Eve library

SAFETY-001 evidence pack — invoked product-path safety

Evidence

Revision 1

On this page

Status: repository-complete / closed. The invoked Hathor dialogue, Metis tutoring/agent/tool, V8 multimodal, and V7 Forge/Moremi paths enforce their safety decisions before the governed action and fail closed when required evidence or authority is unavailable. This status does not claim live third-party moderation, production deployment, population-level calibration, a general learned safety classifier, arbitrary-JavaScript sandboxing, or human approval.

Work-package charter#

  • Task ID: SAFETY-001
  • Owner: V2 Hathor, V9 Metis, V8 Ariadne/Yemaya, and V7 Moremi/Forge maintainers; implementation owner: Elikem Agudogo
  • Accountable products: V2 Hathor NPC dialogue, V9 Metis tutoring and agent runtime, V8 Ariadne/Yemaya case production, and V7 Forge Assist/Moremi realm execution
  • Priority: P0
  • Target milestone: 2026-07-15 SOTA gap-closure program
  • Risk tier: Critical generated-content, learner-safety, agent-tool, multimodal-media, and realm-capability authority risk
  • Audit finding: safety configuration, confidence, memory attribution, policy screening, media moderation, and capability inspection could exist as disconnected metadata or static proxies without governing the invoked product action
  • Dependencies: PROMPT-001 through PROMPT-003, V2 dialogue/memory contracts, V9 agent-runtime safety records, V8 G6/Sekhmet, V7 Moremi/Ixchel, AUTH-001, and product-specific safety policy
  • Affected schemas and content slots: Hathor dialogue confidence, abstention, memory attribution, and adapter output; Metis safety screening, classifier, trust-zone, agent-plan, tool receipt, and tutor metadata; V8 scanner evidence and Loom dependencies; Forge capability authority and Moremi admission receipt. No rendered UI or database migration changed.
  • Linked task IDs: SAFETY-001, PROMPT-001, PROMPT-002, PROMPT-003, AUTH-001, AUTH-003, PROOF-001, V2-003, V7-001, V8-005, and V9-003
  • Rollout strategy: fail-closed library/service boundaries with versioned, additive records; semantic media and live provider edges remain explicitly configured
  • Rollback strategy: revert each product boundary with its fixtures and downstream contract changes. Do not partially restore permissive fallback, fabricated confidence, unbound memory IDs, static-only sandbox admission, or caller-asserted safety state.
  • Evidence-pack location: this file

Closure matrix#

Ledger requirement Invoked product path Current closure proof
Invoke Hathor safety and fail closed DialogueGenerator.generateResponse Input is filtered before provider dispatch; output is validated before delivery; required evaluation throws or abstains when its filter/evidence is unavailable.
Calibrated Hathor confidence and abstention dialogue result and V2 adapter Provider, safety, constraints, and memory evidence feed versioned monotonic regression calibration; a score below 0.62 returns an explicit named abstention.
Real Hathor memory attribution MemorySystem.recall → prompt → marker validation Only retrieved IDs can be attributed; invented IDs and required-but-unused memories abstain; memoriesTriggered contains validated used IDs only.
Successful V9 safety record before planning/generation/tools tutoring, plan creation/run, and DAG tool execution Each path requires the current tenant/input/policy-bound record; missing, stale, drifted, unevaluated, review, blocked, or approval-required state denies the governed action.
V9 policy classifier plus adversarial evaluation Metis agent safety screening Versioned intent classification and deterministic hard tripwires both run; the seven-case adversarial gate must pass completely, while classifier exceptions fail closed.
V8 real multimodal providers and honest offline floor Ariadne G6 through Loom production dependencies Actual text/image/audio/video/3D refs are coverage checked; local binary and remote Sekhmet providers are wired; the text-keyword floor cannot claim media semantics or silently satisfy production.
V7 computed capability calls in a real sandbox Forge → AUTH-001 authority → Moremi admission → Wasmtime Constant expressions compile to exact Wasmtime imports; unsupported/dynamic syntax, invalid payloads, missing or widened authority, source drift, and host-call/authority-use trace drift all deny before an allowed receipt.

