollbackRule {
  string rule_id = 1;
  string category = 2;
  openagents.common.v1.FeedbackLatencyClass tfb_class = 3;
  openagents.common.v1.SeverityClass severity = 4;

  bool require_rollback_plan = 10;
  bool allow_compensating_action_only = 11;
}

message MarketCoverageRule {
  string rule_id = 1;
  string category = 2;
  openagents.common.v1.FeedbackLatencyClass tfb_class = 3;
  openagents.common.v1.SeverityClass severity = 4;

  bool require_coverage_binding = 10;
  openagents.common.v1.Money min_coverage_cap = 11;
  uint32 max_premium_bps = 12;
  uint32 min_distinct_underwriters = 13;
  double max_coverage_concentration_hhi = 14;
  double max_implied_fail_probability = 15;
  double min_market_calibration_24h = 16;
  repeated string disallowed_correlation_groups = 17;
  bool allow_safe_harbor_relaxation = 18;
}

message CostProofRule {
  string rule_id = 1;
  string category = 2;
  openagents.common.v1.FeedbackLatencyClass tfb_class = 3;
  openagents.common.v1.SeverityClass severity = 4;

  bool require_cost_proof_bundle = 10;
  openagents.common.v1.CostAttestationLevel min_cost_attestation_level = 11;
  double max_cost_variance_ratio = 12;
  double min_provider_cost_integrity_score = 13;
}

enum PolicyEvalFailureMode {
  POLICY_EVAL_FAILURE_MODE_UNSPECIFIED = 0;
  FAIL_CLOSED = 1;
  FAIL_BOUNDED_FALLBACK = 2;
}

message PolicyEvaluationGuardrails {
  uint32 max_rule_evaluations = 1;
  uint64 max_eval_time_ms = 2;
  PolicyEvalFailureMode failure_mode = 3;
  string limit_exceeded_reason_code = 4; // e.g. POLICY_EVAL_LIMIT_EXCEEDED
}

message PolicyBundle {
  string policy_bundle_id = 1;
  string name = 2;
  string description = 3;

  int64 created_at_ms = 10;
  string version = 11;

  // Default static requirements
  repeated TierRule tier_rules = 20;
  repeated ProvenanceRule provenance_rules = 21;

  // Capital rules
  repeated EnvelopeRule envelope_rules = 30;
  repeated BondRule bond_rules = 31;

  // Identity / monitoring / pricing / certification controls
  repeated AuthenticationRule authentication_rules = 32;
  repeated MonitoringRule monitoring_rules = 33;
  repeated RiskPricingRule risk_pricing_rules = 34;
  repeated CertificationRule certification_rules = 35;
  repeated RollbackRule rollback_rules = 36;
  repeated MarketCoverageRule market_coverage_rules = 37;
  repeated CostProofRule cost_proof_rules = 38;

  // Dynamic gating
  AutonomyThrottlePolicy autonomy = 40;
  PolicyEvaluationGuardrails evaluation_guardrails = 41;

  // Non-normative metadata
  map<string, string> tags = 50;
}

2.3 Policy semantics (how services must apply it)

This is the “expectations” part: how the kernel must interpret the policy bundle.

A) Matching rules

Every rule uses (category, tfb_class, severity) matching with wildcards. Matching precedence MUST be deterministic:

  1. exact category + exact tfb + exact severity
  2. exact category + exact severity + wildcard tfb
  3. exact category + wildcard severity + wildcard tfb
  4. wildcard category + exact tfb/severity
  5. wildcard everything (global default)

If multiple rules match at the same precedence, the kernel MUST choose deterministically (e.g., lexicographic by rule id) and include that rule id in hash-bound receipt evidence and/or hash-bound receipt payload (not only non-normative tags).

Idempotency scope alignment requirement:

  • idempotency lookup MUST be scoped by canonicalized authority action identity (stable RPC full-name or normalized HTTP method+route) plus authenticated caller identity.
  • aliases for the same action MUST normalize to the same scope key before lookup.

