VaultBook for Freelance Translators: Confidential Glossaries, Client Notes and Deadlines in One Encrypted Offline Vault

Lena had organised the evening the way she organised most deadline evenings, which is to say she had organised it completely and it had still come apart in the first twenty minutes. The contract amendment for the law firm was on the left monitor, eighty-four pages in the source and, by her count that morning, about seventy-one pages finished in English. The medical-device revision was on the right monitor, instructions for use where a single changed verb in German could mean a changed instruction to a clinician in English, and getting that verb wrong was not a matter of style. The third job, the small one she had almost not counted when she planned the day, sat in an agency portal tab: a short commercial text that was not short once you understood that every headline had to do work the literal words did not do. Three clients. Three unrelated subjects. Three sets of instructions that, in places, directly contradicted each other. And a wind-up kitchen timer shaped like a tomato, ticking on the desk because a phone timer pulled her into messages and a kitchen timer did not.

The term stopped her at 7:42.

It was not a difficult term in the ordinary sense. Lena had been translating German legal and commercial texts into English, and English back into German, for eleven years. She had rendered far harder sentences before her first coffee. The word itself was common enough in the contract, a word about acceptance — about the moment when a delivered thing is formally taken over, examined, and accepted as conforming to what was agreed. In general German commercial usage she would have handled it almost without thinking. In this client’s usage she could not, because this client had a settled preference, agreed months earlier after a long exchange, about exactly how that family of terms was to be rendered in English. The preference existed for a good reason. The firm’s own template contracts distinguished between a physical handover, a practical taking into use, and a formal acceptance with legal consequences, and if a translation blurred those three moments together it created the very ambiguity the firm paid Lena to remove.

She knew she had decided it. That was the maddening part. She could picture deciding it: the afternoon in late spring, the light in this same room, the satisfaction when a fiddly distinction finally settled into a form she could use across a whole suite of documents. She could remember writing a note about why the chosen rendering was right, with a warning about the near-miss rendering that looked plausible and was wrong for this firm. What she could not do, at 7:42 on a Tuesday evening with two other clients waiting in other windows, was find the decision.

She tried the translation memory first, because that was the disciplined place to try. Her computer-assisted translation tool — a full professional system of the Trados and memoQ class, the system of record for her translation memories and for the work itself — held the segments from the spring project, and when she searched there she found the term translated, but translated in two different ways in two different documents, because one of those documents predated the decision and one, she suspected, had been handled in a hurry by pasting from an earlier file before the agreed form had been entered properly. The memory could tell her what she had done. It could not, on its own, tell her which of the things she had done was the thing she had decided to do from then on.

She tried the termbase next. The client’s termbase was supposed to be the clean answer to exactly this question. Approved term, forbidden term, a short note on context. When she opened it the entry was there but incomplete, and the note field — the field that would have held her reasoning — was empty. Further down the same list she found a small cluster of entries she was certain she had built during the spring project, present in a different form than she remembered, and two entries she had relied on in a later job that were not there at all. Nothing dramatic had happened that she could point to. There had been no single crash she could blame. The project had been finalised, files had been returned and cleaned, a new project had been created from an old one, and somewhere in that ordinary sequence of professional housekeeping a handful of carefully made decisions had thinned out without announcing their departure.

The email thread was the next place, because the original discussion with the firm had not happened inside any translation tool at all. It had happened the way so much of a freelance practice actually happens: in a chain of messages, some from the partner who commissioned the work, some from an associate who worked on the documents day to day, one from Lena herself sent late at night with the distinction set out more clearly than she had managed in any note since. She searched her mail for the German term. Too many results. She searched for the English rendering she thought was right. Too few, and none of them the thread she needed. She searched for the associate’s name and the month and found status updates, an invoice query, a changed delivery date, a discussion about a different clause altogether. The thread she wanted existed. She believed it existed. Believing a thread exists, at 7:42 in the evening, is a very different thing from being able to put it in front of a tired pair of eyes and read the sentence that settles the matter.

Then she tried the note she might have made in the cloud notebook she used for everything that did not obviously belong somewhere else. The notebook opened. The list of notes opened. The search turned for longer than a search should turn on a broadband connection in a city apartment, and the note she wanted — if it was the note she wanted; the title was a date and two words that could have referred to three different jobs — showed her a title, a fragment of a first line, and then nothing she could use. The connection in the building had been intermittent since the afternoon. Files that other people imagined as simply being in the cloud revealed, on an evening like this, that the cloud meant a sequence of other people’s computers, a login, a sync state, and a network path between her desk and all of that, and that every link in that sequence had to hold at the exact moment she needed the answer. On a quiet Sunday that is an abstraction. Under deadline it is a practical fact with a clock attached.

By 8:05 she had five places open. The CAT tool with its memory and its incomplete termbase. The mailbox with its hundreds of threads organised by who had spoken last rather than by what had been decided. A desktop spreadsheet whose name included the word final and a version number that was plainly not final, because a second spreadsheet in a downloads folder had a higher one. A client portal in a browser tab, logged out now, holding delivery instructions she could not consult again without a fresh login and a code sent to her phone. And a folder of PDFs — the signed engagement terms, the non-disclosure agreement, a scanned style instruction from the device maker that someone had photographed at a slight angle — in which the answer was almost certainly not present at all, but which she searched anyway because searching somewhere feels better than stopping.

The device maker made the problem sharper. While the legal note was still failing to load, a message arrived about the instructions for use: a reviewer had queried a rendering Lena had used for a term connected with cleaning and reprocessing the device. Had she used that form because the manufacturer’s glossary required it, or because she had carried it over from a previous project for a different maker whose house style looked similar? Those two glossaries must never be blended. A client glossary’s whole value is that it belongs to one client: the manufacturer’s terms were product-specific, tied to its own regulatory filings and its own components, while the other maker’s terms belonged to different products and different legal exposures. To mix them would be worse than having no glossary at all, because a mixed glossary gives wrong answers with an air of authority. She had kept them separate. She had been careful. Care, spread across spreadsheets and portal downloads and attachments saved under names chosen in a hurry, had produced separation without retrievability: the terms were not mixed up, they were simply not findable as a coherent body of knowledge when she needed them.

The third client illustrated the other half of the same problem. The agency job had arrived with a brief that referred to an earlier campaign, a tone the end client liked, two phrases the end client had rejected the last time and did not want to see again, and a deadline that had already moved once that day. All of that information existed. It existed in a portal message, a short document attached to the job, a remark in an email from the project manager who was now, Lena happened to know, on leave, and in Lena’s own memory of having worked on the earlier campaign in the winter. If she had been asked in the abstract whether she kept good records she would have said yes, and she would have meant it. Her invoices were immaculate. Her project folders followed a naming convention. She backed up. What she did not have was a practice memory — a single place where the decisions, preferences, warnings, queries, and commitments that make a freelance practice more than a pair of hands and a dictionary could live together, be searched together, and be found under pressure. She had records everywhere. She had a memory nowhere.

She reconstructed the term the hard way, reading around it in the spring contract she trusted most until the evidence in front of her settled which rendering the firm had approved, and she wrote the answer on a paper index card propped against the monitor — because paper, that evening, had a better retrieval record than any of her systems. She answered the device reviewer’s query the same way, from the manufacturer’s own documentation rather than her recollection of it, and noted that the query and its answer needed keeping: next time that component was described, the exchange would save an hour and might prevent an error. At present the query lived inside a portal she would lose access to when the project closed. An asset with an eviction date is not really an asset.

