Eve library

PROMPT-002 evidence pack — V1–V9 model-boundary migration

Evidence

Revision 1

On this page

Work-package charter#

  • Task ID: PROMPT-002
  • Owner: cross-version prompt-security maintainers; implementation owner: Codex
  • Accountable products: V1 creative autonomy, V2 commentary and Hathor, V6 cognition and Clio, V7 Forge and Mawu, V8 case generation, and V9 content generation and tutoring
  • Priority: P0
  • Target milestone: 2026-07-15 SOTA gap-closure program
  • Risk tier: Critical prompt-injection, authority-confusion, source-scope, privacy, and generated-output risk
  • Dependencies: PROMPT-001, @oshun/prompt-trust, provider contracts, and each product's existing deterministic validation
  • Affected slots: observed signals, briefs, constraints, dependencies, sources, artifacts, critique, memories, commentary cues and drafts, persona, history, world state, retrieval, event facts, realm context, source metadata, canon, learner/orchestration state, prior output, and feedback
  • Rollout strategy: additive migration at each invoked prompt renderer with fail-closed output validation; no persistence migration or UI rollout
  • Rollback strategy: revert source commit fd6ff0b7131df0f5625d80309f294dcf7379dd5b; the dependency and audit fields are additive
  • Evidence-pack location: this file

Boundary closure matrix#

Ledger boundary Invoked implementation Closed behavior
V1 ideation creative-autonomy/ideation/llm-synthesizer.ts Observed signals, categories, user direction, and recent topics are non-instruction zones; recent-topic near-duplicates are explicitly prohibited.
V1 planner/text creative-orchestrator/planner.ts, agent-loop-generator.ts, and BFF generation-producer.ts Briefs, constraints, dependencies, research/tool results, feedback, and guidance are zoned. Citation IDs can originate only in supplied source-bound dependencies/tool data.
V1 critique/revise creative-orchestrator/critique-revise.ts Artifact, rubric, dimensions, prior draft, and critique are zoned; claimed evidence must be a real artifact substring.
V1 CQJ self-refine/Reflexion content-quality-judge/self-refine.ts and reflexion.ts Prior output, critique, and remembered lessons remain advisory data and cannot acquire policy authority.
V2 live commentary calliope/match-commentary/commentary-generator.ts Raw match cues are zoned and bounded without adding a new provider round trip.
V2 post-match review same Calliope boundary Generated commentary is zoned for review; deterministic gender, region, playstyle, and prompt-control findings cannot be cleared by an empty model review.
V2 Hathor hathor/llm-npc/dialogue-generator.ts and maya-slm-runtime Persona, emotion, speech, world, nearby entities, constraints, history, and user text are fenced; logs retain lengths/hashes rather than raw prompt snippets.
V6 cognition/Clio Iris cognition gateway and agent run, V6 Clio narration, and Moirai mount Steward input, retrieval, memory, prior output, critique, tool authority, and event facts are separated; event source scope is carried through Clio; run events record hashes instead of prompts.
V7 Forge/Mawu forge-assist.ts and mawu-studio-ai-creation.ts Brief, realm, critique, source metadata, claims, grounding, and backend output are zoned. Forge truthfully labels regex policy as static admission only, never a sandbox.
V8 proposer/assets case-engine/llm-proposer.ts and case-assets/orchestrate.ts Canon and constraints are zoned; counts, uniqueness, control characters, delimiter leakage, culprit membership, and theme membership are semantically checked. Asset prompts carry renderer/hash provenance.
V9 content/tutor Metis TypeScript content generator and Python tutoring service Source bundles/previews, learner/orchestration fields, prior output, and feedback are zoned. Citations must be source-bound; retries retain audit metadata; Python uses the same protocol/version and logs only renderer/hash metadata.

Common trust and provenance contract#

The TypeScript migrations use @oshun/prompt-trust protocol oshun-prompt-trust.v1. Python uses the parity helper in prompt_trust_helpers.py; Rust uses the same delimiter and immunity semantics at the local SLM boundary. Dynamic data is normalized, delimiter-neutralized, bounded, and fenced. Authority remains in typed system/task/source/tool fields.

Prompt evidence records renderer version, exact prompt SHA-256, zone labels, source references, and transformations. Cognition, Hathor, Forge, Mawu, V8, V9, and tutoring telemetry do not add raw prompt logging. Existing provider response retention remains governed by the owning boundary's privacy policy.