B) Verification gating

When creating a Contract or starting Verification, Aegis MUST:

  • compute the required tier and independence requirement from tier_rules
  • reject (or downgrade) any plan that violates policy
  • record the required tier + independence requirement in the contract terms hash

When finalizing a verdict, Aegis MUST:

  • compute whether the IndependenceReport violates policy
  • if it violates, the verdict may still be recorded, but settlement MUST be WITHHELD unless policy explicitly allows it
  • emit receipts with hints.verification_correlated and hints.achieved_verification_tier

C) Provenance gating

When accepting submission, Aegis MUST:

  • require provenance_bundle evidence if required by the policy or plan
  • compute Pgrade deterministically
  • if Pgrade is below min grade for the WorkUnit’s slice, settlement and/or warranty issuance MUST be withheld or denied per policy.

D) Autonomy throttle (paper implementation)

The autonomy throttle is the runtime bridge from the paper’s variables to behavior.

  • The EconomySnapshot job computes sv, sv_effective, xa_hat, delta_m_hat, correlated share, and drift alert pressure per slice.

  • TreasuryRouter and/or the kernel services MUST read the latest snapshot.

  • When conditions match, actions apply in a deterministic order:

    1. mode changes (NORMAL → DEGRADED → APPROVAL_REQUIRED → HALT)
    2. raising verification tier / requiring human step
    3. raising provenance grade
    4. tightening envelope limits / halting envelopes
    5. disabling warranties or capping coverage

All triggered actions MUST be:

  • recorded in a breaker/throttle state object
  • included in receipts as policy notes
  • bound to the exact snapshot_id/snapshot_hash
  • surfaced in /stats (“autonomy mode” and breaker states)

E) Envelope enforcement

Hydra MUST enforce envelope rules using:

  • WorkUnit category/severity from trace links (WorkUnit referenced by Contract)
  • current autonomy mode and throttle actions
  • latest EconomySnapshot values

If policy says “halt_new_envelopes” for a category, Hydra MUST respond WITHHELD and emit a receipt with the exact reason and snapshot id.

F) Authentication/personhood enforcement

For actions with an ActorRole and category slice:

  • the selected AuthenticationRule MUST be deterministic using the same precedence rules.
  • kernel services MUST compare ReceiptHints.auth_assurance_level and personhood_proved against policy.
  • if insufficient, action MUST be denied/withheld with stable reason code AUTH_ASSURANCE_INSUFFICIENT, plus hash-bound policy decision evidence.

G) Monitoring/drift enforcement

When drift signals breach MonitoringRule.drift_alert_threshold_24h:

  • services MUST emit deterministic drift receipts:
    • economy.drift.signal_emitted.v1
    • economy.drift.alert_raised.v1
    • economy.drift.false_positive_confirmed.v1
  • idempotency MUST be keyed by (detector_id + snapshot window + receipt_type),
  • then apply threshold_actions in the same deterministic autonomy action order,
  • and include triggering detector ids in hash-bound receipt evidence.

H) Risk pricing enforcement

For contract creation/update:

  • RiskPricingRule MUST compute liability premium deterministically from snapshot values and configured multipliers.
  • contract and settlement receipts MUST separate execution_price, verification_fee, and liability_premium.
  • receipts MUST include pricing_snapshot_ref so underwriters can reproduce pricing inputs.

I) Certification and rollback enforcement

For high-severity or irreversible slices:

  • CertificationRule.require_certification=true MUST block uncertified actions (WITHHELD / border-block reason code).
  • RollbackRule.require_rollback_plan=true MUST require rollback or compensating-action plan references before authority mutation.
  • missing rollback/compensating plans MUST withhold with stable reason code ROLLBACK_PLAN_REQUIRED and policy decision evidence.
  • rollback action receipts MUST use stable types:
    • economy.rollback.executed.v1
    • economy.rollback.failed.v1
    • economy.compensating_action.executed.v1
  • safe-harbor relaxations are allowed only when certification validity + observed risk metrics satisfy policy.

J) Markets and coverage enforcement

Where coverage markets are enabled:

  • MarketCoverageRule.require_coverage_binding=true MUST block warranty-bearing actions unless a valid CoverageBinding exists.
  • coverage relaxations MUST require policy checks for concentration/diversity/calibration/implied-fail-probability/correlation constraints.
  • any verification/provenance relaxation due to market-backed coverage MUST be:
    • receipted as a policy action,
    • bound to snapshot_id/snapshot_hash,
    • explainable via hash-bound policy evidence.

