Every metric published on Mindaeon is reproduced continuously by our evaluation harness across Appliance Sizing, Foundry Verification Gates, AST Mutation Testing, Hybrid Code Intelligence, Multi-Worktree Radar, and Shadow FinOps.
Complete 8-service sovereign stack (Keycloak 683 MiB, PostgreSQL 18 158 MiB, Gitea 93 MiB, SeaweedFS 78 MiB, Control Plane 29 MiB, OpenBao 26 MiB, Edge + OPA 25 MiB) verified at idle.
Blocks test file deletion, .skip/.only injection, assertion removal, and snapshot tampering outside the agent sandbox.
Deterministic changed-surface AST mutations verify that new tests genuinely fail when production logic is broken.
Hybrid Tree-sitter AST + SCIP/LSP call graph + pgvector semantic embeddings fused with Reciprocal Rank Fusion (3.5x fewer tokens than raw grep).
Continuous Shadow Merge dry-runs across parallel git worktrees detect overlapping AST symbols in p95 = 0.01 ms.
Shadow model replay and deterministic prompt prefix ordering maximize cache hit rates without quality regression.
When an autonomous agent shares write access to both production source files and test suites inside a container, it can trivially satisfy a failing test by deleting the test file, injecting .skip / .only, or weakening an assertion. Mindaeon prevents this structurally: the Host Test Integrity Guard evaluates AST diffs outside the execution sandbox (intercepting 100% (7/7) tampering vectors with 0% false positives), and the Changed-Surface AST Mutation Harness injects deterministic syntax mutants on modified lines in p95 = 3.77 ms to prove that tests actually fail when logic is broken.
Agentic loops that rely solely on raw ripgrep / globbing burn massive context windows reading irrelevant files. Mindaeon fuses three complementary indexes using Reciprocal Rank Fusion (RRF): Tree-sitter AST structural scopes, SCIP / LSP exact cross-file call graphs, and PostgreSQL 18 pgvector semantic embeddings—achieving 94% symbol recall (vs. 62% for raw grep) while reducing prompt token consumption by 71.2%.