Eve library

RUN-004 evidence pack — common-model runtime adapters

Evidence

Revision 1

On this page

Work-package charter#

  • Task ID: RUN-004
  • Owner: Oshun common run-protocol adapter maintainer; implementation owner: Codex
  • Accountable product: Shared V1–V9 autonomous-content platform
  • Priority: P0
  • Target milestone: 2026-07-15 SOTA gap-closure program
  • Risk tier: Critical semantic-drift, authority-widening, false-success, evidence-loss, and premature-envelope-deprecation risk
  • Dependencies: RUN-001 common aggregate, RUN-002 durable record semantics, RUN-003 effect boundary, V1 AgentRun and creative-autonomy records, V8 StageJournal, V9 Metis orchestration plan, and V6 cognition envelope
  • Affected schemas and runtime: Portable adapter conformance record and snapshot builder; V1, V8, V9, and V6 source-specific adapters; package exports, dependency manifests, focused CI gate, and conformance tests
  • Rollout strategy: Keep every product envelope authoritative. Produce a digest-bound common snapshot plus an accounting record that maps every declared source field exactly once or names it as unsupported.
  • Rollback strategy: Revert the common conformance contract, all four source adapters, tests, exports, dependency links, CI expansion, audit qualification, and ledger closure together. No product store or migration consumes the adapters yet, so there is no product-data rollback.
  • Evidence-pack location: This file

Shared conformance contract#

RunProtocolAdapterRecord binds one source identity and canonical source SHA-256 to one validated RUN-001 snapshot. Its field inventory must cover every top-level source field. Every inventory path must then appear exactly once in a semantic mapping or an explicit unsupported-field record. Nested field paths are supported for source subobjects such as V1 token accounting.

Exact mappings cannot name a normalizer. Normalized and aggregated mappings must name one. The record builder independently hashes the source and target semantic projections and refuses publication when the hashes differ. Every mapped target path must exist in the common snapshot. Unsupported fields carry a reason, preservation mode, and follow-up task.

Every adapter snapshot contains a designated audit artifact whose content hash equals the complete canonical source digest. The conformance digest covers the adapter/schema identity, complete field accounting, semantic mappings, unsupported declarations, and common run. Schema validation rejects a changed run, source audit artifact, field accounting, semantic digest, or conformance digest.

The common snapshot builder derives immutable RUN-001 identity from the source identity and source digest boundary, constructs legal run and step transition paths, and validates the result with RunProtocolV1Schema. It does not invent attempts, tool grants, side-effect receipts, or product-store events.

V1 AgentRun and creative-autonomy mapping#

V1 intentionally uses three adapters under one conformance ID because the canonical AgentRun, creative-director cycle, and review item are independent source identities. The adapter never fuses an unrelated cycle or review record into an AgentRun.

  • adaptV1AgentRunToCommonRun validates the canonical AgentRun schema and top-level field inventory. It maps tenant/actor/purpose/risk identity, status, plan DAG, intermediate artifacts, evidence verdict, USD-micros and token measurements, output release lifecycle, and captured workflow/model/prompt/ policy bindings.
  • Raw inputs, tool calls, structured error, decision rationale, snapshot-only start/completion timestamps, parent/replay/successor UUIDs, workspace, tags, and free-form metadata remain explicit unsupported fields. In particular, legacy tool calls are not relabeled as common tool events because AgentRun lacks authoritative resolved-grant and reconciliation records.
  • adaptV1CreativeCycleToCommonRun maps finished/aborted status, cycle time, token usage, production/publication/selection/governor/pruning audit artifacts, and complete source identity. Studio spend units remain unsupported because no authoritative USD conversion exists; LLM call count remains unsupported because the common measurement union has no counter.
  • adaptV1CreativeReviewToCommonRun maps autonomy and lifecycle status, compellingness gate verdict and full gate audit, human/system review actions as interventions, creative provenance binding, variant artifact, lifecycle metadata audit, and publication release. It preserves the full review history and tweak directive digest.

V8 StageJournal and case-gate mapping#

adaptV8StageJournalToCommonRun requires one case/seed journal with contiguous sequence from one, monotonic valid timestamps, non-negative integer attempts and costs, known event/key/cost fields, and valid lowercase output hashes. It maps the deterministic seed binding, latest state for every observed stage, stage output hashes as common artifacts, token measurements, and run/step status without treating a partial pipeline as success.