K) Cost-proof enforcement

For compute-like slices:

  • CostProofRule.require_cost_proof_bundle=true MUST require Submission.cost_proof_bundle.
  • min_cost_attestation_level MUST be enforced deterministically by slice.
  • cost anomalies above policy thresholds MUST emit explicit anomaly/drift receipts (for example economy.cost.anomaly_detected.v1) and feed the same deterministic throttle order.
  • min_provider_cost_integrity_score (if set) MAY be used only as a policy prior; it MUST NOT bypass resolution-path requirements.

L) Verifier assignment, payout linkage, and liability posture

Verification services MUST enforce assignment semantics deterministically:

  • VerificationAssignment.mode defaults to POLICY_SELECTED_SET when unspecified.
  • FinalizeVerdict MUST require verification_assignment.assignment_ref (or equivalent hash-bound linkage). Missing linkage MUST fail with stable reason VERIFICATION_ASSIGNMENT_MISSING.
  • When verifier work is compensated, payout receipts MUST link to both assignment and verdict receipts and MUST debit the explicit verification_fee lane.
  • Verifier economic liability is opt-in policy behavior:
    • default posture: non-liable economically, reputationally accountable,
    • if VerifierLiabilityPolicy.economic_liability_enabled=true, verifier-bond/slash receipts MUST be explicit and slash conditions MUST be hash-bound via slash_policy_ref.
  • Worker/verifier discovery and matching are out-of-kernel mutation scope unless an optional market module is enabled; kernel authority semantics remain post-match and receipt-driven.

M) Evidence immutability and sizing semantics

  • EvidenceRef.digest is authoritative identity; EvidenceRef.uri is transport only.
  • If fetched bytes do not match digest, the evidence MUST be rejected for authority decisions and replay.
  • Evidence supersession MUST be append-only (new digest + new receipt linkage), never in-place mutation.
  • Large evidence payloads SHOULD be externalized and referenced by digest/URI; receipts SHOULD carry compact summaries plus references, not bulk opaque payload bodies.

N) Long-latency truth resolution and simulation governance

  • Delayed/external truth inputs MUST be represented as GroundTruthEvidence with source_id, digest-bound evidence ref, received_at_ms, confidence posture, and contract/verdict/claim receipt linkage.
  • Claim/dispute finalization MUST be receipted and may resolve only via objective harness, declared adjudication policy, or explicit human underwriting.
  • External updates MUST NOT silently flip truth state; they must flow through receipted incident/update and claim-resolution transitions.
  • GroundTruthCase -> SimulationScenario promotion MUST preserve derivation governance:
    • receipted derivation linkage (source incident digest(s), generator identity, timestamp),
    • receipted validation linkage (validator identity + posture),
    • export bundles carrying derivation/validation refs for replay.

O) Identity trust-profile ownership semantics

  • AuthenticationRule.trust_profile_id indicates kernel-owned issuer/revocation verification policy for the slice/role.
  • If trust verification is external, require_external_trust_proof_ref=true requires hash-bound proof references in authority receipts.
  • Missing required trust-profile linkage or missing external trust proof references MUST deterministically deny/withhold with stable reason codes.

P) Policy complexity guardrails and deterministic fallback

  • PolicyEvaluationGuardrails MUST bound evaluation complexity (max_rule_evaluations, max_eval_time_ms).
  • Guardrail breaches MUST produce deterministic failure behavior according to failure_mode with stable reason code (limit_exceeded_reason_code).
  • Authority actions MUST default to fail-closed unless policy explicitly configures a bounded fallback and that fallback remains receipted/auditable.

Q) Reputation and outcome-export boundary semantics

  • Reputation formula/decay/weighting are policy-defined and non-universal; interoperable consumers must not assume one global trust function.
  • Any authority decision materially depending on reputation MUST include measurement-window and metric references in hash-bound receipt linkage.
  • Outcome registry public exports remain aggregate-safe; restricted exports carry linkage needed for independent audit replay.

The remaining sections introduce optional compute market primitives while staying consistent with the kernel invariants and your existing packages.

These additions do not modify the previous files. They add a new package and a few cross-references so the system composes cleanly.