Lena is not careless. She tried. The shape of freelance work defeated the shape of her tools. A law-firm engagement generates an agreement in a PDF, instructions by email, terminology in a termbase, queries in a spreadsheet returned by the client, and delivery conditions in a portal. A device engagement generates a glossary from a regulatory department, a style instruction from technical writing, reference manuals as PDFs, review comments in yet another system, and a retention obligation that starts counting from a date recorded in an email. An agency engagement generates almost everything inside the agency’s own environment, which she cannot treat as her own archive afterwards. Each system is defensible alone. Together they made the person professionally responsible for remembering the only integration point in the arrangement. Her memory, tired at eight in the evening, was the database. Everything else was storage.

There was a confidentiality dimension she felt more strongly than the inconvenience, once she stopped to name it. To find her own decision she had searched in five places, which meant fragments of several clients’ worlds were open on one screen: a clause from an unpublished contract, a device description from a manufacturer’s documentation set, an agency brief naming an end client whose campaign had not launched. She was entitled to have all of it open — that was the job. But the number of services those fragments had travelled through, and in which copies still sat, had grown without her ever deciding it should. The term discussion was in a mailbox. A glossary copy was in a synced folder. The note was in a cloud notebook. Portal copies sat on agency servers under agency retention rules. None of it was scandalous. All of it was ordinary — and ordinary practice had quietly given her clients’ confidential material a far wider geography than she would ever have drawn if anyone had asked her to draw it.

What she wanted, when she articulated it after the amendment was delivered the following morning, was not a new way to translate. She wanted a home for everything around the translation: the decisions and the reasons for them; the client preferences that contradicted each other and had to be kept apart and labelled clearly; the queries and their answers; the names of the documents she was holding, what she was allowed to do with them, and the dates by which some had to be deleted; the deadline view that showed all three clients at once without merging their material; the note she could open on a bad connection, in a train, in a client’s city with no signal worth the name, and trust to be exactly as complete as the note she had left at her desk. She wanted that home to be hers in a physical sense — a thing on her machine, under her keys, that no vendor could reorganise, analyse, or lose on her behalf. She already had a professional system for translation itself, and intended to keep it as the authority for her translation memories and live projects. What she lacked was a private practice vault alongside it.

The Week Your Notes Went Through Someone Else’s AI Vendor

In the first week of November 2026, that abstract worry about geography — where the fragments of a confidential practice have travelled, and through whose hands — acquired a specific, documented shape. Notion published a security advisory on its own site, titled “Fireworks AI incident”, worth reading in its own words and on its own timeline, without exaggeration, because the lesson for a translator does not need exaggerating. The updated version is dated 10 November 2026, and it is available from Notion itself at https://www.notion.com/en-gb/security/advisories/security-incident-involving-a-third-party-ai-vendor. What it describes is not a failure of Notion’s own core product in the way users usually imagine a breach. It describes something more instructive for anyone who works with other people’s confidential words: an incident that arrived through a third-party artificial intelligence vendor in the chain behind a feature.

Follow the sequence as Notion reports it. On the evening of 2 November, Fireworks AI notified Notion that a small dataset had been accessed by malicious actors. At that first stage, the dataset was assessed as Notion AI synthetic testing data, with no customer data involved. On 3 November, Notion disconnected from Fireworks and deleted data held there — a prompt, sensible severing of the connection once a vendor reported unauthorised access. On 6 November, the picture changed: Fireworks notified Notion that malicious actors may have accessed potential customer data. On 8 November, Fireworks confirmed the final impacted dataset, and that dataset, as Notion reports it, included Notion AI chat transcripts used for the Insights feature. Notion adds the qualifications that matter, and this article will keep them in full view: for the vast majority of those transcripts, the content was unattributable, and the dataset contained no tool-call results. On 9 November, Notion contacted the customers it had identified as impacted. The incident, Notion says, affects fewer than one per cent of workspaces. Its advice to affected customers is concrete: treat the data as exposed, and rotate passwords or API keys that were pasted into Notion AI.

Sit with those qualifications before drawing any moral. Fewer than one per cent of workspaces is a small proportion. Most of the transcripts in the confirmed dataset could not be attributed. No tool-call results were included. A reader who turns this incident into a claim that a major notes service carelessly spilled its customers’ work wholesale would be misrepresenting the source, and translators — whose entire trade is fidelity to what a source actually says — should be the last people to do that. Notion responded by disconnecting from the vendor and deleting data there, and it notified impacted customers within a day of the final dataset being confirmed. Those are the facts as the vendor itself reports them, and they describe a contained incident handled with reasonable speed once its true scope was known.

And yet the structural lesson survives every one of those qualifications intact, and it is the structural lesson, not the scale of this particular incident, that should matter to Lena and to anyone in her position. The exposure did not begin with a notes application failing at being a notes application. It began one step removed, with a vendor behind an AI feature — a company whose name most affected users had probably never chosen, never evaluated, and possibly never heard before the advisory landed. A user decides to use a notes service. She may, if she is diligent, read that service’s security pages, consider its reputation, and make a considered choice to trust it. What she cannot do in the same way is choose every company that service subsequently engages to process the things she types into its AI features. That sub-chain forms behind the product, changes as the product’s architecture changes, and becomes visible to her, if it becomes visible at all, at exactly the moment something in it has gone wrong.

Now translate that structure into a translator’s working day, because the fit is uncomfortably exact. What does a freelance translator paste into an AI chat attached to a cloud workspace? She pastes the things she is stuck on. A paragraph whose register will not settle. A clause she wants to test two renderings against. A term with its surrounding sentence, because a term without its sentence is untranslatable in any responsible sense. A client’s name, because the question she is asking is about that client’s preference. A passage from a style discussion. The very habit that makes an AI feature useful — bring it the real material, in context, and ask a real question — is the habit that places the most sensitive fragments of a confidential practice into a transcript. A transcript is not the document. Lena would never upload a client’s full unpublished contract to a general chat window and call that professional. But a transcript built up over months, question by question, can hold the contract’s hardest clause, the device’s unreleased description, the end client’s unlaunched slogan and the reasons it was changed, the glossary decision and the client’s name attached to it. It holds them in the order of her curiosity, which is to say in an order that maps precisely onto what was difficult, novel, or commercially sensitive in each job. Routine text she translated without help. The transcript collects the exceptions. The exceptions are the practice’s crown jewels.

Notion’s own advice in the advisory makes the mechanism plain in a single line of guidance: affected customers were told to rotate passwords and API keys they had pasted into Notion AI. That advice exists because pasting is what people do with a helpful text box. They paste the thing they are working on, including, in a pressured moment, things that function as secrets. A translator’s equivalent is not usually a password. It is something with a longer and quieter life: an unpublished product name, the wording of a clause a firm has not yet filed, a term decision that reveals how a client’s template allocates risk, a patient’s details inside a document presented for certified translation. You cannot rotate a client’s unpublished contract the way you rotate an API key. Once a confidential source segment is treated as exposed, the remedy on the translator’s side is notification, apology, and whatever the engagement terms and the law provide — consequences of an entirely different weight from generating a fresh credential.

