Eve library

V7-001 Mawu and Forge executable-boundary closure

Evidence

Revision 1

On this page

Work-package charter#

  • Task ID: V7-001
  • Owner: V7 Mawu, Forge Assist, Moremi realm-runtime, and shared authorization maintainers; implementation owner: Elikem Agudogo
  • Accountable product: V7 Mawu/Moremi creator and realm platform
  • Priority: P0
  • Target milestone: 2026-07-15 SOTA gap-closure program
  • Dependencies: PROMPT-002/PROMPT-003 trust zoning, AUTH-001 capability grants, SAFETY-001 static/invocation safety, Wasmtime 45, the Mawu semantic grounding contract, and the authenticated internal Moremi boundary
  • Affected schemas and content slots: Mawu output-grounding and provenance, Forge restricted-language/admission receipts, AUTH-bound invocation records, and Moremi Wasmtime host-call traces. No production realm, player-visible content slot, or cooked artifact changed.
  • Scope closed here: semantic source-to-output grounding, honest procedural-backend labeling, per-refinement and invocation-time Forge policy checks, AUTH-001 limit binding, and complete execution of the supported restricted Forge program
  • Scope deliberately excluded: arbitrary JavaScript is not an accepted Forge language and is rejected rather than partially projected
  • Risk tier: Critical creator-code execution, delegated authority, source faithfulness, and provenance
  • Source commits: 18b6b990a5 (Forge Wasmtime admission), 2af8c9f1d6 (semantic grounding), b8f637a8d7 (AUTH-bound invocation), and 16d3588f78 (complete restricted-program execution)
  • Rollout strategy: Keep the procedural and semantic-reference backends explicitly labeled, fail closed without an injected production backend, and admit Forge programs only through the complete restricted grammar, exact AUTH-001 bindings, and Wasmtime host-call trace equality.
  • Rollback strategy: Revert this evidence commit to reopen the five closed checklist bullets. Reverting source requires reverting the applicable source commit; no persistence migration is involved.
  • Evidence-pack location: This file

Mawu semantic grounding is output-bound#

The Mawu Studio generation route no longer treats the presence of Sophia source metadata as proof that an output is grounded. After generation, an injected MawuStudioSemanticGroundingVerifier must describe the factual claims made by the actual generated descriptor and return claim/citation assessments. The route builds the portfolio semantic proof bundle over:

  1. retrieval proofs for the exact approved Sophia passages;
  2. passage-match proofs binding citations to those passages; and
  3. claim-entailment proofs binding generated-output claims to cited passages.

The canonical evaluateSemanticGroundingProofBundle gate must pass before an artifact is emitted. Missing verifier configuration returns semantic-verifier-not-configured; unsupported output claims or invalid proof coverage return semantic-grounding-failed. Both outcomes return a null artifact. The owning regression also injects an unsupported “made of gold” claim and proves it cannot become an artifact.

This is semantic source-to-output evidence, not source-list decoration. The proof bundle is preserved on the successful artifact and its summary reports the entailed-claim numerator and denominator.

The reference backend is explicitly procedural#

The local deterministic backend identifies itself as isis.mawu-studio.reference-procedural-backend.v1. That exact identifier is written into artifact provenance and asserted by the owning generation-policy eval. The route never silently substitutes it for an advertised model backend: when no backend is injected, the result is backend-not-configured and no artifact is fabricated.

The reference semantic verifier is likewise labeled isis.mawu-studio.reference-semantic-verifier.v1. It exists for deterministic evaluation of the procedural descriptor; it is not described as a production model or independent human assessment.

Forge policy is rechecked across generation and invocation#

Forge Assist applies its static capability policy inside generateProposal. The initial attempt and every quality refinement therefore traverse the same policy, AUTH-001, and sandbox-admission sequence. The regression records the source hash authorized for a low-quality attempt and its refined replacement and proves both distinct hashes received independent authorization.

The Moremi admission service then:

  • accepts only constant string bindings and direct top-level oshun.invoke(capability, strictJsonPayload); statements, rejecting every unsupported statement before instantiation;
  • resolves literal, concatenated, bound, and template-computed capabilities; unresolved runtime-dynamic, aliased, optional, indexed, or hidden access fails closed;
  • independently enforces realm-only and declared-capability policy over the resolved runtime call set;
  • parses each payload as strict JSON, recursively rejects credential- and secret-shaped fields, canonicalizes it, and applies a 16 KiB per-call limit;
  • compiles every accepted statement, in order, into a real core Wasmtime 45 module whose data segments contain the exact canonical payload bytes and whose run export invokes only declared, policy-clean, authority-bound imports;
  • rechecks expiry and wall time inside each actual host callback and records the corresponding AUTH-001 use ID after reading and revalidating the payload from guest memory; and
  • requires the host-call, authority-use, and SHA-256 payload traces to match the compiled program.

The HTTP boundary requires configured bearer authentication and binds the v3 receipt to artifact ID, source SHA-256, declarations, program-module SHA-256, host calls, AUTH-001 uses, and payload hashes. Forge Assist refuses substituted use IDs, malformed hashes, missing module identity, or trace-length mismatches.