This lets the Economy Kernel support:

  • spot compute markets
  • forward capacity sales
  • GPU futures
  • compute options
  • delivery proofs
  • compute indices
  • hedging instruments

while remaining receipt-native and policy-bounded.

3. Optional compute market proto additions

New package:

proto/
  openagents/
    compute/v1/
      compute_products.proto
      compute_capacity.proto
      compute_instruments.proto
      compute_indices.proto
      compute_delivery.proto

These packages implement the Section 8 compute commodity market spec.

They integrate with:

  • common/v1
  • policy/v1
  • economy/v1
  • aegis/outcomes/v1

3.1 Compute product definitions

proto/openagents/compute/v1/compute_products.proto

This defines standardized compute commodities.

syntax = "proto3";

package openagents.compute.v1;

import "openagents/common/v1/common.proto";

option java_multiple_files = true;

enum ResourceClass {
  RESOURCE_CLASS_UNSPECIFIED = 0;

  GPU_A100 = 1;
  GPU_H100 = 2;
  GPU_B200 = 3;

  GPU_GENERIC_HIGHMEM = 10;
  GPU_GENERIC_STANDARD = 11;

  CPU_STANDARD = 20;
  CPU_HIGHMEM = 21;

  STORAGE_IO = 30;
}

enum SettlementMode {
  SETTLEMENT_MODE_UNSPECIFIED = 0;
  PHYSICAL_DELIVERY = 1;
  CASH_SETTLED = 2;
  BUYER_ELECTION = 3;
}

message ComputeProduct {

  string product_id = 1;

  ResourceClass resource_class = 2;

  string capacity_unit = 3;
  // e.g. "gpu_hour", "gpu_second", "gpu_month"

  string region_spec = 4;
  // e.g. "any", "us-east", "europe"

  string performance_band = 5;
  // e.g. memory / bandwidth / accelerator spec band

  openagents.common.v1.CostAttestationLevel min_cost_attestation = 6;

  bool index_eligible = 7;

  SettlementMode settlement_mode = 8;

  int64 version = 20;

  map<string,string> tags = 30;
}

This allows the kernel to standardize products like:

GPU_H100_US_GPU_HOUR
GPU_GENERIC_HIGHMEM_GLOBAL_GPU_SECOND

which become the base layer for futures and indices.

3.2 Capacity lots (physical inventory)

proto/openagents/compute/v1/compute_capacity.proto

Represents actual compute inventory offered into the market.

syntax = "proto3";

package openagents.compute.v1;

import "openagents/common/v1/common.proto";

option java_multiple_files = true;

enum CapacityLotState {
  CAPACITY_LOT_STATE_UNSPECIFIED = 0;
  OFFERED = 1;
  RESERVED = 2;
  ASSIGNED = 3;
  DELIVERING = 4;
  COMPLETED = 5;
  CANCELLED = 6;
}

message CapacityLot {

  string lot_id = 1;

  string provider_id = 2;

  string product_id = 3;

  uint64 quantity = 4;

  int64 delivery_start_ms = 5;
  int64 delivery_end_ms = 6;

  openagents.common.v1.Money min_price = 7;

  string region_override = 8;

  openagents.common.v1.CostAttestationLevel attestation_level = 9;

  CapacityLotState state = 10;

  map<string,string> tags = 20;
}

Example:

lot_id: "lot_h100_cluster_0123"
product_id: "GPU_H100_US_GPU_HOUR"
quantity: 10,000
delivery_start: 2026-06-01
delivery_end: 2026-06-30

These lots back physical settlement.

3.3 Compute financial instruments

proto/openagents/compute/v1/compute_instruments.proto

Defines spot, forwards, futures, options.

syntax = "proto3";

package openagents.compute.v1;

import "openagents/common/v1/common.proto";

option java_multiple_files = true;

enum InstrumentType {

  INSTRUMENT_TYPE_UNSPECIFIED = 0;

  SPOT = 1;

  FORWARD_PHYSICAL = 2;

  FUTURE_CASH = 3;

  FUTURE_PHYSICAL = 4;

  CALL_OPTION_RESERVATION = 5;

  PUT_OPTION_RELEASE = 6;

  SWAP_FIXED_FLOATING = 7;

