4 — Evaluate an ecosystem seam

Evidence before a migration

The practice-plan example is invented. The repositories below are real, but a repository’s architecture description is not proof that every dependency complies with it. The linked README snapshots were consulted for this course; no ecosystem application was built, migrated or benchmarked here.

Repository and documented context	Candidate seam to investigate	Evidence required before adopting it
GA: five layers, music domain and application hosts	One host-neutral operation shared by two existing entry points	Exact callers, input/output parity fixtures, layer references and deployment routes
Gaia: coordination and evidence workflows	Separate a transition decision from its external effect adapter	Replay fixtures, claim ownership, duplicate handling and authority checks at the effect boundary
IX: Rust algorithms and tools	Stable algorithm inputs/results with tool-specific parsing outside	Numerical equivalence, versioned fixtures, allocation and latency measurements on a fixed workload
Demerzel: governance framework	Separate evidence validation and policy evaluation from publication	Schema compatibility, provenance, negative permission tests and explicit publication authority

These are candidate experiments, not verified refactors or declarations that a repository has adopted clean or hexagonal architecture. The pins come from this site’s existing courses and Streeling lockfile; they are study baselines, not assertions about current production.

A bounded GA decision, worked

Pain: two entry points may encode the same rule differently. Assumption: they intend identical semantics; verify that before unifying them. Candidate: expose one application operation while preserving the established layer rules. AI behavior stays at its permitted layer; “put everything in the core” is not a migration plan.

Strongest counterargument: the endpoints serve different clients and intentionally differ in streaming, history or errors. Hidden costs: model conversion, cancellation, diagnostics, deployment ownership and compatibility. Simpler alternative: share the pure rule while retaining separate orchestration.

Probe: pin a commit, select ten explicit inputs covering success, invalid input, permission failure and cancellation, and run both existing paths. Ten is a proposed fixture budget, not a reported pass count. Then extract only the smallest common operation in an isolated candidate and replay the same fixtures.

Acceptance: no unexplained observable change, no new forbidden layer reference, and the common rule can run without starting a web host. Falsifier: a supposedly shared input needs incompatible business semantics. Rejection: abandon unification if preserving those semantics creates more conditional branching than the existing shared-rule alternative. Revisit: a new caller or a reproduced parity bug changes the evidence.

Keep authority separate from architecture

For Gaia or Demerzel, a clean interface does not grant permission to push, merge or publish. Keep the decision result separate from the effect, and validate authority again where the effect occurs. A fake adapter proves only how the caller handles its simulated response; it does not prove the real provider’s permissions, concurrency or failure semantics.

Cross-repository artifact contracts also carry version, identity and provenance obligations. Before changing a field, identify producer and consumers and pin their schema revisions. A common interface name is not evidence that those contracts match.

Exercise — Write a rejectable decision

Pick one row. Write six sentences: observed pain, assumption, candidate, strongest objection, measurable acceptance, and rejection/revisit trigger. Specify which evidence is currently missing.

Worked solution: IX

“We suspect tool parsing and numerical computation change together; this is not yet measured. We assume the algorithm has a stable deterministic input/result boundary. We will expose one existing computation behind that boundary and leave parsing at the tool edge. The objection is extra copying on a hot path. Accept only if fixed-seed numerical fixtures stay within the existing tolerance and measured allocations/latency remain within a budget agreed before running. Reject if conversion dominates or semantics differ; revisit when a second caller appears.” Missing evidence includes the exact implementation, baseline workload, tolerance, measurements and caller inventory. This is a testable proposal, not an adopted IX design.

Continue with the journal to distinguish completed course validation from experiments still to run.