ENGINEERING VERSION LOG · 2026.07

Launch audit: where beta ends and production begins

An HTTP 200 does not mean the product is ready to charge customers. Website, identity, inference, billing, and distribution each require separate evidence.

InkTyper Engineering2026.07.30
Controlled betaAudit conclusion at the time
60 / 60Recent transcription requests returned HTTP 200
322msObserved average inference time
4.3sObserved average Worker end-to-end time
ScopePre-release auditInstaller and payment verification remain continuous
EvidenceReal production-path smoke tests plus release gatesObserved results, targets, and estimates are kept separate.
StatusBeta-usable; not every production gate is closedThis log records a dated engineering state, not an evergreen promise.

01 · WHAT PASSED

Public pages and account flows were working

  • English and Chinese home, account, privacy, terms, refunds, and support returned 200.
  • Email login and Google OAuth had production success records.
  • The Supabase project reported healthy.
  • Free, Pro Monthly, and Pro Yearly pricing APIs worked.

02 · REAL INFERENCE

Production must be tested through the real route

A 3.258-second WebM clip travelled through the Worker, Cloudflare Tunnel, and RTX 3080 SenseVoice CUDA node. The usage ledger recorded 3258 audio_ms. AI polish also completed a real call and recorded usage.

03 · RELEASE GATES

Installer identity must match product identity

The audit found that the public DMG and the tested Electron Cloud-first client represented different product generations. A release requires a stable Bundle ID, Developer ID signature, notarization ticket, and first-run verification on a clean Mac.

04 · PAYMENT

Creating checkout is not a completed payment loop

Monthly and yearly products created real Dodo Checkout URLs, and unsigned webhook traffic was rejected. A controlled real payment, refund, and cancellation still require separate acceptance evidence.

EVIDENCE · DIAGNOSIS

What the evidence establishes—and what it does not

The audit separates an HTTP 200 from a sellable product. Account, real ASR, AI, and usage writes crossed the production path, but signing, notarization, first install on a clean machine, real payment, refund, cancellation, and recovery are independent gates. Success in one does not prove another.

Reading ruleProduction observations describe the observed path. Controlled benchmarks isolate a component. Targets remain targets until a sustained distribution proves them.

OPERATIONS · OPEN WORK

Failure behavior and the next verification gate

Every macOS replacement must verify the stable bundle ID, Team ID, entitlements, architecture, and strict signature, while preserving a backup. The Windows artifact needs Authenticode before it becomes a trusted public release. Billing requires controlled real transactions and webhook replay tests.

Failure visibility
Every fallback needs a reason code and stage timing. A successful final transcript must not erase evidence that real-time, AI, or paste failed first.
Release evidence
Keep request ID, actual model, node, queue, upload, ASR, AI, and delivery timing together so a regression can be reconstructed end to end.