V2 Hathor dialogue boundary#

DialogueGenerator owns the complete safety sequence around the invoked provider:

  1. evaluate the current user utterance with configured filterContent;
  2. retrieve bounded memories from the injected Hathor MemorySystem;
  3. fence persona, world, constraints, memories, and user text as untrusted prompt data;
  4. invoke the configured provider;
  5. parse only [[memory:<retrieved-id>]] markers and strip them from surfaced text;
  6. evaluate generated text with validateNPCResponse;
  7. validate dialogue constraints and memory attribution;
  8. calculate versioned evidence and either answer or return a named abstention.

The default is a local SafetyFilter. Passing null is permitted only when the request explicitly makes evaluation optional. Required missing evaluation or filter exceptions fail closed. Unsafe input never reaches the provider; unsafe output never reaches the player.

The response exposes hathor.dialogue-confidence.2026-07-16.v1, its raw evidence, calibrated score, threshold, and components. This is bounded monotonic regression calibration over repository evidence, not provider certainty or population calibration. Explicit states are answered and abstained, with named reasons for safety, constraints, memory attribution, and low confidence. The V2 adapter subsequently validates the surfaced safety output rather than trusting a bridge-shaped object.

V9 Metis governed execution boundary#

Metis safety screenings are service-owned records, not booleans supplied by a request. An admitted record binds:

  • screening and tenant identities;
  • canonical request-input hash;
  • metis.agent-runtime.safety policy identity and version 2.0.0;
  • successful classifier evaluation and adversarial-gate result;
  • clear/allow decision;
  • a trust zone without outstanding human approval.

Plan creation screens first and embeds an immutable projection. Run execution reloads the service record and rejects missing state, tenant/hash mismatch, policy or classifier drift, failed evaluation, review/blocked state, binding drift, or required approval. The DAG executor independently screens immediately before each tool call and records the screening ID in its receipt. Tutoring performs the same governed screen before provider generation and returns learner-visible safe degradation when admission fails.

The policy has two independent surfaces: exact hard tripwires for control tokens, private-key material, destructive commands, and unsafe network targets; and a versioned deterministic intent classifier for prompt override, secret disclosure, cross-tenant access, destructive mutation, unsafe network access, assessment compromise, and policy bypass. Every screen evaluates seven paraphrased, obfuscated, and benign cases. All must match their expected blocked/review/clear result. Classifier exceptions and adversarial regression produce separate critical findings and fail closed.

V8 Ariadne multimodal G6 boundary#

G6 scans actual asset-manifest inputs rather than using transcript text as proof for media. Results name provider, version, mode, covered refs/kinds, flags, and limitations. requireSafetyClear rejects absent scanners, malformed evidence, missing refs/kinds, or any flag.

The providers intentionally make different claims:

  • KeywordSafetyScanner is a text-only offline floor and refuses image, audio, video, and 3D coverage.
  • ContentAnalyzerSafetyScanner invokes the local @yemaya/safety binary analyzer on fetched bytes. Its claim is integrity, metadata, steganography, and known-hash analysis, not semantic vision or 3D understanding.
  • SekhmetHttpSafetyScanner sends actual text or base64 media bytes to the configured governed semantic provider and validates provider identity, coverage, refs, flags, and verdict consistency.
  • CompositeSafetyScanner unions independent provider coverage and flags.

Production Loom construction requires Sekhmet configuration and an explicit media-origin allowlist. Fetches constrain protocol, initial and redirect origin, per-asset and batch bytes, streamed response size, and timeout. Production cannot silently fall back to the keyword floor.

The committed 12-case offline benchmark deliberately retains semantic misses and benign-context false positives:

Metric Result
True positives 4
False negatives 2
False positives 2
True negatives 4
Precision 0.6667
Recall 0.6667
False-positive rate 0.3333
Semantic/multimodal coverage false

This measures the keyword floor honestly; it is not evidence that media was semantically moderated.

V7 Moremi/Forge capability boundary#

Forge Assist does not treat static scanning as execution authority. Every static-clean candidate first earns source- and capability-bound AUTH-001 bindings, then submits source, declarations, authority, and hashes to Moremi's authenticated admission endpoint. Missing configuration, transport failure, malformed or foreign receipts, missing component hashes, and trace mismatch fail closed.

Moremi accepts a deliberately restricted Forge program:

  • constant string bindings;
  • direct oshun.invoke calls;
  • string literals, parenthesized constants, concatenations, and constant template interpolation for capability names;
  • strict JSON payloads without secret-bearing keys.