The common run succeeds only when publish succeeded and every latest stage is succeeded, repaired, or skipped. Latest verify, eval, and hitl states become typed gate verdicts with their output artifact or journal evidence; failed gates block and escalated gates remain pending. Free-form detail, the complete attempt timeline, and job-count costs remain explicitly unsupported pending product-store event consumption and authoritative cost conversion.

V9 dependency, governance, and agent-plan mapping#

adaptV9AgentPlanToCommonRun is locked to the checked-in Python service OpenAPI AgentOrchestrationPlanResponse property inventory. It validates source hash/time identity, unique selected agents, exact selected-agent/DAG agreement, positive node stages, unique node/edge/evidence/gate IDs, known references, complete dependency-edge pairs, and an acyclic graph.

The adapter maps the Metis build/tenant/actor/objective identity, planning or execution status, DAG nodes and dependencies, evidence-ledger artifacts, pending approval gates, and registry/plan/source/agent version bindings. It normalizes arbitrary storage locators into stable common references while retaining content hashes. Event-topic transport configuration, armed kill-switch declarations, and the replay manifest remain explicit unsupported fields until the product runtime emits executed events, interventions, and durable checkpoints under RUN-005.

V6 cognition events and child authority#

adaptV6CognitionRunToCommonRun validates the AgentRunEnvelope field and usage inventories, non-negative usage counters and costs, unique allowed tools, bounded and monotonic event times, and outcome timestamps. Every cognition event becomes a content-hashed common audit artifact; final text becomes an output artifact; provider/model and allowed-tool authority become versioned bindings; tokens, run cost, and reserved child cost become measurements.

A child requires an explicit source-parent/common-parent mapping. Its allowed tools must be a subset of the parent's tools before adaptation. The common parent identity, delegated actor evidence, and authority binding are retained; a missing mapping or widened child authority fails loud. Executing, completed, killed, quarantined/budget-exceeded, and truncated/failed/max-iteration states normalize without inventing success. LLM/tool call counts remain explicit unsupported counter fields.

Automated scenario matrix#

Scenario Conformance result
Complete field accounting Hidden, missing, duplicate, or unaccounted source fields fail.
Semantic equivalence Source and target semantic digests must match before record creation.
Target drift A mapping to a nonexistent common path fails.
Source preservation Every adapter retains the full source SHA-256 in its designated audit artifact.
Tamper resistance Changed common data, source audit identity, or conformance digest fails validation.
V1 independence AgentRun, creative cycle, and review lifecycle receive separate common identities.
V1 review Gate, human approval, provenance, lifecycle audit, and publication release survive normalization.
V8 gate evidence Latest verify/eval/HITL outcome becomes a typed verdict linked to hash/journal evidence.
V8 false success Publish plus complete healthy latest-stage state is required for success.
V9 schema drift OpenAPI response properties must equal the adapter inventory.
V9 dependency integrity Unknown, unmatched, duplicate, or cyclic DAG relationships fail.
V6 child authority Missing parent mapping or tool widening fails before a child common run is emitted.
Outcome normalization Kill, quarantine, failure, timeout, and truncation never become success.
Determinism Repeated source input yields the same record; semantic source change changes its digests.

