Bringing the README to recruiter caliber — and deleting the claims the code didn't back
Rewrote Mimi's README to a deep, portfolio-grade reference modeled on a sibling project. The part worth recording is the verification: several security/infra claims copied from the template (presigned URLs, HTTP Range) weren't true for Mimi, and got caught by grepping each one against the actual code before committing.
The README is the recruiter's front door, so I rewrote it to real depth — a table of contents, an engineering summary for reviewers in a hurry, the full API surface, an eleven-table data model, an architecture diagram, a security table, honest limitations. I modeled the structure on a sibling project's README that I'd already polished.
That's exactly where the interesting failure was.
Copying a strong README from another repo imports its claims, not its code. The draft confidently said Mimi serves recordings via presigned S3/R2 URLs with HTTP Range handling for iOS audio — both true in the sibling project, neither true here. In Mimi, S3RecordingStorage.load() returns a streamed ResponseInputStream and RecordingController.stream() hands it to the client through an InputStreamResource: bytes stream straight through the backend, no presigning, no range support.
Before committing, I grepped every security/infra line against the source. Six presigned/range claims were corrected and an entire "iOS audio range" row was deleted. What survived is only what's traceable to code: BCryptPasswordEncoder, the token_version revocation check in the JWT filter, findByIdAndUserId per-user isolation, the two distinct rate limiters (per-IP on auth, per-user on the AI endpoint), the env-gated local↔S3 storage seam, and the JSON clip export. I also did not claim Terraform — Mimi has a bootstrap runbook and an ECS task definition, not IaC, and saying otherwise would be the same lie in a different spot.
This is the same anti-fabrication rule that bit earlier in this project (planning docs that listed shipped features as out-of-scope), now in a new shape: cross-repo copy-paste drift. A commit message, a planning doc, and a borrowed README are all claims — grep against this repo's source is the only evidence. Recorded against 328f68a.
(Follow-up, same day: split the README into an English-primary README.md with a separate README.ko.md, since the audience is a Toronto job search — English leads, Korean is one click away.)