ENGINEERING VERSION LOG · 2026.07

Cache, sync, and data governance baseline

Data governance does not mean retaining everything forever. It means knowing what each layer stores, why it exists, how long it lives, and how it is deleted.

InkTyper Engineering2026.07.30

30 minutesAudio diagnostic cache and model idle timeout

1,000 itemsDesktop history ceiling

30 daysDefault login session lifetime

60 secondsBrowser callback code lifetime

ScopeGovernance baselineBidirectional sync and automatic retention remain open

EvidenceData classification plus lifecycle decisionsObserved results, targets, and estimates are kept separate.

StatusPartially implementedThis log records a dated engineering state, not an evergreen promise.

01 · LOCAL

The desktop keeps only what the product experience needs

Settings, encrypted sessions, and the most recent 1,000 history items stay local. Credentials use AES-256-GCM, business data is atomically replaced, and desktop audio is not retained long term.

02 · CLOUD

Identity, quota, and performance must be traceable

D1 stores accounts, devices, sessions, plans, and subscription state.
The usage ledger is idempotent by request ID.
Raw transcript, AI output, model, language, and timing are recorded.
Performance and AI editing diagnostics are structured events.

03 · GPU NODE

Audio exists only as a short diagnostic artifact

Inference nodes retain model weights and loaded state. Diagnostic audio defaults to 30 minutes, directories use 0700, files use 0600, and cleanup rolls continuously.

04 · OPEN WORK

Sync is not yet a complete two-way system

The current path is mainly client-to-cloud. Sync cursors, versions, conflict resolution, deletion propagation, local SQLite migration, and automated D1 retention remain open work.

EVIDENCE · DIAGNOSIS

What the evidence establishes—and what it does not

The system distinguishes local settings and sessions, account and quota state, performance events, and short-lived diagnostic audio. They cannot share one vague history lifecycle: credentials require encryption, ledgers require auditability, performance events need aggregation, and raw audio must be short-lived and least-privileged.

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

The current 30-minute diagnostic cache exists to reproduce short-output and model failures, with restricted directory and file permissions plus rolling deletion. Retention jobs, deletion audit, sync cursors, conflict versions, user export, and per-class access control remain required.

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.