Artifact identity#

Artifact SHA-256
libs/contracts/src/run-protocol/adapter-conformance.ts 3bb90afb91ce2a38c7e7430c056bb6bc68bc4162b24106f37b0327b86ba0920c
libs/contracts/src/run-protocol/adapter-conformance.spec.ts 476b5264ede6798b0f76a3f88da2128f162fa648e3252380fbfd54fb8ac2d57b
libs/oshun/creative-autonomy/src/common-run-adapter.ts 96a2b5667de6773760e00a2646d05c8080cff866bd3530ed93390f6d951149c2
libs/oshun/creative-autonomy/src/common-run-adapter.test.ts 57c24dc57fe1ae20d832dbe187fe3cd96625f4c81a494b94b6c51946b16a6fda
libs/yemaya/case-pipeline/src/common-run-adapter.ts 3cf752b6255fc244a4954a01f8c68350fb8e03dcba5b3dfb457357023ce79bb1
libs/yemaya/case-pipeline/src/common-run-adapter.test.ts 5a321dd25cd4381582494cf3db7f1cf35e4fae5b3bd388107490aea8b00b7a3a
libs/metis/api-client/src/common-run-adapter.ts 7dc421403efaf011ef753b05afb58b23162e02adb1fd6e17fba24243e55768bc
libs/metis/api-client/src/common-run-adapter.test.ts dc3d6b25886a166b1b2d397964ffd6977cdb463ca59840bc7b6229832aa617ec
libs/iris/agents/core/src/agentic/common-run-adapter.ts 0cb98c262332bb4b63fa84d8e0f30abc66927d90cd9bdc48c0fbeddb8451f4a4
libs/iris/agents/core/src/agentic/common-run-adapter.test.ts 517aa3621073a3522f91fc8d1d6c70a66418406c76894b075b9e4d20bc04e4fd
package.json e7d9bf89abf451ac36f88586fc9fff85a578ec3783be856358a4cd60e020898d
pnpm-lock.yaml 627422a4715c85693b8fccf921ee7c615f0f002911f8506758769e5f6bbbc04e

Automated and adversarial verification#

  • pnpm run verify:contracts:run-protocol passes targeted ESLint with zero warnings, 49/49 common protocol cases with one worker, and contracts source typechecking. RUN-004 contributes four shared conformance cases.
  • Focused adapter suites pass 19/19 cases: V1 5/5, V8 4/4, V9 5/5, and V6 5/5. They cover deterministic mapping, source/target digests, unsupported accounting, source and OpenAPI drift, invalid hashes, broken/cyclic DAGs, false-success outcomes, token dishonesty, time regression, parent lineage, and child authority widening.
  • Full affected-package regressions pass sequentially with one worker: V1 creative autonomy 253/253, V8 case pipeline 26/26, V9 Metis API client 738/738, and V6 Iris agents core 206/206 (1,223/1,223 total).
  • Production TypeScript checks pass for contracts, V1 creative autonomy, V8 case pipeline, V9 API client, and V6 agents core. The V9 package-wide spec typecheck still reports pre-existing Vitest mock-generic errors in unrelated auth/http/cache/error tests; the new adapter source typecheck and full 738- case runtime suite pass.
  • Changed production files and the common protocol gate pass ESLint with zero warnings. Package-wide lint discovery exposes existing unrelated debt: seven V8 non-null warnings; two V6 non-null warnings and two existing module- boundary errors; three V9 non-null warnings; and 87 V1 non-null warnings. No changed adapter file contributes an error or warning.
  • NX_DAEMON=false pnpm exec nx run @oshun/contracts:build --parallel=1 --outputStyle=static succeeds, and runtime probes confirm the adapter schema through the root flat export, RunProtocolContracts namespace, and @oshun/contracts/run-protocol subpath.
  • pnpm install --frozen-lockfile --lockfile-only --offline --ignore-scripts passes after the three local workspace dependency links were recorded.
  • The broader gates ran sequentially after the required resource check. The host had 15 GiB RAM, 31 GiB swap, about 10 GiB available RAM, and 38 GiB free disk. No broad build matrix or parallel package verification was run.

Deprecation boundary and negative proof#

  • No existing V1, V8, V9, or V6 envelope is deprecated. The adapters are one-way, in-memory conformance functions; product records remain the source of truth.
  • Several adapters deliberately report unsupported fields. No product store writes a common record, no migration or dual-write has run, no product replay has compared a migrated record with the source runtime, and no rollback has restored a product from common state. The conditional deprecation gate is therefore closed.
  • Tests and artifact hashes are synthetic repository evidence, not a deployed run, production migration, live provider receipt, learner interaction, case release, or operational rollback drill.
  • RUN-005 owns product consumption, atomic approval/execution lifecycle, and replay/rollback integration. Envelope deprecation remains prohibited until migration, replay, and rollback evidence passes with no unsupported semantic loss.
  • No user-facing UI, mobile flow, media/model pipeline, database migration, or Unreal asset changed. Playwright/mobile/browser accessibility, provider/media/database, and Unreal verification are inapplicable.

Publication identity#

  • Implementation and first remote-integrated tested source commit: 0f68a5e5a6a72cef4f32fbc3e96db8e732c5c00c
  • Publication metadata commit: a70125182c315d6560c9a9fddefa2bb364476f17