It is also worth being precise about whose decision created the risk. Lena did not choose Fireworks. No translator using an AI feature inside a notes product chose the inference provider or the security posture of a company she had no contract with and no reason to know existed. That is not an accusation against any single firm; it is a description of how layered software now works. Each layer is, very often, competently run. The chain as a whole is longer than the user’s trust decision, and the user bears a professional duty — to her law firm, her device maker, the agency’s unseen end client — that she cannot delegate down a chain she cannot see. When her non-disclosure agreement says she will safeguard confidential information, it adds no asterisk for information safeguarded by a vendor of a vendor. The duty stays with the person who signed. The architecture should match the duty.

Against that background, consider what changes when the practice vault makes zero network requests. That is a plain, checkable property of VaultBook: it is a local-first, single-file application that does not phone home, does not sync through a server, and does not route the contents of an entry through a third party in order to be useful. There is no AI vendor in its chain because there is no chain. When Lena searches her own glossaries in it, the search runs locally. When she uses its built-in transcription to turn a client’s recorded briefing into text she can file and search, the transcription runs locally, with a local Whisper model, on her own machine — the client’s voice does not take a trip to a transcription service to become a note. When she runs optical character recognition over a scanned style instruction or a photographed document so that its text becomes findable, that OCR — a local tesseract engine — happens where the file already is. The features a cloud product often delivers by sending your material away are, in this design, delivered by computation that stays in the room.

That difference is sometimes described as a preference for offline tools, as if it were temperament, like preferring paper. For a translator under non-disclosure obligations it is better understood as custody: who holds the material, on what machines, under whose keys, for how long, and with what other parties able to reach it. A local vault gives a short answer to every part of that question. The material is in one file, on machines Lena controls, protected entry by entry with encryption she unlocks herself. There is no workspace for a vendor to include in a percentage. There is no transcript of her questions accumulating as a by-product of asking them, because the questions are put to her own data, locally, and the asking creates no copy elsewhere. An incident that begins in somebody else’s vendor relationship has no equivalent opening in an architecture with no vendor relationship to begin it.

None of this requires believing that cloud notes are reckless, or that the people who run them are careless — the advisory itself shows a company detecting, severing, and disclosing, which is what responsible handling looks like after an incident begins. Nor does it require believing that a translator should never use an online service for anything; Lena’s clients run portals and commission work through systems she does not control, and she uses those systems because the work requires it. The argument is narrower and stronger for it. For the one body of material that is hers to govern — her practice memory, which she alone is responsible for keeping both available and confidential — she should choose an architecture in which the number of other parties holding it is zero. Not small. Not under one per cent. Zero, by construction, on every ordinary day. The Notion and Fireworks week is useful precisely because it was, by the reporting, a limited incident: it shows the shape of a risk clearly while the stakes were still modest.

Why a Translator’s Notes Are Not Ordinary Notes

Spend an hour reading how translators talk about their tools — in the professional forums and the long, practical threads where freelancers compare what broke, what vanished, and what nearly cost them a client — and a pattern emerges that no feature list captures. The anxieties are rarely about translation in the abstract. They are about custody of the apparatus around it: the glossary that is no longer complete, the memory that no longer agrees with itself, the deadline visible in one system and invisible in three others, the queasy pause before pasting a client’s paragraph into some free online tool because the paragraph is the only way to explain the problem and also the thing she has promised not to circulate. Lena’s Tuesday evening was not an exotic failure. It was the ordinary failure, whose ingredients recur across the working lives of people who translate for several masters at once.

Begin with glossaries and termbases, because they look like the simplest part of the practice and turn out to be among the most fragile. A glossary begins life clean: a client sends a list, or a translator builds one during a first project, and for a while it is a small, trustworthy object. Then the work continues. A second project adds terms. A reviewer corrects three entries and the corrections come back inside a file whose name resembles the original closely enough to invite confusion. A project is finalised inside the CAT tool, and the housekeeping that finalising performs — returning files, cleaning projects, preparing the next job from the bones of the last — quietly changes what is attached to what. Entries a translator distinctly remembers creating are not in the list she opens under pressure. Duplicates appear, with rival spellings and no note to say which was approved later. Translators describe versions of this with a recognisable exasperation: not the anger of a dramatic data loss, but the slower corrosion of trust. Once a termbase has surprised you twice, you stop consulting it first. You consult your memory, then a past translation, then the client file, then — last, and with resentment — the glossary that was supposed to settle the question in seconds. A reference you do not trust is worse than an untidy one, because its failures surface exactly when there is no time to route around them.

Translation memories carry the same risk at larger scale, and the forum conversations return again and again to one arrangement: the memory kept in a folder that is also a synced cloud folder, so it will be available on a second machine and safe if the first dies. The intention is sound and the failure mode subtle. A large memory is a working database, updated constantly while a project is open; sync software sees a file in motion and does its own work on its own schedule; two machines, or a machine and a sync process, treat versions of the same resource as candidates to be reconciled by rules that know nothing about translation. The outcomes translators report range from the irritating to the frightening: a memory that opens with errors, a memory rolled back to an earlier state, duplicated resources with divergent contents, a project behaving as if last month’s segments had never been confirmed. The lesson the pattern teaches is plain. The resources a translator can least afford to corrupt must be treated as precious local objects — backed up deliberately, copied deliberately, never left to be negotiated over by software whose job is to make folders match rather than to keep a body of linguistic decisions intact.

Crashes form the third recurring theme, arriving with a timing so consistent it would be comic if the tellers were not so tired. A CAT tool that has behaved for weeks chooses the final evening of a long project to stall on a large file, mishandle an update, or reopen a project in a state that needs twenty minutes of untangling before a single sentence can be translated. Experienced translators build rituals around this: save compulsively, export early, keep the delivery copy separate from the working copy, refuse updates mid-commission. Lena has all of them. Notice what they protect, though: the translation itself, the deliverable. They do not protect the surrounding knowledge — the note about why a term was chosen, the query sent and not yet answered, the instruction about which documents must be deleted in ninety days. When a tool crashes at 9 p.m., every scrap of emergency effort flows toward the deliverable, and the practice notes in other windows are saved only if they happen to save themselves.

The deadline problem looks simpler and is just as corrosive. A freelancer with one client can hold a deadline in her head. A freelancer with a law firm that thinks in court-adjacent dates, a manufacturer that thinks in release milestones, and two agencies that think in portal countdowns — each of which may move, one belonging to a project manager who has gone on leave — needs a single view no individual client system will ever give her, because every client system quite properly shows only its own work. Translators talk about this in the language of juggling and dread: the job quiet for a week turns urgent on the same afternoon as the job that was always urgent; a delivery date buried in the fourth message of a thread is remembered as later than it is; a small job accepted as a favour expands into the space reserved for a large one. The failure is rarely a missed deadline outright — professionals protect delivery dates fiercely, and Lena has never missed one. The failure is the cost of the protection: evenings restructured at short notice, and the low-grade vigilance of a person who knows no single screen in her working life shows her the whole of what she has promised.

