EVIDENCE CASE
Hardening a translation provider after a production incident
A real portfolio incident became a regression-first fix that protects the Groq boundary from a retired model override while preserving supported explicit configuration.
01
Problem
A production error showed a retired Groq model reaching the translation boundary through a stale environment override. Fix the proven failure path without disabling legitimate overrides or changing the live site before verification.
02
Methods applied
- 01Selecting Working MethodsPrevents skill overload and overlap while preserving clear ownership between discovery, implementation, verification, review, and delivery.
- 02Building Regression TestsMakes bug fixes durable and proves the new test actually detects the defect instead of merely passing alongside the fix.
- 03Testing Integration BoundariesFinds serialization, persistence, protocol, and failure-mode defects that mocks and isolated unit tests cannot prove away.
- 04Shipping GitHub and Vercel ChangesMoves repository changes toward review and release without mixing unrelated work, duplicating pull requests, or overstating deployment readiness.
Decision record
- 01
Route by ownership
Use method selection to keep regression testing, integration-boundary verification, and delivery as distinct responsibilities instead of stacking overlapping workflows.
- 02
Reproduce before fixing
Capture the actual outbound provider model in a boundary-level regression test and require the known retired override to fail before implementation.
- 03
Guard only the proven legacy value
Map the exact retired model to the supported default while preserving other explicit model overrides.
- 04
Keep production isolated
Merge and verify in the development branch first. No production promotion is claimed, and the missing Vercel development deployment remains documented as a limitation.
03
Verification
Treat this report as evidence for the methods and verification shown here. Related-pack overlap alone is not proof that a pack was used or fully validated.
04
Result
- 01
161/161 tests passed on the development merge commit.
- 02
Changed-source lint and CI typecheck passed after merge.
- 03
The retired override now resolves to openai/gpt-oss-20b while other explicit overrides remain configurable.
- 04
No Vercel Preview was emitted for the development merge, so this case does not claim production runtime verification.