Eve library

V7-002 measured headline-backend closure

Evidence

Revision 1

On this page

Date: 2026-07-25

Source commits:

  • RL/meta: bdf6e7113ce0aca52e0e5c3081c04226637f0465
  • terrain/vision: 8e887a83cc2fe5382164f8ec3f9fe73d6bcc95e0
  • product generation gateway: 53f1c31d29d9f3317c5d4f42633fcf55b6db0bfd
  • Themis/parity integration boundary: a95341090610bfef0010fdc9c2d6e5aed62a44fb
  • baseline/fallback governance: 171d5d624624df6f9f287f80702cb76a477d24bd
  • operational-evidence admission contract: 31f4ee1b6e7cf3e673230c45d6f6f516b52abc58

Scope and claim boundary#

This evidence closes four V7-002 tasks: measured RL/meta discovery, real local ML terrain/reference-image execution, a product-backend integration boundary for all four Mawu generation tools, and explicit governance of procedural and statistical cores as measurable baselines/fallbacks. The learning implementations are genuine, bounded CPU baselines. They are not evidence for PPO, deep/distributed RL, Earthbender, TerraFusion, Mesa, diffusion, CNNs, or product-scale models. Those capabilities do not inherit the bounded baselines' results.

The remaining V7-002 Themis/full-engine and per-backend operational evidence tasks remain open.

Measured RL path#

crucible-agents now depends directly on crucible-core. The core exposes learned:conservative, learned:balanced, and learned:aggressive strategies, each with explicit damage/survival trade-offs at equal skill (700). This lets policy choice affect simulated gameplay without using a hidden skill advantage.

train_rl_agent accepts only the tabular-q-learning algorithm for local execution. It uses an epsilon-greedy, one-state/three-action policy and updates Q values from completed crucible-core matches. Reward components are derived from measured win, damage, objective, survival, and resource outcomes. Training budgets are bounded to 600–250,000 frames; the owning fixture executes 6,000 frames.

The produced checkpoint records:

  • learned Q parameters and selected policy;
  • the reward curve and exact frames trained;
  • 50 held-out matches with derived wins and win rate;
  • separate SHA-256 training and evaluation digests.

Held-out evaluation alternates the learned policy between red and blue sides. Checkpoint validation recomputes win rate from wins/matches and rejects missing learned parameters, reward observations, or run digests.

Match-based meta discovery#

run_match_based_meta_discovery requires even per-candidate match counts and runs each candidate on both sides. The fixture evaluates a population of four over three generations with four matches per candidate: 48 completed matches. Measured wins select the top half of each generation as parents; deterministic mutations form the next generation. Every match result identity contributes to a SHA-256 aggregate digest, and final ranking uses a Wilson lower bound rather than an asserted win-rate formula. The local execution ceiling is 20,000 matches.

Reference-image vision#

loom-inverse schema v2 accepts actual decoded RGB8 pixels rather than deriving an estimate from a reference ID. It measures mean RGB/luminance, luminance standard deviation, horizontal/vertical edge energy, and saturation across all pixels. train_reference_vision_model trains an 8→12→4 sigmoid MLP that maps those visual measurements to normalized aridity, forest cover, elevation bias, and rockiness.

Training and validation are caller-supplied, separate slices. The trainer rejects duplicate reference IDs across both slices, invalid pixels/labels, fewer than 8 training or 2 validation images, more than 4 million combined pixels, and unsafe hyperparameters. Pixel features are extracted once before the bounded 10–5,000 epoch learning loop. The retained model contains every parameter, counts, epochs, measured holdout MAE, and an explicitly labeled FNV-1a-64 integrity digest. Inference recomputes that digest and rejects tampered parameters.

The deterministic test fixture uses 16 synthetic RGB training images and four distinct synthetic held-out images for 900 epochs. Its measured holdout MAE is 0.032510392, with parameter digest 84ac94caf29660f1. A complete retraining produces the same digest and metric.