Runtime-dependent capability expressions return DYNAMIC_CAPABILITY_UNRESOLVED; aliased, optional-chain, computed Oshun syntax, uncompiled statements, and unsupported control flow return fail-closed parser violations. Missing, expired, payload-mismatched, or widened AUTH-001 bindings never reach Wasmtime.

For admitted source, Moremi creates a Wasmtime component whose imports derive from exact resolved capabilities. The linker exposes only declared realm capabilities, applies authority resource limits, then instantiates and runs the program. Every host call consumes the exact authority-use identity and records its payload hash. The receipt binds source/component hashes, compiled and observed call sequence, authority-use sequence, payload hashes, runtime/version, verdict, and violations. A TCP test crosses authentication, HTTP parsing, component instantiation, and host execution.

Adversarial fixture matrix#

Fixture Result
Required Hathor filter is absent or throws Provider action is denied or explicit abstention is returned
Provider invents a memory marker Attribution fails; invented ID is never surfaced as triggered memory
Required memory is retrieved but unused Explicit memory_required abstention
Metis record tenant/input/policy drifts Planning, tutoring, or tool execution is rejected
Metis classifier throws or adversarial set regresses Critical finding; governed action fails closed
V8 scanner omits an asset ref/kind G6 admission rejects incomplete coverage
Keyword floor receives non-text media Coverage is refused, not guessed
Media redirects outside the allowlist Byte fetch fails before scanner admission
Forge source hash or authority payload binding changes Moremi denies before Wasmtime
Forge capability is aliased, unresolved, or undeclared Parser or missing Wasmtime import rejects
Extra invocation argument or secret-bearing payload INVALID_INVOCATION_PAYLOAD before Wasmtime
Executed host-call, authority-use, or payload trace drifts Admission receipt is denied

Automated verification#

Current admitted value ledger#

The owning suites were run sequentially. The Forge and prompt rows precede the source-remediation commit, but neither surface changed in that commit; Hathor and Moremi were rerun after their exact changes.

UTC interval Files/targets Result Surface
2026-07-29T15:27:10.483Z–2026-07-29T15:27:13.314Z 10 110/110 full Hathor llm-npc package
2026-07-29T15:31:11.747Z–2026-07-29T15:31:15.887Z 5 38/38 Metis safety/runtime/learning slice
2026-07-29T15:32:10.210Z–2026-07-29T15:32:12.159Z 3 38/38 Yemaya case pipeline
2026-07-29T15:32:17.476Z–2026-07-29T15:32:34.513Z 10 41/41 + 3 skips V8 Loom service
2026-07-29T15:32:40.623Z–2026-07-29T15:32:41.552Z 1 18/18 Yemaya case contracts
2026-07-29T15:27:35.615Z–2026-07-29T15:27:44.924Z 1 82/82 full Moremi Rust library target; binary/doc targets contain 0
2026-07-29T15:19:47.476Z–2026-07-29T15:20:06.735Z 3 25/25 full Forge Assist package
2026-07-29T15:20:12.464Z–2026-07-29T15:20:23.587Z 1 67/67 exact V6/V7 prompt-boundary regression
Unique total 34 419/419 plus three explicit Loom skips; no assertion is counted twice

The former 378/378 total was accurate for the smaller historical suite set. The @yemaya/safety package currently has no local test files; its configured --passWithNoTests gate exited zero and is not counted as passing assertions.

Representative commands:

text
VITEST_MAX_WORKERS=1 pnpm --dir libs/hathor/llm-npc exec vitest run
PYTHONPATH=apps/metis/service/src /tmp/metis-safety-venv/bin/python -m pytest -q <five Metis files>
VITEST_MAX_WORKERS=1 pnpm --dir libs/yemaya/case-pipeline exec vitest run
VITEST_MAX_WORKERS=1 pnpm --dir apps/v8/loom-service exec vitest run
VITEST_MAX_WORKERS=1 pnpm --dir libs/yemaya/case-contracts exec vitest run
cargo test --locked --manifest-path apps/v7/Cargo.toml -p moremi-realm-server -- --nocapture
VITEST_MAX_WORKERS=1 pnpm exec vitest run --config libs/maya/forge-assist/vitest.config.ts
VITEST_MAX_WORKERS=1 pnpm exec vitest run --config testing/prompt-injection/vitest.config.ts testing/prompt-injection/src/v6-v7-boundaries.spec.ts