AUTH-001 is bound into the executable request#

Auth001ForgeInvocationAuthority is a narrow adapter over the canonical PortfolioCapabilityBroker; the broker structurally satisfies the adapter port without creating a reverse package dependency from Maya into Agentic Studio. For every declared capability, it resolves:

  • exact agent, actor, tenant, user, run, pipeline/session, tool, and purpose;
  • the exact forge-capability:<capability> resource and invoke action;
  • source and target trust zones plus risk ceiling;
  • grant liveness and scope specificity; and
  • the grant's nested forge fuel, memory, wall-time, and process limits.

The adapter requests no filesystem path, executable, child process, network, or secret. It emits no credential or bearer token. Its one-shot binding contains only the grant/use identity, artifact and source hashes, capability/resource/ action, timestamps, explicit zero-egress/zero-secret constraints, and resource limits. The binding lifetime is capped at 15 seconds even when the parent grant lives longer.

Admission schema v7.moremi.forge-assist-admission.3 rejects missing, duplicate, expired, overlong, source-swapped, undeclared, egress-enabled, secret-bearing, or nonzero-process bindings before instantiation. Moremi applies the narrowest bound fuel and memory limits to its Store, keeps process authority at zero, and checks binding expiry and elapsed wall time in the host callback. Forge Assist records succeeded, denied, and failed outcomes back through AUTH-001.

The binding travels only over the bearer-authenticated internal Moremi boundary. It is not presented as an independently signed public token; exposure of the internal service API key remains an infrastructure credential incident, not a supported caller path.

Supported-language completeness and dynamic-call handling#

Forge no longer presents arbitrary JavaScript-like text as executable. The system prompt names the restricted grammar, positive fixtures use it, and Moremi's parser consumes the entire source modulo whitespace and comments. Assignments from invocations, functions, handlers, control flow, trailing statements, object shorthand, unquoted JSON keys, extra arguments, and embedded invocations are rejected before Wasmtime. There is therefore no capability-only projection of otherwise unexecuted behavior.

Literal and constant-computed capabilities execute through the same Wasmtime host boundary. Truly runtime-dynamic capabilities cannot be proven or linked under the supported language and fail closed as DYNAMIC_CAPABILITY_UNRESOLVED; aliased or computed property access fails as UNSUPPORTED_OSHUN_SYNTAX. This is mediation by explicit admission or denial, not static-scan permission followed by an ungoverned call.

Automated verification#

All checks were run sequentially on 2026-07-25:

Boundary Command Result
Complete Moremi package cargo test --manifest-path apps/v7/moremi-realm-server/Cargo.toml 82 passed, 0 failed
Moremi Rust lint cargo clippy --manifest-path apps/v7/moremi-realm-server/Cargo.toml --all-targets -- -D warnings passed
Moremi Rust format cargo fmt --manifest-path apps/v7/moremi-realm-server/Cargo.toml -- --check passed
Forge Assist pnpm exec vitest run --config libs/maya/forge-assist/vitest.config.ts --maxWorkers=1 25 passed across 3 files
Forge Assist types pnpm exec tsc --noEmit -p libs/maya/forge-assist/tsconfig.json passed
Forge Assist lint pnpm nx lint @maya/forge-assist --parallel=1 passed
V6/V7 prompt boundary pnpm exec vitest run --config testing/prompt-injection/vitest.config.ts testing/prompt-injection/src/v6-v7-boundaries.spec.ts --maxWorkers=1 67 passed
Prompt-boundary types pnpm exec tsc --noEmit -p testing/prompt-injection/tsconfig.json passed
Mawu 3D pipelines pnpm exec vitest run --config libs/isis/three-d-pipelines/vitest.config.ts --maxWorkers=1 23 passed
Mawu types pnpm exec tsc --noEmit -p libs/isis/three-d-pipelines/tsconfig.lib.json passed
Changed source lint pnpm exec eslint libs/maya/forge-assist/src libs/isis/three-d-pipelines/src/mawu-studio-ai-creation.ts --max-warnings=0 passed
Patch hygiene git diff --check passed

2026-07-29 evidence refresh#

The refresh ran against repository commit 85f77dd06059c2b2e67d2e5fb6aa4617415ab723 on app-dev-server, Linux 6.8.0-134-generic x86_64, Node.js v24.15.0, pnpm 10.25.0, TypeScript 6.0.3, Rust 1.94.1, Cargo 1.94.1, and Vitest 4.1.7. No model/provider credential or production realm was used.

