---
title: Validation and limits
description: What has been verified, how to reproduce it, and what has not been tested.
sidebar:
  order: 7
---

The implementation is validated against [PLAN.md](https://github.com/Stephanvs/archevolt/blob/main/PLAN.md) through nine gates in `scripts/verify.ps1`. The [acceptance record](https://github.com/Stephanvs/archevolt/blob/main/.audit/acceptance.md) holds the evidence.

## Verified

- Release build with zero warnings, and 127 tests, on Windows and on Ubuntu 24.04.
- Storage, entity identity, migration, reservations, and disposal, including a fixed-seed randomized comparison against a dictionary model.
- Generated queries, namespace placement, diagnostics `AV001` through `AV010`, and incremental generator caching.
- Planner ordering, disjoint query proofs, global resource conflicts, cycles, and barriers.
- Parallel execution under worker-pool stress across budgets of 1, 2, and 4.
- Packed-package consumers under JIT, trimming, and NativeAOT.
- Zero managed allocations on warmed traversal, filter, and dispatch paths.

## Reproduce

```powershell
./scripts/verify.ps1
```

Logs are written to `artifacts/logs`. For a Linux check, build, test, and pack in a clean copy:

```bash
dotnet build Archevolt.slnx -c Release -warnaserror
dotnet test --solution Archevolt.slnx -c Release
dotnet pack src/Archevolt/Archevolt.csproj -c Release -o artifacts/packages
dotnet run --project tests/PackageConsumer -c Release
```

## Known limits

- A forced worker `Thread.Start` failure is not tested. Partial-start cleanup was inspected, but that failure path has no runtime proof.
- Linux NativeAOT was not run. NativeAOT evidence is Windows x64.
- Structural churn allocates, and parallel dispatch has material overhead on tiny workloads.
- Benchmark baselines are exploratory. They do not establish numeric release regression limits or cross-platform performance guarantees.