Type, lint, build, and contract gates#

  • Hathor, Forge Assist, and prompt-boundary TypeScript configurations passed; the commit hook also passed Hathor, V2 adapter, prompt, and BFF affected typechecks, with BFF reporting 0 known and 0 new app-owned errors.
  • Seven current V8 configurations passed: case-pipeline library/spec, Loom, case-contract library/spec, Yemaya safety, and Daedalus.
  • Hathor, Forge, prompt, and the core V8 scanner/Loom/contract/compiler surfaces have zero ESLint errors and warnings. The broader @yemaya/safety package lint exits zero with 12 pre-existing warnings in ethics/content-filter managers; this refresh does not relabel them clean.
  • Hathor Prettier and tsup build passed. Yemaya safety tsup and V8 docs drift passed.
  • Metis Ruff check/format passed on 11 safety source/test files; strict Mypy passed on six source modules. OpenAPI JSON and generated TypeScript drift checks passed with the compatible dependency set named below.
  • Moremi Rust format, locked Cargo check, strict all-target Clippy, complete library tests, and zero-test binary/doc targets passed.
  • git diff --check passed at the source checkpoint.

Refresh-discovered red evidence and remediation#

The refresh retains setup and gate failures rather than converting them into passes:

  1. System Python collected zero Metis tests because fastapi and pydantic_settings were absent. An isolated environment installed the declared apps/metis/service[dev] extra; the exact five-file suite then passed 38/38.
  2. A fresh unconstrained install resolved FastAPI 0.141.0 with Pydantic 2.13.4. Tests passed, but OpenAPI drift correctly failed because that Pydantic version collapses two previously separate input/output schema pairs. FastAPI 0.141.0, Pydantic 2.11.7, pydantic-settings 2.10.1, and Starlette 0.47.3 reproduce the checked-in JSON and generated TypeScript exactly. The service's broad lower-bound-only dependency ranges remain a reproducibility caveat; no application-schema drift is hidden.
  3. Hathor's first package-wide Prettier check found ten files. Commit 6299a9fc53213fcd0dd02ebd2a6596ea3021f50a mechanically formatted them; full tests, typecheck, lint, formatter, build, and downstream hook typechecks then passed.
  4. Rust 1.95.0 strict Clippy found one collapsible parser branch and two manual checked divisions. The same commit split duplicate-comma handling into explicit fail-closed match arms and used checked division with the existing zero-denominator fallbacks. Moremi 82/82 and all Rust static gates then passed.

Loom's green run emitted missing-source-map warnings for checked-in shared database JavaScript. Its 41 assertions passed and three integration cases were explicitly skipped. The Yemaya no-test gate also reported a ten-second close timeout before exiting zero; neither message is represented as a test pass.

Artifact identity#

Hashes were recomputed from published implementation tip 6299a9fc53213fcd0dd02ebd2a6596ea3021f50a.

Artifact / path SHA-256
Hathor dialogue boundary — dialogue-generator.ts 6e8a7896ac6c814726b56280b522470ab4e9240b451b0183a1b47380fdcb18f9
Hathor dialogue fixtures — dialogue-generator.spec.ts b0b23cff1ca57ed73961dbcd196199357b9b9d654b717ee2408ee226c877598b
Metis policy classifier — agent_safety_helpers.py 2a4e08b136b76e2c403c70c1f1e7f75adb965ee1801767e5724417566b65196e
Metis runtime record gate — agent_runtime_service.py c420873700d7fda575b3cb2f5396a69572d017dc8845d8a4b6253a3fd1fa91f1
Metis tool gate — agent_dag_executor.py e85cc9236273d89efb77080c2a39080255c8ae9f785d3b3f4a5fcd663864633f
Metis tutor gate — tutoring_service.py a413391549a5e36f8457ab6deefa920704413135a31ef1978ec74c81238a310f
V8 scanner/providers — case-pipeline/src/safety.ts ae980a44bf6fb8a83f1f4c59783033e358ed14b2a284566aeee0c5d122c0c68c
V8 keyword benchmark — keyword-safety-benchmark.ts 77f946298d6b78f7c93a48565f72a8cefca8a64df5cbed66898c0445fcf8c8ab
V8 production wiring — loom-service/src/deps.ts ab5c76c6581c28ea42438a43b90ad33e4ba10b6dc79da5345bac0db317fd8c5e
Moremi Wasmtime admission — forge_assist_sandbox.rs 02e54e26fae90e17ba80d03000192f30a78a7ffc0d11f0272022a6c4124283e6
Forge admission client — sandbox-admission.ts 97dfdc5bd86b4779f4d4f539fbb05488c7da2b31b5af41b16fb49042a34557d0
Forge invoked boundary — forge-assist.ts ea5c5a893be8a686d525c37a2038196f484c877aaf5550cc3b1ba3b1483b70a4
Forge authority boundary — invocation-authority.ts dfcc21c131973a5fadcb9a8af65f6af2b5bf370ede737395829899e17276e11c
Metis OpenAPI artifact 7a0aca2325ddd0c102aa0463ffd29ab103df0f38b7aded4dff6adbd8c4c7a925
Metis generated TypeScript contract 1966aea9128c5f6e715eb5e192414c805ef1dff9e18604e25f71656a71afbd4f
V7 Cargo lock 4d7d0181cda4c0f1ef947d1d326bfcd1a1bcb8bd7ce8b6220754bd6e291eb94d
Workspace dependency lock 08b055786e9b8a6104d5313601d7eef3205dedcf395066fd9a8264c9859d690f