Underneath all of it sits the confidentiality anatomy, where a translator’s notes part company decisively from ordinary notes. Consider what Lena holds in any given week. Source texts that have not been published: a contract amendment whose very existence signals a transaction a counterparty may not yet have announced; device instructions describing functions that will reach the market under regulatory controls; a campaign text whose launch date is itself a commercial secret. Material that sits next to patents and product development without being a patent itself — descriptions detailed enough that a competitor reading them would learn the direction of a product line. And, in certified and legal work, personal data of the most literal kind: names, dates, certificates, disputes, the paper trail of individual lives submitted for translation because an authority in another country requires them in another language. Ordinary notes contain what their writer thought. A translator’s notes contain what her clients are, what they are doing, and what has happened to private individuals — almost none of it hers to reveal, and much of it hers to protect under a signed agreement, under data-protection law, and under the retention and deletion duties that form part of a serious engagement.

The query log deserves its own paragraph, because outsiders underestimate it. When Lena asks a client a question — does this clause mean the supplier or the purchaser bears the cost if acceptance is delayed; is this component cleaned by the user or only by service personnel — the question itself is information. A set of queries about a contract reveals where the contract is ambiguous, which is to say where it is vulnerable. A set of queries about a device reveals which instructions the manufacturer itself found hard to state unambiguously. Queries reveal strategy in aggregate even when each looks innocuous alone, and the answers are client decisions that may govern dozens of future documents. Lena’s query exchanges are among the most valuable and most sensitive records in her practice, and in a conventional arrangement among the worst housed: split between sent mail, portal threads that close with the project, and spreadsheets returned in formats that match nothing else she keeps.

Glossaries, finally, are not merely useful. They are client intellectual property in a concentrated form. A mature client glossary encodes how a firm names its own products, services, obligations, and distinctions; it may have been built by the client over years, or built by Lena under an engagement that makes the resulting terminology the client’s asset. A style guide is the same kind of property wearing different clothes, and the guides in Lena’s practice contradict each other cheerfully: one client wants a formal, spare English with every defined term capitalised and no contractions anywhere near the page; another wants contractions, sentence fragments in headings, and a voice that sounds like a person at a trade fair rather than a person in a courtroom; the device maker wants terminology so controlled that a synonym is treated as an error, while the agency’s end client asks her, in terms, to vary the language so the campaign does not sound translated at all. A translator who keeps these guides in one undifferentiated pile of notes risks the cardinal sin of her profession: applying the right rule to the wrong client. Separation is not filing tidiness. Separation is the substance of the service.

Retention closes the circle, because confidentiality for a translator is not only about keeping things safely. It is also about ceasing to keep them, on schedule and provably. Engagements end. Agreements specify periods after which source material should be deleted, or returned, or retained for a defined time and then deleted. A translator who cannot say, client by client and job by job, what she holds and when each holding expires does not fully control her own compliance — she is hoping that folders named years ago contain what she thinks they contain. The anxious version of this problem is familiar from translators’ discussions of pasting client text into free machine-translation or convenience tools: a moment’s practicality weighed against a promise, with no good record afterwards of what went where. The mature version is quieter and more demanding. It asks for a practice in which every client’s material, decisions, and dates are kept apart, kept findable, kept encrypted where they rest, and kept only as long as they are supposed to be kept. Lena did not need a more powerful way to translate. She needed her notes to be treated, by the tool that held them, as what they actually were.

One Vault Per Practice, One Entry Per Client

The design Lena settled on begins with a decision about what VaultBook is and what it is not, and she states it to herself, and to any client who asks about her tooling, without hedging. Her CAT tool remains her translation system of record. Translation memories live there. Live projects are translated there. Segmentation, concordance, quality checks against the memory — that is the CAT tool’s country, and VaultBook does not enter it. VaultBook is the private practice vault alongside it: the home for everything the CAT tool was never designed to hold and the mailbox, the cloud notebook, the spreadsheets, and the portals held badly. Keeping that boundary plain matters, because a tool adopted for the wrong job gets blamed for the wrong failures, and because Lena’s clients are entitled to a straight answer about which system governs their translated text. The vault governs her practice. The CAT tool governs her translations. Each makes the other more reliable by not pretending to be it.

VaultBook itself is disarmingly simple to describe, and the simplicity is the security story. It is a single-file HTML application — one file of roughly one and a half megabytes — that runs locally in a browser and makes zero network requests. There is no account to create because there is no server keeping a copy. There is no sync to configure because there is nothing to sync with; moving the vault means moving one file, deliberately, the way you would move a physical folder from one cabinet to another, and knowing exactly what you moved. Inside that one file lives a complete working environment: entries organised with labels, wiki-style connections between notes, and ordinary notes; a full search that reaches across the vault, including semantic question-and-answer search over her own material, optical character recognition through a built-in tesseract engine for scanned pages, and indexing of attachments so that a PDF she filed months ago can be found by what it says rather than by what she happened to name it. Around the entries sit the built-in tools of a small office: a File and Folder Analyzer for seeing what a delivered document set actually contains; PDF merge, split, and compression for the constant reshaping of client files; an MP3 cutter and joiner and a local Whisper transcription tool for recorded briefings and dictated notes; a Photo and Video Explorer that reads EXIF data, for the site photographs and document images clients send without context; a Kanban board; a Reader for long reference texts; data visualisation for the figures she keeps on her own workload; a password generator; and an Obsidian import for anyone arriving with an existing body of notes. A built-in calendar, a sidebar that surfaces what is recent, due, and expiring, dynamic tabs for moving between clients without losing her place, related-notes connections, a private analytics sidebar, inline audio and video players, and version history for entries complete the working picture. None of that list requires a connection. All of it is in the file, on the machine, under her control.

The feature that changes the moral character of the arrangement is per-entry encryption. VaultBook encrypts individual entries with AES-256-GCM, with keys derived through PBKDF2 using the Web Crypto API — in ordinary language, a client entry is not merely stored in a private place but sealed, so the vault file by itself, without Lena’s passphrase for that entry, does not disclose what it contains. That granularity fits a translator’s obligations exactly. Her practice is not uniformly sensitive: a note on German commercial idiom is a professional asset but hardly a secret, while the law firm’s entry holds its preferences, the people who gave them, the warning about the near-miss term, and engagements that are themselves confidential, and the device maker’s entry holds an unpublished product’s terminology and its review history. Encrypting entry by entry lets Lena match protection to duty, client by client. The vault can sit on her machine as one portable object while its most confidential contents remain individually closed.

Her structure inside the vault follows the shape of the practice, and its first rule is the one in this section’s heading: one entry per client, encrypted, and one client per entry. The law firm has an entry. The device maker has an entry. Each agency has an entry, and within agency work she is careful — an agency is her client, but the end client’s material she learns through the agency is treated with the same separation as if the end client had commissioned her directly, because the confidentiality runs with the material, not with the invoicing. Inside each client entry, the client’s world is allowed to be complete: who commissions, who reviews, who answers queries and how quickly, which documents govern the engagement, what the non-disclosure agreement requires, and when it requires it. Beside the client entries sit the glossary entries — one per client, never merged — so that the firm’s controlled renderings and the manufacturer’s controlled renderings can each be searched, found in seconds, and never mistaken for one another. The glossary entry is where the missing piece from that Tuesday evening now lives in permanent form: approved term, rejected near-miss term, and, crucially, the reason, in the sentence Lena wrote when she reconstructed it, explaining the distinction between physical handover, taking into use, and formal acceptance, so that a future Lena, tired at 7:42 on some future evening, is handed the reasoning and not merely the conclusion.