Learned terrain generation#

loom-terrain-ml schema v2 adds a separate learned path without relabeling its named procedural fallbacks. train_local_terrain_mlp trains a 10→12→1 sigmoid MLP from caller-supplied TerrainTrainingTiles. Each pixel's normalized position, canny/mountain/lake/road controls, aridity, forest cover, elevation bias, and rockiness become a predicted height. generate_local_ml_terrain executes the learned model and returns procedural_stand_in: false.

The trainer requires at least two training and one validation tile, rejects duplicate tile IDs across both slices, validates every control/target value, caps supplied data at one million pixels and learning at 50 million updates, and records a measured validation MAE. Learned parameters are retained and bound to a labeled FNV-1a-64 digest that inference revalidates.

The deterministic fixture trains six synthetic tiles (384 pixels) for 600 epochs and evaluates two distinct held-out tiles (128 pixels). Its measured holdout MAE is 0.013766719, with parameter digest 5993d80611afaaa6. Complete retraining produces the same digest and metric.

Mawu product generation gateway#

runMawuStudioAiCreationWithProductBackend provides the asynchronous product execution path for the exact text-to-3d, text-to-texture, text-to-world, and text-to-npc tools. It evaluates policy before dispatch, passes a trust-zoned prompt and audit material to the configured backend, and routes a successful provider result through the existing semantic-grounding, attribution, provenance, content-hash, and output-trust finalization path. Denied requests never dispatch. Missing configuration returns backend-not-configured; typed provider failures return backend-failed; both produce no artifact.

createMawuStudioHttpProductBackend maps each tool to a separately configured endpoint and provider identity. Non-loopback endpoints require HTTPS and a token. URLs containing credentials, queries, or fragments are rejected. Requests use POST, forbid redirects, carry the exact request, tool, model, and pipeline identifiers, and include the fenced model prompt plus its audit. Responses must be JSON and stay within a bounded streamed byte limit before parsing. The gateway validates the echoed identities, exact expected media type, job and artifact references, and tool-specific payload structure. It recomputes usability locally rather than trusting a provider's claim and binds provider execution metadata and the configured provider identity into artifact provenance.

Timeout, network, HTTP, content-type, response-size, and schema failures are classified into structured codes with explicit retryability. Environment resolution fails on partial configuration; it returns no backend only when all four route groups are absent.

The product-backend specification exercises every tool through a mocked real fetch boundary, including exact routes and authorization, request trust metadata, provider provenance, semantic grounding, a structured 503, denied policy with zero dispatch, missing and insecure configuration, malformed responses, and the response-size ceiling. The owning library suite passes 27/27 tests.

This closes the integration task at the executable product gateway, not at the external-provider evidence layer. No live provider credentials or endpoints were available for this slice, so it does not claim a credentialed generation, generated model bytes, provider model/data licensing, product quality, cost/latency SLOs, or failover performance. Those claims remain assigned to the open per-backend operational-evidence task.

Themis dispatch and full-engine parity boundary#

Eunomia now separates route construction from delivery. The existing routeProposalAppealToThemis() continues to return pending-dispatch. dispatchProposalAppealToThemis() accepts only an already valid appeal route, constructs the exact Themis case and appeal reference, and invokes an injected dispatcher. Invalid lifecycle input returns not-routed without calling the dispatcher. Typed failures return dispatch-failed with a stable code, retryability, and optional HTTP status.

createThemisHttpAppealDispatcher() is the concrete product boundary. Non-loopback endpoints require HTTPS and a bearer token; endpoint credentials, queries, fragments, and redirects are rejected. Timeout and response sizes are bounded. Successful HTTP responses require JSON and must echo the configured Themis service identity plus the exact case and appeal references with a non-empty receipt and valid acceptance timestamp. The specification exercises the real fetch boundary, exact request/auth contract, successful bound receipt, zero dispatch after lifecycle rejection, retryable 503, identity mismatch, response-size enforcement, and partial/insecure configuration.