  STRIP = 8;
}

enum InstrumentState {

  INSTRUMENT_STATE_UNSPECIFIED = 0;

  CREATED = 1;

  OPEN = 2;

  FILLED = 3;

  ACTIVE = 4;

  DELIVERING = 5;

  CASH_SETTLING = 6;

  SETTLED = 7;

  DEFAULTED = 8;

  EXPIRED = 9;

  CANCELLED = 10;
}

message CapacityInstrument {

  string instrument_id = 1;

  InstrumentType type = 2;

  string product_id = 3;

  uint64 quantity = 4;

  openagents.common.v1.Money strike_price = 5;

  int64 delivery_start_ms = 6;

  int64 delivery_end_ms = 7;

  SettlementMode settlement_mode = 8;

  string buyer_id = 9;

  string seller_id = 10;

  InstrumentState state = 11;

  string compute_index_id = 12;

  string capacity_lot_id = 13;

  map<string,string> tags = 20;
}

Examples:

type: FUTURE_CASH
product: GPU_H100_US_GPU_HOUR
delivery_month: 2027-01

or

type: CALL_OPTION_RESERVATION
strike: \$2.50/gpu_hour

3.4 Compute indices

proto/openagents/compute/v1/compute_indices.proto

Used for cash settlement and /stats pricing.

syntax = "proto3";

package openagents.compute.v1;

import "openagents/common/v1/common.proto";

option java_multiple_files = true;

message ComputeIndex {

  string index_id = 1;

  string product_id = 2;

  int64 observation_start_ms = 3;

  int64 observation_end_ms = 4;

  double price_p50 = 5;

  double price_p95 = 6;

  uint64 observation_count = 7;

  double quality_score = 8;

  string methodology_version = 9;

  string index_hash = 10;

  map<string,string> tags = 20;
}

This index powers:

cash-settled GPU futures
swap floating legs
market pricing signals

3.5 Delivery proofs

proto/openagents/compute/v1/compute_delivery.proto

Physical settlement proof.

syntax = "proto3";

package openagents.compute.v1;

import "openagents/common/v1/common.proto";

option java_multiple_files = true;

message DeliveryProof {

  string delivery_id = 1;

  string instrument_id = 2;

  string capacity_lot_id = 3;

  uint64 delivered_quantity = 4;

  int64 delivery_start_ms = 5;

  int64 delivery_end_ms = 6;

  openagents.common.v1.EvidenceRef cost_proof_bundle = 7;

  openagents.common.v1.EvidenceRef runtime_attestation = 8;

  bool accepted = 9;

  openagents.common.v1.TraceContext trace = 10;
}

This links:

CapacityInstrument → CapacityLot → DeliveryProof → CostProofBundle

making delivery verifiable and auditable.

4. Cross-package integrations

4.1 WorkUnits referencing compute markets

outcomes_work.proto can reference compute instruments:

string compute_instrument_id = N;

This allows:

hedged compute → execution job

4.2 Policy rules extended for compute markets

Add fields to PolicyBundle:

message ComputeMarketRule {

  string product_id = 1;

  uint64 max_open_interest = 2;

  double max_paper_to_physical_ratio = 3;

  double min_deliverable_coverage_ratio = 4;

  bool allow_cash_settlement = 5;

  bool allow_options = 6;

}

4.3 Economy snapshot additions

Add compute market metrics:

double compute_open_interest = X;
double paper_to_physical_ratio = Y;
double deliverable_coverage_ratio = Z;
double hedged_compute_share = W;

5. Resulting economic stack

With these additions the kernel supports:

Base layer

WorkUnits Contracts Verification Settlement

Risk layer

Bonds Warranties Claims

Market layer

Coverage markets Prediction signals

Compute commodity layer

Spot compute Forward capacity GPU futures Options on compute Delivery proofs Compute price indices

6. Why this architecture is powerful

You now have one unified substrate for:

  1. AI labor markets
agent work
verification
liability
  1. Compute markets
GPU supply
capacity futures
price discovery
  1. Risk markets
coverage
prediction
insurance

All share:

policy bundles
receipts
identity assurance
provenance
evidence

So the system becomes something very unusual:

a general economic kernel for machine work

not just a marketplace.