Style-guide entries follow the same logic, one per client, because contradiction between clients is a feature of the practice rather than a problem to be averaged away. The firm’s guide can demand capitalised defined terms and forbid contractions in the same vault where an agency client’s guide demands contractions and fragments, and the vault keeps both honest by keeping them labelled, separate, and attached to the client they govern. When Lena opens a job, she opens the client’s entry, glossary, and guide together — dynamic tabs make that movement quick — and related-notes connections surface the older decisions she linked when she made them. Query-log entries complete the record: one running entry per client, each query dated, with its context, the answer when it arrived, and the documents that answer now governs. That log is, within a month, the most consulted object in her vault, because so much of translation quality is not knowing a language — it is knowing what this client decided last time this exact ambiguity appeared, and being able to prove, if challenged, when and by whom it was decided.

Deadlines and retention live in the same structure rather than in a separate system that might drift out of agreement with it. Each job’s delivery date goes onto the vault’s built-in calendar from the client entry where the job is described, so the calendar is not a fourth place to maintain but a view over the records she already keeps — and for the first time in her freelance life, one screen shows the firm’s court-adjacent date, the manufacturer’s milestone, and both agencies’ countdowns in a single honest picture of the week. Retention dates travel with the material they govern: the date a project’s source files must be deleted is written in that project’s record, and the sidebar’s Due and Expiring views bring such dates forward before they arrive, turning deletion from an annual act of archaeological guesswork into a short list of specific obligations. When a retention date passes, Lena deletes, records that she deleted, and lets the entry’s version history preserve the fact and date of the deletion without preserving the material itself — an audit trail of compliance that does not defeat its own purpose.

What this feels like in daily use is less dramatic than the evening that caused it, and that is the measure of its success. A query arrives from the device maker’s reviewer; Lena opens the manufacturer’s entry, finds the earlier query on the same component in the log, and answers in minutes with the previous decision quoted back from her own record rather than reconstructed from a closing portal. The firm sends a new amendment in the same suite as the spring contracts; the glossary entry gives her the acceptance term, the rejected alternative, and the reason before she has translated a paragraph. An agency project manager she has never worked with asks whether she can take a rush job; the calendar answers a question that used to require opening three systems and doing hopeful arithmetic, and answers it while she is on a train with a connection too poor to load a cloud notebook — because the vault was never waiting on a connection in the first place. The work itself has not changed. The sentences are as hard as they were. What has changed is that the knowledge around the sentences now lives in one place, sealed where sealing is owed, searchable as a whole and separable by client, in a file that is hers in the same uncomplicated sense that the index cards on her desk were hers — except that, unlike the index cards, it can find the acceptance term before the timer runs out.

Glossaries That Survive the End of the Project

The moment a translation project ends is, for most terminology, the moment it quietly starts to die. Lena learned that the hard way long before VaultBook. Her CAT tool does exactly what a CAT tool should do: it stores translation memories, it suggests matches while she works, and it holds a termbase that is invaluable inside an active job. What it does badly — because it was never designed to do it at all — is preserve the reasoning behind a term after the job is finalised, the files are delivered, and the project folder has been archived, moved, or in one early-career case, corrupted when a laptop drive failed during a Windows update.

A termbase entry in a CAT tool is usually just a pair: source term, target term, perhaps a domain tag if she remembered to add one. It tells her what she used. It does not tell her why she used it, who asked for it, what the alternative was, or whether the client changed their mind three months later. For legal and commercial work, that missing context is not a nicety. It is the difference between reusing a term confidently and re-opening an old email thread at midnight to check whether “Gesellschafterversammlung” had been rendered as “shareholders’ meeting” or “general meeting” for that particular law firm, and whether the partner who reviewed the articles of association had insisted on the second form for a reason that would also apply to the contract she is translating now.

In VaultBook, the glossary lives differently. Lena keeps one entry per client, and inside the practice of using that entry — entries, labels, notes and wiki-style links working together — she builds a glossary record for each significant term that is closer to a case note than a dictionary line. A typical record is short to write and enormously valuable to reread. It has the German term as she encountered it in the source. It has the English rendering she delivered. It has the sentence, or at least the clause, where the term mattered, because a term without its sentence is a term she will misjudge later. It has the client’s decision in the client’s own logic: approved as house style, approved only for this document type, rejected alternative with the reason given. It has the date. And it has who approved it — a named reviewer at the firm, an agency project manager relaying a client’s in-house counsel, the medical-device maker’s regulatory lead — not as gossip, but as authority. If a term was approved by regulatory, she does not override it on a marketing manager’s casual preference without flagging the conflict.

Labels do the sorting work that folders used to do badly. She labels glossary notes by client and by domain, so a search for the medical-device client filtered by the regulatory label surfaces only the terms that were decided under regulatory scrutiny, not the looser language from a product brochure for the same client. The same German word can, and does, have two different approved English forms for two different clients, and a third form she uses when neither client’s preference applies and she is following standard legal usage. That is not inconsistency. That is professional translation. The vault is the first place she has kept those parallel decisions without them collapsing into one “correct” answer that was, in fact, correct for nobody.

Search is where the investment pays back. Six months after a project closes, an agency sends her a follow-up: a short amendment to a distribution agreement she translated in the spring, for the law firm client, under the same NDA. Her CAT tool has a new, clean project. The old termbase, if she can find the exported file at all, gives her pairs without provenance. In the vault she searches the client name and two or three distinctive terms from the amendment. Full search brings back the glossary notes, the query log entries attached to them, and the note where she recorded that this firm prefers “herein” to be kept sparingly, prefers British spelling even in documents governed by German law, and changed its rendering of one central concept halfway through the spring job after a partner review. Semantic search helps when she remembers the idea but not the exact term — she can ask the vault, in effect, what was decided about liability language for this client, and be led to the relevant notes rather than having to guess the keyword she used in March.

Version history turns that record from a snapshot into a history. Clients change their minds. A medical-device maker rebrands a component, a law firm merges and adopts the other firm’s house terminology, a regulator updates the preferred English name for a procedure and every document after a given date must follow it. In a flat termbase, the new term simply replaces the old one, and the past disappears. In the vault, the glossary note carries its history with it. Lena can see that a term was rendered one way until a specific date, that the change came from a specific reviewer with a specific reason — a new compliance manual, a trademark clearance, a complaint from a UK distributor — and that documents before that date should not be “corrected” retrospectively. When a client later asks why a 2024 document and a 2026 document differ, she does not have to reconstruct an answer from memory or apology. She can point, privately in her own records, to the decision, the date and the authority, and reply with calm precision.

The direction of travel matters, and she states it plainly to younger translators who ask: the CAT termbase is now rebuilt from the vault, not the other way round. At the start of a new job for a returning client, she opens the client’s vault entry, filters the glossary notes by domain, and exports — mentally, or by copying the approved pairs — the terms that belong in this job’s termbase. The CAT tool gets the clean pairs it needs to prompt her while she types. The vault keeps everything the CAT tool cannot hold: context, authority, disagreement, change. If the CAT project is lost, finalised badly, or never reopened because she has moved to a different tool for the next agency, nothing essential is lost with it. The termbase is a disposable working copy. The vault is the record.

