# Phase 169 locally actionable completion evidence

Date: 2026-07-15

## Completion audit

Phase 169 defines 160 implementation and validation requirements across eleven
compute areas and ten validation requirements. Twelve focused Rust workspaces
under `libs/neith/compute-*` contain the portable implementation. The fourteen
open ledger entries require target GPU runtimes or physical release evidence,
but their previously missing repository-local integration and grading work is
now complete.

The seven native backend enum labels no longer terminate in an unconditional
`DriverNotLinked` result. `compute-abstraction::native_backend` defines a
versioned executable-provider manifest and wire protocol for Vulkan, Metal,
D3D12, WebGPU, CUDA, HIP, and SYCL. It performs real configured-provider
discovery, process invocation, probing, device enumeration, and shader dispatch.
The core validates backend/protocol/nonce identity, the backend-specific shader
format, entry point and dispatch dimensions, unique bindings, byte limits,
device metadata, immutable inputs, output shape, and non-zero provider timing.
An absent or malformed provider remains unavailable rather than becoming a
fabricated device.

The former conformance `gates` module could only name unavailable resources and
had no evidence input or passing path. It now provides serde-backed acquisition
contracts, strict graders, negative and decision fixtures, and the
`compute-validate` operator CLI. Synthetic evidence can exercise every metric
decision but can never set `certified: true`.

## Requirement-to-evidence matrix

| Requirements       | Authoritative package | Passing tests |
| ------------------ | --------------------- | ------------: |
| 169.1.1–169.1.15   | `compute-abstraction` |            50 |
| 169.2.1–169.2.15   | `compute-lang`        |            51 |
| 169.3.1–169.3.15   | `compute-resources`   |            49 |
| 169.4.1–169.4.15   | `compute-dispatch`    |            39 |
| 169.5.1–169.5.15   | `compute-wave`        |            55 |
| 169.6.1–169.6.15   | `compute-memory`      |            43 |
| 169.7.1–169.7.15   | `compute-ml`          |            73 |
| 169.8.1–169.8.15   | `compute-sim`         |            77 |
| 169.9.1–169.9.10   | `compute-render`      |            49 |
| 169.10.1–169.10.10 | `compute-debug`       |            41 |
| 169.11.1–169.11.10 | `compute-sdk`         |            38 |
| 169.12.1–169.12.10 | `compute-conformance` |            23 |

All 588 Rust tests pass with zero failures and zero ignored cases. Every package
passes canonical rustfmt, strict all-target clippy, tests, and a release build.

## Open external acceptance gates

| Requirements | Locally complete contract                                                                                                                                                                                                                                                                              | External evidence still required                                                               |
| ------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------- |
| 169.1.2–.8   | Backend-specific provider manifests bind SPIR-V, AIR, DXIL, WGSL, PTX, AMDGPU, and Level-Zero/SPIR-V artifacts to strict probe/enumerate/dispatch operations. All seven paths pass the same spoofing, identity, buffer-integrity, and output validation.                                               | A production provider binary and identified compatible device on each target runtime/platform. |
| 169.12.2–.3  | Paired candidate/reference captures require the same SHA-256-addressed kernel suite, device/build provenance, at least 30 non-zero samples per path, and grade median reference time divided by median candidate time against the immutable 0.85 floor.                                                | Matched live CUDA and HIP reference/candidate benchmark captures.                              |
| 169.12.4     | Browser name/version and physical adapter identity are mandatory; software adapters, incomplete totals, failures, and skips cannot pass.                                                                                                                                                               | A live browser WebGPU conformance capture.                                                     |
| 169.12.5     | The soak grader requires an ordered interval of at least 86,400 seconds, at least one dispatch, and zero driver crashes, device losses, or validation errors.                                                                                                                                          | An identified live GPU run lasting at least 24 hours.                                          |
| 169.12.6     | Scaling evidence binds a SHA-256-addressed workload to at least two uniquely identified devices and grades `multi throughput / (single throughput × device count)` against the immutable 0.80 floor.                                                                                                   | A live run on two or more peer-linked GPUs.                                                    |
| 169.12.9     | Dogfood evidence binds SHA-256 identities for `libs/neith/renderer`, `libs/neith/sim-core`, and `libs/nous/core`; requires ordered engine boot, simulation dispatch, inference dispatch, engine consumption, and clean shutdown; and enforces ten minutes, 100 dispatches, and zero validation errors. | An identified live Phase 42 + 157 + 47 integrated session.                                     |
| 169.12.10    | Console evidence requires platform/devkit/OS identity, a complete test total, no failures/skips, and no device loss.                                                                                                                                                                                   | A live Phase-137 console devkit capture.                                                       |

Malformed hashes/provenance, mismatched reference backends, short or unequal
sample arrays, zero/non-finite metrics, software browser adapters, incoherent
test totals, reversed or short soak intervals, duplicate GPU identities,
incorrect source paths, reordered dogfood milestones, incomplete console runs,
and provider response spoofing all fail closed.

## Operator entry points

Native providers are configured independently per backend with
`NEITH_COMPUTE_<BACKEND>_ADAPTER`, whose value is the absolute path to its JSON
manifest. The release build also emits a file-oriented validator:

```bash
compute-validate <evidence.json>
```

Exit status is 0 only for certified live evidence, 2 for well-formed evidence
that does not certify, and 1 for malformed or rejected input.

## Verification

For all twelve `libs/neith/compute-*/Cargo.toml` manifests, run serially:

```bash
cargo fmt --manifest-path <manifest> -- --check
cargo clippy --manifest-path <manifest> --all-targets -- -D warnings
cargo test --manifest-path <manifest> --all-targets
cargo build --release --manifest-path <manifest> --all-targets
```

No browser, mobile, editor, or customer-facing UI changed, so Playwright and
mobile UI automation are not applicable.