Crucible's parity reference is now schema- and provenance-bound. It records a supported runtime (UnrealEngine55 or assembled MayaEngineCore), engine build, executable SHA-256, run ID, artifact URI, UTC completion, and the fnv1a64-length-framed-v1 integrity digest of the complete scenario and loop configuration. A mismatched digest is rejected before execution. A parity pass requires both exact final-state-hash equality and a margin delta no greater than 1%; a close margin cannot conceal state divergence.

These changes close the executable integration contracts, but not the combined V7-002 evidence task. No THEMIS_ARBITRATION_* credentials or endpoint were available, and /root/workspace/UnrealEngine-5.5/Engine/Binaries/Linux/UnrealEditor-Cmd was absent. Therefore this slice has no external Themis receipt, independent Unreal/Maya run artifact, executable digest, or measured parity result. The unit full-engine fixture tests comparison semantics only and is not published as an engine run.

Baseline and fallback governance#

The V7 generation and simulation cores now expose their implementation class instead of allowing local deterministic behavior to inherit a neural product claim:

  • Crucible Agents supports only bounded tabular-q-learning locally. Its learned policy, completed-match rewards, and held-out results are measured. Named PPO, distributed/GPU, and self-play configurations fail loud. The separate local MLP policy is labeled as a bounded baseline.
  • Loom Terrain marks Earthbender, TerraFusion, Mesa, inpainting, and super-resolution results procedural_stand_in: true; only the trained local terrain MLP returns false, and its retained parameters and held-out MAE are measured.
  • Mawu's local implementation identifies itself as reference-procedural-backend. Product execution requires a configured provider, and absence fails loud rather than promoting the reference output.
  • Loom Agent's module contract identifies its orchestration, city, growth, scene, and edit paths as rule-based planners with no LLM or trained RL agent. City reward values are explicitly design target_* fields, while visual consistency aggregates caller-supplied measured viewpoint scores.
  • Crucible scenario priors are labeled DesignModelPrior, and balance data that has not been calculated is represented as not_computed.
  • Loom Diffgen schema v2 identifies itself as differentiable-procedural-baseline-v1. Prompt guidance is labeled fnv1a32-prompt-hash-target-v1; erosion residuals are labeled procedural-trigonometric-residual-v1; the records explicitly say no trained model or perceptual backend is configured. The prompt baseline executes bounded gradient steps and measures initial/final parameter loss. benchmark_procedural_optimizers executes gradient descent, deterministic evolution strategy, and random search, then derives the winner from their measured final losses.

This closes baseline governance, not neural capability. In particular, the procedural prompt hash is not CLIP, the trigonometric residual is not a neural network, and the optimizer benchmark is not evidence of generated-world quality. Neural/product systems retain their own open operational-evidence obligation.

Operational-evidence admission contract#

V7_HEADLINE_BACKEND_IDS is the exhaustive admission inventory for ten execution boundaries: Crucible tabular-Q and match-meta, Loom reference vision and local terrain ML, four separately admitted Mawu product tools, Themis appeal dispatch, and Crucible full-engine parity. The catalog evaluator requires exactly one passing record for every ID and rejects missing or duplicate records.

Each record is content-digest-bound and expires after 30 days. The evaluator enforces the backend's fixed implementation class, exact implementation version and SHA-256, and an explicit model/data disposition. Learned and external-model paths require exact model version/digest plus verified model and dataset license receipts; they cannot waive undisclosed data as not applicable. Genuinely model-free or dataset-free paths must say so with a reason. Every path still requires a verified output-license receipt bound to the exact artifact.

Admission also requires exact run/trace/request/artifact provenance; sampled USD cost plus monotonic p50/p95/p99 latency; a passing, internally consistent safety scenario count; threshold-derived quality metrics over a digest-bound evaluation set; and an exercised recovery path. Recovery must match the implementation class (provider failover, deterministic retry, or engine restart), inject at least one failure, recover inside the declared objective, and report zero data loss. Evidence changed after sealing is rejected.