That reversal also changes how she handles the end of a first project for a new client. She used to treat terminology work as job overhead, captured hastily because the next deadline was already pressing. Now the last half hour of terminology tidying is not overhead at all. It is the moment she converts a fortnight of decisions — many of them made by other people, in emails, in tracked changes, in a two-minute phone call — into an asset she owns. A client who returns eighteen months later does not find a translator starting from zero. They find a translator whose first draft already sounds like their own documents, because she kept the decisions after the project that produced them had closed.

One small habit protects the whole glossary system from her own haste. When a term is contested — two reviewers disagree, or an agency relays a preference she suspects the end client never actually gave — she does not resolve it by choosing the most recent voice. She records the disagreement itself in the note, marks the term as contested with the date, and raises it as a formal query on the next job for that client. Twice, that restraint prevented her from baking a project manager’s guess into a client’s long-term terminology. A glossary that can hold an unresolved question honestly is far more trustworthy, in daily use, than one that pretends every pair was settled with equal authority.

Style Guides, Query Logs and the Answers Clients Gave You

If glossaries are where a client’s words live, style guides and query logs are where a client’s judgement lives — and judgement, unlike terminology, contradicts itself cheerfully from one client to the next. Lena’s law firm wants sentences broken up, modern punctuation, and a deliberate avoidance of archaic legal filler even when the German source is dense with it. One of her agencies wants the opposite for a different end client: closeness to the source, long sentences preserved, nothing “smoothed” without a query. The medical-device maker wants numbers written one way in instructions for use and another way in clinical summaries. None of these preferences is wrong. All of them are fatal to apply to the wrong job.

She keeps each client’s style rules in that client’s vault entry, as a living note rather than a PDF buried in an email attachment. When a client sends a formal style guide, she does not simply file the document; she works through it once and writes the rules she will actually need at the keyboard, in the order she will need them: spelling variant, capitalisation habits, how headings are handled, how citations and cross-references are rendered, what to do with German abbreviations on first use, whether EUR amounts keep the German formatting or adopt English conventions, how warnings are phrased. Where the guide is silent — most guides are silent on exactly the points that arise on page nine — she adds the rulings she obtains job by job, dated, so the note grows into a guide that reflects the client as they actually review, not only as they once briefed.

Keeping the guides apart is the whole point. Labels by client, and the discipline of one entry per client, mean a rule never drifts across a boundary in her own system the way it can in a single sprawling document called “style notes” that she kept in her second year freelance. That early document still embarrasses her to remember: three clients’ preferences interleaved, amendments untracked, and one memorable delivery where she applied a US agency’s serial-comma enthusiasm to a London firm’s document because she had searched, found a comma note, and applied it without checking whose it was. The vault does not make her more careful by magic. It makes carelessness harder to sustain, because the client label is on the note itself, visible every time she opens it.

The query log is the most disciplined habit VaultBook has enforced, and the one she would now defend most strongly to any translator who works under NDA. A query in legal and commercial translation is not a confession of ignorance. It is a professional instrument: this segment is ambiguous, these are the two readings, this is what turns on the choice, which does the client intend? The problem has never been sending queries. The problem is what happens to the answers. Answers arrive in agency portals, in email threads with six other subjects, in a reviewer’s marginal comments, in a call she takes while making tea. Each answer is authoritative for its segment and, very often, instructive far beyond it: a client who explains that a clause is about indemnity, not warranty, has taught her how that firm uses two neighbouring concepts across all its documents.

Her log format never varies, which is why it works under time pressure. Question sent, with date and the segment reference as it appeared in the job. Answer received, paraphrased faithfully in her own words, with who gave it and when. Segment it affects, identified precisely enough that she can find it again in review. And then — the step she used to skip — the consequence: did this answer change a glossary term, confirm a style rule, or close off an interpretation she must not use again for this client? If it did, she links it. Related notes in the vault let a query answer sit beside the glossary entries it changed and the style note it amended, so months later, opening the term, she also sees the query that settled it. The knowledge forms a small, navigable web inside the client entry instead of a chronological swamp she has to wade through from the top.

Reviewer feedback gets the same treatment, captured once and reused deliberately. Every experienced translator knows the sinking pattern: a reviewer corrects the same preference on three separate jobs because it was never written down anywhere the translator actually looks. A reviewer at the law firm consistently changes her rendering of a particular German construction into a shorter English form — not because hers is wrong, but because the firm has decided, internally, that the shorter form reads better to its English-speaking clients. In the old workflow that correction lived and died in returned files. In the vault it becomes a dated note in the style section, phrased as the firm’s preference, with an example pair. She checks it before delivery, not after correction. Reviewers notice. Two of her steadiest working relationships improved measurably, she believes, not because her German improved — it was already the reason they hired her — but because the same conversation stopped happening twice.

There is a confidentiality dimension here that is easy to miss. Query logs and reviewer comments can be as sensitive as source texts. A query about a contract clause reveals that the clause exists. A reviewer’s blunt internal comment — “legal says we cannot admit this wording externally” — is not hers to store on a shared platform, a collaborative workspace, or any service where an agency or a future integration might surface it to the wrong eyes. In the vault, those notes sit inside the client’s own encrypted entry, on her machine, under the same per-entry encryption she uses for terminology. Capturing feedback properly and capturing it safely turned out to be the same decision. She writes more down now, and writes it more frankly, because she knows exactly where it lives and who cannot read it.

Over a year, the compound effect is striking. A new brief arrives from the medical-device maker: a document type she has not handled for them before. Before she translates a sentence, she spends twenty minutes in the entry — their style note, the last three query logs, the reviewer feedback from the previous autumn. She starts the job already knowing how they want warnings voiced, which terms regulatory has settled, and which ambiguity to query early rather than guess at late. That twenty minutes is invisible to the client and decisive for the delivery. It is also, she notes without sentimentality, the reason she can quote a realistic deadline while a translator starting cold would need longer and still deliver a first draft further from what the client will accept.

The log also changes how she handles her own uncertainty in front of clients. Because every answer she has ever been given is findable, she queries less often on repeat ground and more precisely on genuinely new ground — and clients can tell the difference. Her queries now routinely open by showing she has done the work: the segment, the two readings, a reference to how a neighbouring question was settled on an earlier job, and the specific decision she needs. Reviewers answer faster and more fully when a query respects their time in that way. The discipline is circular, in the best sense: a good log produces better questions, better questions produce clearer answers, and clearer answers make the log more valuable still.

Source Documents, PDFs and Scanned Pages Without Uploading Them

Every translator knows the quiet, grubby moment in a job when a small practical need collides with a large confidentiality duty. The client has sent a 140-page bundle: a main contract as a proper PDF, exhibits as scanned pages, an annex that is really three documents merged in the wrong order, and a reference agreement she needs to compare clause by clause. She needs to split one file, merge two others, compress a bundle before returning it through an agency portal with a size limit, and read the scanned exhibits as text so she can search them instead of squinting at images for four hours. The internet offers a dozen free tools for each of those tasks, bright and instant and utterly unsuitable, because every one of them begins with the same instruction: upload your file.