Generated output is not trusted merely because a model returned structured JSON. The migrations add source-ID, exact-quote, evidence-substring, semantic case-field, deterministic commentary, and retry/fail-closed checks at the relevant boundaries.

Adversarial and negative proof#

The committed tests exercise literal, percent-encoded, Unicode/escaped, and base64-style delimiter material in user, source, retrieval, canon, memory, prior-output, critique, tool-result, world, persona, event, and backend-output positions. They also prove intended rejection for invented citations, minted tool source IDs, fabricated evidence quotes, poisoned canon, invalid suspect graphs, unsafe commentary with an empty reviewer response, and malformed V9 output after retry.

This is boundary-level negative proof for PROMPT-002. The complete held-out, provider-change portfolio and multimodal attack corpus remains PROMPT-003.

Automated verification#

Final post-merge verification used one Vitest worker and the integrated tree at 6af7cb3a6dc25aaea166c9c88d8d2517b37a2d63; the subsequent remote wave touched only unrelated Hathor/shared-math files plus the lockfile. The final combined lock was verified and published at 7da910b22a356e2393bd4fc4be706b9b1a65119f.

  • Consolidated migrated-boundary Vitest selection: 188/188 across 18/18 files.
  • BFF autonomy bindings: 30/30 across 1/1 file.
  • Complete Clio package: 13/13 across 2/2 files.
  • Complete Moirai kernel package: 15/15 across 2/2 files.
  • Complete prompt-trust package: 8/8 across 1/1 file.
  • Metis Python prompt/tutoring selection: 25/25 across 4/4 files.
  • Maya SLM runtime: 17/17 integration tests; library and doc-test targets also passed.
  • Total behavior regression: 296/296 tests across 29/29 test files.
  • Fourteen scoped TypeScript configurations passed with zero errors: creative autonomy, creative orchestrator, CQJ, Calliope, Hathor, Iris, Clio, Moirai, Forge, Isis, both V8 case libraries, Metis LLM client, and prompt-trust.
  • Nx builds passed for Clio and its dependency chain (including prompt-trust) and for Isis three-d-pipelines with prompt-trust.
  • Targeted ESLint passed with 0 errors. It reported 60 non-blocking inherited warnings, chiefly ignored test files and existing Metis nullish/non-null style warnings.
  • Prettier passed for every task source/config file. Ruff check passed; Ruff format passed for all three changed Python files. Cargo fmt passed.
  • Frozen offline pnpm lock validation passed after each remote reconciliation.

Representative commands, all with exit code 0:

text
pnpm exec vitest run --reporter=dot --maxWorkers=1 <18 migrated test files>
pnpm --filter @oshun/bff exec vitest run --reporter=dot --maxWorkers=1 src/agentic/autonomy-bindings/autonomy-bindings.test.ts
pnpm --filter @oshun/clio-story exec vitest run --reporter=dot --maxWorkers=1
pnpm --filter @oshun/moirai-kernel exec vitest run --reporter=dot --maxWorkers=1
pnpm --filter @oshun/prompt-trust exec vitest run --reporter=dot --maxWorkers=1
pnpm exec tsc --noEmit -p <each of 14 scoped configurations>
PYTHONPATH=services/metis/src python3 -m pytest -q services/metis/tests/test_prompt_trust_helpers.py services/metis/tests/test_tutoring_api_contract.py services/metis/tests/test_tutoring_orchestrator_helpers.py services/metis/tests/test_source_tutoring_helpers.py
cargo test --manifest-path libs/maya/engine-core/crates/maya-slm-runtime/Cargo.toml
pnpm exec eslint <changed TypeScript files>
pnpm exec prettier --check <changed source/config/document files>
python3 -m ruff check <Metis prompt/tutoring files>
python3 -m ruff format --check <three changed Metis Python files>
cargo fmt --manifest-path libs/maya/engine-core/crates/maya-slm-runtime/Cargo.toml --check
pnpm install --lockfile-only --offline --ignore-scripts --frozen-lockfile --filter @oshun/testing-infra

The final post-merge behavior/typecheck window ran from 2026-07-16T17:52:04Z through 2026-07-16T17:57:21Z on Linux 6.8.0-110-generic x86_64 with Node 24.14.1, pnpm 10.25.0, Python 3.12.3, and Rust 1.94.1. Model tests used scripted local provider doubles; no live provider credential or production model response is claimed.