The contract tests use synthetic records only to exercise validation and are not operational evidence. No retained production/provider/full-engine record was available for any of the ten entries, so the V7-002 operational-evidence checkbox remains open. This slice prevents missing evidence from being silently treated as readiness; it does not manufacture the evidence needed to close the task.

Fail-loud boundaries#

  • ppo returns RlTrainingBackendNotConfigured.
  • Distributed/GPU counts above one, self-play, unsupported baselines, invalid algorithms, and out-of-range budgets return TrainingConfigInvalid.
  • Python training.py is an orchestration and checkpoint-ingest seam. It does not fabricate training results and directs local execution to the Rust baseline.
  • The inverse ID-only and product-CNN training seams still refuse without pixels/a configured external backend.
  • Earthbender, TerraFusion, Mesa, inpainting, and super-resolution remain explicitly procedural fallbacks; ONNX records are serving plans, not runtime exports.
  • The terrain/vision fixture data are synthetic. They establish executable training, held-out measurement, and inference—not production-corpus quality, model/data licensing, or artist preference.
  • FNV-1a-64 model digests are labeled integrity checks, not cryptographic identities.
  • The Mawu product gateway is executable and contract-tested, but no external provider run was made. Its reference backend remains explicitly procedural and evaluation-only.
  • The Themis HTTP boundary has no credentialed delivery receipt, and the parity boundary has no independent full-engine receipt on this host.
  • No result here establishes full-engine parity; that V7-002 task remains open.

Verification#

Check Result
cargo test --manifest-path libs/maya/crucible-core/Cargo.toml 11 passed
cargo test --manifest-path libs/maya/crucible-agents/Cargo.toml 12 passed
PYTHONPATH=libs/maya/crucible-agents/python python -m unittest discover -s libs/maya/crucible-agents/python/tests -v 4 passed
python -m compileall -q libs/maya/crucible-agents/python passed
cargo fmt --manifest-path libs/maya/crucible-core/Cargo.toml -- --check passed
cargo fmt --manifest-path libs/maya/crucible-agents/Cargo.toml -- --check passed
cargo clippy --manifest-path libs/maya/crucible-core/Cargo.toml --all-targets -- -D warnings passed
cargo clippy --manifest-path libs/maya/crucible-agents/Cargo.toml --all-targets -- -D warnings passed
cargo test --manifest-path libs/maya/loom-inverse/Cargo.toml 7 passed
cargo test --manifest-path libs/maya/loom-terrain-ml/Cargo.toml 7 passed
cargo fmt --manifest-path libs/maya/loom-inverse/Cargo.toml -- --check passed
cargo fmt --manifest-path libs/maya/loom-terrain-ml/Cargo.toml -- --check passed
cargo clippy --manifest-path libs/maya/loom-inverse/Cargo.toml --all-targets -- -D warnings passed
cargo clippy --manifest-path libs/maya/loom-terrain-ml/Cargo.toml --all-targets -- -D warnings passed
pnpm exec tsc -p libs/isis/three-d-pipelines/tsconfig.lib.json --noEmit passed
pnpm exec vitest run --config libs/isis/three-d-pipelines/vitest.config.ts 27 passed
production-file pnpm exec eslint --max-warnings=0 … passed
V6/V7 prompt-injection boundary suite 67 passed
pnpm exec nx run @isis/three-d-pipelines:build --parallel=1 passed
Eunomia production and specification TypeScript checks passed
Eunomia governance service Vitest suite 18 passed
Eunomia production-file ESLint passed
pnpm exec nx run @v7/eunomia-governance-service:build --parallel=1 passed
cargo test --manifest-path libs/maya/crucible-core/Cargo.toml after parity hardening 11 passed
Crucible Core format and warning-denied clippy after parity hardening passed
cargo test --manifest-path libs/maya/loom-diffgen/Cargo.toml 5 passed
Loom Diffgen format and warning-denied clippy passed
pnpm exec vitest run --config libs/isis/three-d-pipelines/vitest.config.ts after evidence admission 35 passed
V7 evidence admission production/test strict ESLint passed
pnpm exec tsc -p libs/isis/three-d-pipelines/tsconfig.lib.json --noEmit after evidence admission passed
pnpm exec nx run @isis/three-d-pipelines:build --parallel=1 after evidence admission passed