Human authority and external deployment boundary#

Repository publication of deterministic policies and fixtures requires no fabricated human signoff, and none is claimed.

  • Hathor calibration is regression calibration over explicit product evidence, not population calibration, licensed human-gold validation, or a live-provider reliability estimate.
  • Metis uses a deterministic versioned policy-intent classifier and bounded seven-case gate. It is not represented as a general learned semantic model; live-model red teaming remains a release activity.
  • V8's local analyzer covers byte/media integrity, not semantic multimodal understanding. Semantic production coverage depends on configured Sekhmet; no live moderation response is claimed.
  • Moremi executes the admitted restricted Forge program, not arbitrary JavaScript. Unsupported or runtime-dependent syntax is denied rather than guessed.
  • AUTH-001 records and local TCP fixtures prove repository authority behavior, not a signed remote attestation, production realm deployment, or human approval.
  • No production provider call, customer outcome, operator decision, or deployment is claimed.

Resource-safe verification and UI scope#

The refresh host was app-dev-server, Ubuntu 24.04.4 LTS, Linux 6.8.0-134-generic x86_64, with 15 GiB RAM, 31 GiB swap, and 16 GiB free disk. The toolchains were Node.js v24.15.0, pnpm 10.25.0, TypeScript 6.0.3, Vitest 4.1.7, Python 3.12.3, pytest 9.1.1, Rust/Cargo 1.95.0.

An unrelated checkout's long-running Vitest worker and this checkout's Nx daemon were active. Verification remained targeted and sequential except for the bounded repository hook. No full-workspace Nx sweep, production web build, Expo export, Tauri packaging, Unreal build, Playwright run, or mobile automation was started.

This refresh changes backend/runtime formatting, Rust gate compatibility, and evidence only. It does not change customer/admin/PWA/mobile/Unreal UI, so the frontend skill, Playwright, mobile automation, screenshots, and Unreal verification are inapplicable.

Publication identity#

  • Original Hathor source commit: fee5bba3101af670798333cf71844f7fda779d74
  • Original Metis source commit: c06014ca70aa4e296689770d634827b281071e16
  • Original V8 source identity: bbc441f42a04b7518f832a2b4d903a8c7dde680e
  • V8 source identity on the current graph (identical tree): 7c5d061745168fe759b6c72bfb68e51801992dff
  • Original V7 source commit: 39c039a5713606982ca1b2f3b4c1e6f783fe82f6
  • Historical evidence-pack commit on the current graph: 0073cf4feee86df22f575ac05de3507e9a2d1a01
  • Historical equivalent evidence identity (identical tree): 9453041294efd6888274ca640666276879527427
  • Historical ledger closure commit: b3dce86f52fcde33ed7b913a77e2b2588d7b40e3
  • V2 surfaced-output hardening: 4436396cc24e26b19abe5a3ba4d898ae97df5146
  • V7 invocation-authority hardening: b8f637a8d7817b0e58635fea8e47c19239cbb93d
  • V7 complete restricted-program hardening: 16d3588f7824d44cd31dbe6b9a54cf1722bd4f4d
  • V8 exact eight-gate admission hardening: b0d07ff5a8a7d1ddaa08e5fd297e7354b1667013
  • Current static-gate remediation and admitted source tip: 6299a9fc53213fcd0dd02ebd2a6596ea3021f50a
  • Current evidence-refresh commit: recorded by the task-scoped commit that contains this file.