For documents under NDA, upload is disclosure. It does not matter that a free PDF site promises deletion after an hour, or that its marketing page uses the word “secure” in a reassuring font. The file has left her custody, crossed networks she does not control, and rested — however briefly — on a server whose retention, logging and jurisdiction she cannot verify. Her NDA with the law firm does not contain an exception for convenient websites. The medical-device maker’s supplier terms are even more explicit about subcontractors and processing, and a free web tool she chose alone is, in effect, a processor nobody approved. Lena’s rule for years was simply to refuse the convenience and do things the slow way. VaultBook’s built-in tools changed the trade-off: PDF merge, split and compress run inside the vault, locally, as part of the same single-file application that makes zero network requests. The slow way and the safe way stopped being different ways.

The practical sequence on a new bundle is now routine. First, before she even begins translating, she runs the File/Folder Analyzer over the client’s delivery folder. It gives her a clear inventory of what actually arrived: file names, sizes, types, the duplicate exhibit the agency sent twice under two names, the one file whose extension says PDF but whose content is a scan with no text layer at all. That inventory goes into the client entry as the arrival record. It sounds bureaucratic until the day a client queries whether a particular annex was ever sent — then the arrival audit, written on the day, settles the question in a sentence instead of an afternoon of forensic email-searching. The same tool earns its keep again at the other end of the job, when she audits the folder before deletion and can show herself, file by file, that what she is about to destroy matches what the retention instruction covers.

Scanned pages are where local OCR changes the texture of a working day. The exhibits that arrive as images — stamped, signed, sometimes photographed at a slight angle by someone at a client’s office — used to be second-class documents: readable by eye, invisible to search, impossible to quote from without retyping. With OCR running locally through tesseract inside VaultBook, the text of those pages becomes searchable in the vault without the pages ever leaving her machine. She can search the entire bundle, scans included, for a defined term, a date, a party name, a clause number. She can check whether “the Agreement” in exhibit three is the same agreement as the main contract or an earlier one it supersedes — a question that matters legally and that no amount of careful page-flipping answers as reliably as a search across OCR text. She is honest with herself about OCR’s limits: umlauts on a poor scan, a table that emerges scrambled, a handwritten marginal note that defeats the recogniser entirely. She proofreads against the image where it counts. But a searchable approximation she controls is worth infinitely more than a perfect conversion performed by a stranger’s server on a confidential file.

Attachment indexing extends the same principle across the whole entry. Source documents, reference PDFs, a client’s previous annual report, the regulatory guidance she keeps returning to — once indexed, their content participates in vault search alongside her own notes. A question that arises mid-translation — did this term appear in last year’s report, and how was it handled in the published English version? — becomes a search, not an excavation. Full search and semantic Q&A over indexed material mean she can interrogate a long reference corpus in the way she actually thinks: not “which file did I save that in” but “where did this client last explain their distribution model.” The answers lead her back to the document and the passage, and the document itself never had to be uploaded to a question-answering service to become askable.

For the long reading itself she uses the Reader, which matters more than it sounds. Translators do not read source documents the way other people read them. She reads a contract once for structure, once for obligations, once for the traps — the defined term used undefined in an annex, the date that conflicts between the recitals and the operative clause. Doing that inside the vault, with the document beside her notes rather than in a separate application, keeps one train of thought in one place. Observations made on the first read become notes linked to the client entry immediately, while the passage is in front of her, instead of sticky-note fragments that lose their context by the third read. A 90-page reference document stops being an interruption to the work and becomes part of the same workspace as the terminology and queries it will generate.

Compressing a return bundle illustrates how small the individual acts are and how large the cumulative risk would be if each were outsourced. Individually: a split here, a merge there, an OCR conversion, a compression before upload to the agency — each one trivial, each one a separate upload to a separate free service with separate terms. Together they would amount to a habit of sending an NDA client’s entire documentary life, piece by piece, to whichever website ranked first that morning. Doing them all inside the vault is not a dramatic security posture. It is simply the absence of a hundred small breaches of custody, repeated job after job, which is what real confidentiality mostly consists of.

She also keeps one strict separation inside the vault that the arrival audit makes possible: source material she received is never renamed, reorganised or “helpfully” tidied in place. Working copies are what she merges, splits and compresses; the received bundle stays exactly as delivered until its disposition date, so that any later question about what a document contained on arrival can be answered against the untouched original. The tools make copies freely because copies cost nothing; the discipline about which copy is the received one costs a moment’s attention and has settled two disputes she can name — one about a missing signature page, one about which version of an annex was current — without either becoming an argument about her filing.

Dictated Drafts, Interpreting Notes and Audio That Stays Home

Lena does not only type. On days when a German legal text is particularly knotted — subordinate clauses nested three deep, a sentence that runs for eleven lines — she often dictates her first English draft aloud, walking around the room, because speaking forces a sentence to become a sentence: it has to be sayable in one breath, in an order a reader can follow. She also interprets occasionally for two long-standing clients, short assignments she accepts selectively, and after each one she records a debrief note to herself while the terminology and the tricky moments are still fresh. Both habits produce audio. Both kinds of audio are, in their raw form, among the most confidential material she creates.

Dictation is unguarded in a way typed drafts are not. She thinks aloud in it. She says the client’s name, the parties’ names, the figure under negotiation. She tries a rendering, rejects it, tries a bolder one, wonders aloud whether the German could be read the other way and what that would mean for liability, laughs at a false friend, and leaves hesitations and half-sentences all over the recording. A typed draft is a product. A dictated draft is a mind at work on somebody else’s secrets, and it deserves to be treated as more sensitive than the final text, not less.

The obvious modern convenience — send the recording to a cloud transcription service and receive a tidy transcript in minutes — is therefore exactly wrong for her practice. The recording would leave her custody in its most revealing form, be processed on infrastructure she cannot inspect, and very likely be retained, logged or used under terms written for podcasters and meeting recordings, not for privileged legal work in progress. She used such a service once, early on, for a harmless practice exercise, and reading its data terms afterwards ended the experiment. The question was never whether the transcript was good. It was excellent. The question was what she had handed over to get it.

Local Whisper transcription inside VaultBook removed the dilemma rather than merely forbidding the convenience. She records her dictation, transcribes it on her own machine, and works with the transcript as a draft inside the client entry. The raw audio stays beside the transcript, playable in the inline audio player, so when the transcription flattens something important — a German proper name rendered phonetically, a number she spoke quickly — she can listen to the exact moment instead of guessing. The transcript is a starting text, not a record she trusts blindly: she corrects names, figures and negations with particular care, because a dropped “not” in a dictated liability clause is the kind of error that survives a casual read-through. Knowing the processing is local changes how freely she dictates. She says the names. She explores the alternative reading. The thinking-out-loud that makes dictation valuable no longer has to be self-censored for the sake of the tool that will hear it.

The MP3 cutter and joiner sound like humble utilities until an interpreting assignment shows their use. Reference audio accumulates around her work: a client’s recorded product briefing she was sent to prepare from, a pronunciation check she made for herself of names in a witness statement, her own debrief recordings that run long because she recorded the drive home as well as the debrief. Cutting lets her excerpt the ninety seconds that matter — the passage where the medical-device trainer pronounces the new component name and defines it — and file that excerpt in the entry where future-her will find it in context, instead of as an hour-long file she will never reopen. Joining lets her combine the two halves of a debrief recorded either side of a phone call, so the account of the assignment exists as one note, in order, rather than two fragments she must mentally stitch together months later.