Published artifact identities#

Artifact SHA-256
libs/maya/crucible-agents/src/lib.rs 08015a9d57cb30b795b1a11b3a32e04dcd44d40bad9713242884a149e1e20dc6
libs/maya/crucible-agents/Cargo.toml 30efb651c36ea8fbb29026163162b01cf2e4afcf8575400403f0f387c06206a4
libs/maya/crucible-agents/Cargo.lock a23da9a939f12ea8066234262cd962abebbd8452232933369b3ab63f15ace00c
libs/maya/crucible-agents/python/crucible_agents/training.py 988350888a729ce2f494a0d992c5a02c123d65c502bc0c056e243e0dfdba9183
libs/maya/crucible-core/src/lib.rs 2affd742ca261137f91821b7a40504d34d6d0e1556555ccdd665690e58ea28e4
libs/maya/loom-inverse/src/lib.rs 8d07e6c4d0e80e85b34f47795f2ccb753bb150a3c4bc4aeea45201855fbb9ebe
libs/maya/loom-inverse/Cargo.toml bca5e423a9ba6933b83655023f3280a332cbd9573c0f01f38b15bd86e12fce45
libs/maya/loom-inverse/Cargo.lock af65d65fb12f02bf9f064c4e562ee535520c325c8e3041a0457e982507c4dd5f
libs/maya/loom-terrain-ml/src/lib.rs 4a5a8a745e08472f1a37dcb959ef224fda0a260ffa82426ca0a607f08005712b
libs/maya/loom-terrain-ml/Cargo.toml eb15b992b3b8ad644ae8c94fc5f3a21667a06a19c6da8a4adc5938422f230857
libs/maya/loom-diffgen/src/lib.rs 87473f091d2acb0e6330be1784deadef25711bf88456e383d9dcf8969b778233
libs/isis/three-d-pipelines/src/v7-backend-operational-evidence.ts a6cf709fb25e02016df20759c53bea97ee5db2a9ba3cbd547c8db0370abcf217
libs/isis/three-d-pipelines/src/v7-backend-operational-evidence.spec.ts c36189fc3d9cf4ac22b81dbfcc249e8e26ed10a199e6a9848bc88dafae58f5c9
libs/maya/loom-terrain-ml/Cargo.lock 2677707393b730d1fd6d6e35f2a8cb24018d51635186bdb8a4ec4dddd1570e28
libs/isis/three-d-pipelines/src/index.ts ea8dfe84173e271e4d96c0cc1e738431e9d0a6c63a80777e0a2a8280aa831dc6
libs/isis/three-d-pipelines/src/mawu-studio-ai-creation.ts a32fe7a000e5b29f20f01f463c85707d47ee06c79fda35d4e9b30f2de73cc0a0
libs/isis/three-d-pipelines/src/mawu-studio-product-backend.ts c37097fff9ec7cce1c8c8b8f9cd79e56eea3a8248806126162f1a7fff200eef1
libs/isis/three-d-pipelines/src/mawu-studio-product-backend.spec.ts 4460b0a8bfecb72199a4b2de935c58226389a869595e338cfce0dd716d9f292c
apps/v7/eunomia-governance-service/src/themis-dispatch.ts d4d70c0860a286a852cd3dec8a3c6330980444fa8061465bc4d60b6b8dc922e9
apps/v7/eunomia-governance-service/src/themis-dispatch.spec.ts 004ba96d31edb578307ddbc91feca01ecb0bdb92e95f18b9a5c9680be907b960

This slice changes no user-facing UI, so browser or accessibility automation is not applicable.