# Glossary & Translation Memory: Let Translations Follow Your Rules
When translating technical documentation, legal contracts, or localized marketing material, relying on default machine output almost always introduces inconsistencies that break style guidelines or violate client terminology mandates. Generic engines frequently prioritize speed over precision, mixing crowd-sourced phrases with algorithmic guesses that rarely align with professional workflows. To eliminate this friction, you need a translation environment that learns exclusively from your explicit corrections rather than opaque ranking systems. This is the foundational design principle behind JoesanTrans, which treats your editorial decisions as the sole authority for repetition management.
Navigating the Five-Layer Verification Chain
Every submission enters a fixed sequential evaluation pathway: local cache, offline dictionary, translation memory, offline neural model, and finally cloud AI. The engine tests each tier in strict order and stops at the first successful match. This architecture serves two practical purposes. First, it drastically reduces latency because lightweight, highly specific lookups resolve before heavier computational routines engage. Second, it conserves resources by avoiding unnecessary network transmission. You will immediately notice this logic reflected in the interface, where each generated segment carries a transparent label indicating its origin layer. If a badge reads cache or dictionary, the response executed entirely on your hardware. If it displays translation memory, you are seeing a direct recall of a previous version you personally approved. This visibility removes guesswork from quality control and lets you trust the pipeline without auditing backend routing.
The separation between local operations and external processing is intentional and rigid. Only the final AI layer transmits your text to remote servers, while the preceding four tiers operate completely offline. Consequently, vocabulary queries, cached matches, dictionary lookups, and memory retrievals never consume network bandwidth, trigger authentication handshakes, or draw down usage allocations. The client application intentionally omits any API key storage mechanism, centralizing cloud access through a secure server relay instead. This isolation guarantees that extended drafting sessions, batch processing tasks, and travel workflows remain uninterrupted. When you modify a glossary entry or revise a memory segment, the system automatically reroutes future identical inputs through the appropriate local tier, bypassing the AI fallback entirely. Mastering this progression allows you to predict response times and allocate your daily cloud character allowance exclusively for genuinely novel content.
Shaping Memory Through Deliberate Corrections
A persistent weakness in conventional translation software is its tendency to harvest loosely matched segments from public datasets or algorithmic scoring, creating a bloated repository filled with inconsistent phrasing. To counteract this, the memory module records only the exact strings you manually edit and confirm. The underlying logic is straightforward: when you override an output line, the application captures your revised version and binds it to the original source. The next time that identical source appears, the engine retrieves your customized phrasing before attempting broader fuzzy matching or neural reconstruction. This approach prevents terminology drift and locks recurring technical terms, stylistic preferences, and approved nomenclatures directly into your active workspace.
Practically, this means you should treat every inline edit as a permanent directive rather than a provisional suggestion. Once you finalize a correction and advance, the memory repository permanently adopts your wording as the authoritative match. If project requirements evolve and a previously accepted term needs replacement, you can simply overwrite the existing entry in place. The updated phrasing immediately supersedes the older record without requiring database cleanup or manual synchronization. The system deliberately refrains from scoring, voting, or auto-filtering your contributions, recognizing that professional context cannot be accurately quantified by automated metrics. Over multiple iterations, this creates a self-reinforcing feedback loop where your file outputs grow increasingly predictable. You stop allocating time to second-guessing algorithmic fluctuations because the pipeline consistently honors your explicit overrides, even when operating in completely disconnected environments.
Deploying Regional Dictionaries and On-Demand Models
Terminology consistency extends beyond memory tracking; it requires comprehensive lexical coverage across diverse language directions. The embedded offline dictionary supports seventy-two directional combinations, deliberately incorporating regional linguistic variants such as Cantonese syntax and simplified versus traditional character distinctions. You are not required to pre-install every possible language pair to maintain baseline functionality. Instead, the application follows an on-demand acquisition model, allowing you to fetch resources precisely when your current project demands them. Individual directional neural models range from thirty-seven to seventy-nine megabytes, while a universal cross-language variant occupies approximately four hundred ninety-seven megabytes. By downloading only the configurations you actively utilize, you preserve local storage capacity and minimize initial setup overhead.
Subscription tiers align directly with this modular infrastructure. The free plan grants unrestricted access to all offline capabilities, meaning dictionary queries, cache retrieval, memory lookups, and model installations never encounter artificial caps. Daily character limitations apply exclusively to the cloud AI component, preserving continuous offline productivity for extensive translation batches. The premium tier removes remaining constraints entirely, enabling unlimited model deployments, boundless memory and glossary expansion, plus unrestricted document translation and screenshot recognition routines. Regardless of whether you operate on macOS or Windows, the interface adapts fluidly across ten localized layouts, ensuring that configuration panels and terminology dashboards remain intuitive. Because the client architecture completely isolates network communication, you retain absolute authority over when external services interact with your dataset. Trusting a strictly local pipeline transforms terminology management from a constant maintenance burden into a stable foundation for rule-bound, repeatable output.