Verification constraints and resource safety#

The host had 15 GiB RAM and 31 GiB swap. Verification was targeted, sequential, and single-worker. A 4 GiB BFF typecheck attempt exhausted its heap; the 6 GiB diagnostic run showed the BFF's existing unrelated monorepo baseline, and the sole task-scoped fixture error was repaired. Focused BFF tests and lint then passed.

The pre-commit affected gate expanded shared config changes to 199 projects and reported unrelated Lilith/testing-infra baseline failures. It was stopped instead of violating the repository's less-than-32-GiB broad-gate rule. The same task-scoped files/configurations were already covered by the passing lint, 14 typechecks, package builds, and 296 behavior tests above. No full Nx sweep, Next build, Expo export, Tauri packaging, Unreal build, or UI automation ran.

Artifact identity#

Boundary artifact SHA-256
generation-producer.ts 9b5b132d64e84e6ebea543c062992070fd663c9c78d6ae06c3af507dc8d2e549
llm-synthesizer.ts 61e86e0b3a8ffe8c5c6dbac722498b770611c29a945310dd29e1c2d1a5719a24
planner.ts 31b2b47acfd9859c5a7b7216b9dda999e69e8d0f1afd21486a49fbd4c1caaf54
agent-loop-generator.ts 23f68485d3c32be61b4f271687cdc489942594c1cbf53fb3999acfe78d493595
critique-revise.ts fb8343eead47c393ca96db0db3339b67d9bad30db09dd2a6b32638b7b03007c3
CQJ self-refine.ts / reflexion.ts ac8fbfd28b68c5d054b8f6247d5b3c9e92589cf48be810b6d7a04f0cd8a2b684 / 9276fdb9a53092032537dceb6a3912073c69782c3ad8d1d2f2b6817dd2d26588
Calliope / Hathor dialogue be904d3a71595d0de864b2f0693b5df73e283fcd6d55dcaf1abd8e8ea72ef5e4 / d5717b79cd1fb592bcbaae90a10c42ec5a8dacf8c4ea9adc504a45df64edae28
Maya SLM runtime ea70025a196b7c1cb0b6faa6339246a0d0ead44577c8d6840376da340c0724d1
Iris cognition / Clio / Moirai mount 81fabc8bfa9f8fff89bf3ff7c82b33024557228db312ded6c474c0ebefc481c3 / e65ba16cd338f5f17060aea603c5b45bb97425ad7405ce52df821a662e1c868b / 75d9ade796f469c7b4473d9a91d0db26a5c9743830c1f64f32e1ac6b608d1148
Forge / Mawu 7c9b1eb1b82605ca73c059ea384945347dc9f5fb19809c590743f3a2d5958edf / 0557fd62c2bda7e89dce6f512d1e95861c719d700f2c3357379e6aececa64520
V8 proposer / assets 83825684fe823eb68042666c7e86059f7a465afad633807b6b3e1acbb708f047 / 924944fbca08047a643dd1b367417910293c9804553157dc4e69898a97ea2922
V9 TypeScript content generator ff7c2e9f5624f68b04d183e03c066cc3912c3b2d93544b7b77966e7600c39488
Python trust helper / tutor de7a14b651e1692b04646f0d1b7bbded100101e1e2a4d424ba15779fba0d6bd1 / ee2cdb7bb5e08c22ec8102af31533ba6b232ca63af65dde433ce26e8bed75208

Scope and remaining seams#

  • PROMPT-002 closes the identified prompt-construction boundaries; it does not claim that prompt structure alone makes a model immune to all attacks.
  • PROMPT-003 remains open for the complete held-out portfolio/provider-change attack suite.
  • Semantic grounding beyond source-ID membership remains in PROOF tasks. Real executable capability sandboxing for Forge remains AUTH/V7 work; this change explicitly prevents the static regex admission check from being called a sandbox.
  • Existing Hathor safety invocation/calibration, durable V6 state, live providers, production model placement, and product acceptance slices remain under their named SAFETY, V6, V7, V8, and V9 packages.
  • No rendered web, PWA, mobile, admin, or Unreal UI changed, so frontend skill, Playwright, mobile automation, screenshots, and Unreal verification were inapplicable. No release decision requiring a human signature was made.

Publication identity#

  • Authoritative boundary source commit: fd6ff0b7131df0f5625d80309f294dcf7379dd5b
  • Source publication tip on origin/main: 7da910b22a356e2393bd4fc4be706b9b1a65119f
  • Closure evidence commit: bd9c0079efac4d966d994e00e2557f09320880e6