UTC interval Exit Exact command and result
2026-07-29T11:24:38.802Z–2026-07-29T11:24:58.507Z 0 cargo test --manifest-path apps/v7/moremi-realm-server/Cargo.toml — 82/82 library cases passed; main/doc targets contained no tests
2026-07-29T11:25:28Z, 12.38 s 1 The first pnpm exec vitest run --config testing/prompt-injection/vitest.config.ts testing/prompt-injection/src/v6-v7-boundaries.spec.ts --maxWorkers=1 reached 41 passing V7 cases but 26 V6 cases crashed during setup because ignored April compiler output under libs/shared/tracing/src/*.js shadowed the newer tracked TypeScript and called the removed OpenTelemetry Resource constructor. This host-artifact failure is retained and is not counted as a product pass.
2026-07-29T11:26:51.831Z–2026-07-29T11:27:03.619Z 0 After moving exactly 24 ignored tracing .js, .js.map, and .d.ts.map artifacts to /tmp/oshun-tracing-stale.Cxdqit, the same prompt-boundary command passed 67/67 against tracked source.
2026-07-29T11:27:13.146Z–2026-07-29T11:28:19.183Z 0 pnpm --dir libs/shared/tracing exec vitest run --maxWorkers=1 (73/73); pnpm exec vitest run --config libs/maya/forge-assist/vitest.config.ts --maxWorkers=1 (25/25); pnpm exec vitest run --config libs/isis/three-d-pipelines/vitest.config.ts --maxWorkers=1 (45/45); production typechecks for tracing, Forge Assist, and Mawu all passed

The stale files were ignored, untracked compiler output and do not exist in a clean checkout. No tracked tracing compatibility patch was necessary: tracer.ts already uses OpenTelemetry 2.x resourceFromAttributes.

The protected working lockfile retains its exact pre-task SHA-256. A frozen lockfile probe still reports the pre-existing root-manifest drift (@changesets/cli removal and @fastify/cors specifier mismatch); this slice does not change a package manifest or claim to repair that unrelated protected state.

Artifact identity#

Artifact SHA-256
apps/v7/moremi-realm-server/src/forge_assist_sandbox.rs fb56e8f3d9935c0cbd5ce6358f63f75f71d5a9bcc8584ce48a7d80826b8a0354
apps/v7/moremi-realm-server/src/lib.rs f2a633a0ce45935440e48dd29d67edff1b1777c6bf1ac40bac0724d8b8397c5a
libs/maya/forge-assist/src/invocation-authority.ts dfcc21c131973a5fadcb9a8af65f6af2b5bf370ede737395829899e17276e11c
libs/maya/forge-assist/src/invocation-authority.spec.ts 120663ef93d50953cbdea0d345a4ae38859b5d21721d7ddcb72728c6d7625429
libs/maya/forge-assist/src/forge-assist.ts ea5c5a893be8a686d525c37a2038196f484c877aaf5550cc3b1ba3b1483b70a4
libs/maya/forge-assist/src/sandbox-admission.ts 97dfdc5bd86b4779f4d4f539fbb05488c7da2b31b5af41b16fb49042a34557d0
libs/maya/forge-assist/src/capability-policy.ts 132a347814b895ef28baa366d1063ba6cdc1d8a1486415f4e7b7ed62fc6c3ab3
testing/prompt-injection/src/v6-v7-boundaries.spec.ts e5bf69f17c41eca5c3e0ad4de19822d626ae5f1cd583cf3818ad7fc7e7f29e23
libs/isis/three-d-pipelines/src/mawu-studio-ai-creation.ts 5b2e353a12f3f39d2a2d9b475b496bae1ee540019a1428526d362917dc5f57c5
libs/isis/three-d-pipelines/src/three-d-pipelines.test.ts 01c2348a10668b6fbac1401fd5ebd24bd7f09ca5dd59aa1ccef6b6a3fab15dc1

Resource and applicability note#

The host reported 15 GiB RAM, 31 GiB swap, 12 GiB available memory, and 15 GiB free disk. Another long-lived UI audit was present, so verification remained targeted and sequential. This slice changes backend security/provenance code, not rendered web, mobile, admin, or Unreal UI, so frontend-skill, product Playwright UI, mobile, and Unreal automation are not applicable.

Authority, blocker, and generated-status disposition#

  • The negative/adversarial suites reject unsupported or hidden syntax, dynamic and aliased capability access, secret-bearing or oversized payloads, undeclared imports, widened/expired/source-swapped AUTH bindings, falsified Wasmtime traces, missing semantic verification, and unsupported output claims. Each failure produces no executable or Mawu artifact.
  • Test grants, bearer tokens, semantic verifiers, and actors are fixtures. No policy requires a human release signature to implement this restricted execution boundary, and no real realm/player approval is fabricated.
  • No V7-001 implementation blocker remains. Credentialed headline backends, full-engine parity, cooked placement, human play/review, canaries, and best-authored comparisons remain explicit V7-002/V7-003 work.
  • Generated capability/status views must not convert the labeled procedural Mawu backend or local Wasmtime tests into production readiness. The completed-work-package admission verifier derives V7-001's evidence linkage from the ledger and this task-specific pack; no generated production-green claim changed.

Publication identity#

  • AUTH-bound Forge source: b8f637a8d7a8719495990544897c9e59a3d9fc60
  • Complete Forge-program source: 16d3588f7824d44cd31dbe6b9a54cf1722bd4f4d