Interpreting debriefs deserve a word of their own, because they are the notes translators most often fail to keep at all. After an assignment, the professional value is perishable: which terms the client’s own bilingual staff actually used in the room (sometimes not the glossary terms, and she needs to know that before the next session), where communication snagged, which speaker’s idiom she should prepare for, what she rendered in the moment that she would render better now. A voice debrief recorded immediately, transcribed locally, and corrected while memory is warm captures all of that in fifteen minutes. Filed against the client, labelled as interpreting rather than translation, it becomes preparation material for the next assignment and — handled carefully — terminology evidence: if the client’s team consistently says one English term aloud while their documents insist on another, that discrepancy is worth a tactful query, and she has the note to raise it from.

What unites dictation, debriefs and reference audio is the principle she now applies to every file type, learned first from PDFs and confirmed here: the more candid the material, the more important its custody. Final translations are polished, reviewed, and ultimately seen by many eyes. The audio behind them is seen by no one, was never meant to be, and contains more of the client’s reality per minute than any delivered document. A workflow that protects the deliverable while streaming the dictation to a cloud service has the sensitivity ranking exactly backwards. Hers, finally, has it the right way round — the roughest, frankest recordings she makes are the ones that never leave the room she made them in.

A final detail of the audio workflow is easy to overlook: she decides the fate of each recording at the moment she files it, not later. Dictation audio for a delivered job is working material, noted for deletion with the job’s other working copies once she is satisfied the transcript and final text are secure. A debrief note, by contrast, is practice knowledge — its transcript is edited to remove anything that should not persist, and the edited note is what she keeps, with the raw recording deleted on the schedule she recorded. Deciding at filing time means no recording lingers in a grey zone, kept by inertia, its status dependent on a future decision she might never consciously make.

Her photo and video material follows the same custody logic, on the rare jobs where it exists at all. A site-visit photograph from an interpreting assignment, or a short client video whose spoken content she must render in writing, is examined with the Photo/Video Explorer inside the vault — file details and EXIF information checked locally, frames reviewed without uploading a still to any external viewer. It is a minor part of her practice and she treats it as a telling one: even the file types she touches twice a year pass through the same rule, that client material is opened with tools that keep it home.

Every Deadline, Retention Date and Licence Renewal in One Sidebar

Ask a freelance translator what their project management system is and most will laugh, then describe a patchwork: deadlines in email subject lines, a wall calendar, a notebook list, a phone reminder for the one client who always needs delivery before a board meeting in another timezone. Lena’s patchwork worked until the month it didn’t — not a missed client deadline, she is careful to say, but a near-miss close enough to frighten her, when an agency job and a direct-client job both quietly converged on the same Friday because one due date lived in an email thread and the other lived in her head. The freelancer’s structural problem is not disorganisation. It is fragmentation: each client provides their own portal, their own thread, their own confirmation, and nobody — except the translator — ever sees the whole board.

VaultBook’s built-in calendar, together with the Recent, Due and Expiring sidebar, gives her the single view her working life otherwise lacks. Every dated commitment she records in a client entry lands in one place: delivery deadlines, staged deadlines for large jobs (terminology check, first half, full draft, review return), the date a query must be answered by before she proceeds on an assumption, invoice due dates, and the slower dates that are easiest to lose — NDA retention expiries, deletion deadlines, professional renewals. The sidebar’s value is that it does not ask her to go and look. Recent shows what she touched lately, Due shows what is approaching, Expiring shows what is about to lapse or must be acted on. A retention date recorded eight months ago resurfaces on its own, in the week it matters, instead of depending on her remembering that a spring job carried a six-month destruction clause.

Retention dates deserve emphasis because they are the deadlines clients rarely remind you about and auditors always ask about. Her NDAs and agency terms vary: some require deletion of source and target files a set period after delivery or payment, some require retention for a period for quality or legal reasons and then deletion, some distinguish between the deliverable — which the client obviously keeps — and her working copies, memories and drafts. Each instruction is recorded in the client entry at the start of the job, with its date computed and set as an expiring item, the clause reference noted beside it. When it surfaces in the sidebar, she acts on it as part of the end-of-job machinery rather than as a surprise. Before VaultBook, these dates were scattered through signed PDFs she would have had to reread one by one to reconstruct her own obligations. Now the obligations announce themselves.

The professional layer of her life sits in the same view, deliberately. Her professional membership renewal, her indemnity insurance renewal, the continuing professional development evidence deadline for her association, the renewal of a specialist legal-translation qualification she maintains — each is unglamorous, each is consequential if missed, and each used to live in a separate mental drawer from client work, which is precisely why such dates get missed by busy freelancers whose client deadlines are impeccably kept. In the sidebar they take their place alongside job deadlines as what they are: dated commitments of the same practice. The week her insurance and a major delivery coincide is visible in January, not discovered in the week itself, and she plans the delivery schedule around the renewal paperwork instead of colliding with it.

The Kanban board is her pipeline picture, and she keeps it ruthlessly current because its whole value is truthfulness at a glance. Columns run enquiry, quoted, in progress, delivered, invoiced, and finally retention clock — that last column is her own addition to the usual freelance sequence, and she considers it the most professional thing on the board. A card in “in progress” is a job. A card in “retention clock” is an obligation: the work is delivered and paid, but the card stays visible, ageing towards its deletion or archive date, so finished jobs do not vanish from her awareness at the moment the invoice clears. Enquiry and quoted columns do quieter work. An enquiry she quoted three weeks ago and never heard back from is a prompt for one polite follow-up, not a permanent open loop in her mind. A cluster of cards in “quoted” warns her that next week may suddenly become very full if they confirm together — a forecasting insight no single client’s portal could ever give her. She moves cards as part of the same small rituals that close a working day, and the board repays her with the thing freelancers most lack: a picture of the whole practice that is not held together by memory and anxiety.

Dynamic tabs solve the crunch-day problem that the calendar cannot. On a heavy day she may have three jobs genuinely open at once: the law firm contract in revision, the medical-device document awaiting a query answer (draft parked at the segment the answer will unlock), and a short agency job she is fitting between them. In a browser-and-folders workflow, each switch costs her reorientation — finding the files, the notes, the place she stopped. In the vault, each client’s work sits in its own tab inside the one application: the entry, its glossary notes, its style note, its document bundle, all where she left them. Switching clients becomes switching tabs, and switching back restores not just a document but a context. She is under no illusion that tooling makes three simultaneous deadlines comfortable. It makes them navigable without the low-grade chaos that used to generate errors — the style rule from the wrong client checked in the wrong window, the query drafted against the wrong segment — in the final hours of a long day.

There is also, tucked into this, the private analytics sidebar, which she uses in a modest, unobsessive way at month’s end. A simple view of her own activity — what she worked on, what the vault holds, how her records are distributed across clients — supports two practical acts: invoicing cross-checks (does every delivered card have its invoice facts recorded?) and a sanity check on the balance of her practice (is one agency quietly becoming half her income, a dependency she prefers to notice early?). It is her data about her work, computed locally for her eyes, and that is exactly the scale of analytics a one-person practice needs: enough to run the business deliberately, with none of it reported anywhere else.

Once a month she gives the board and sidebar a sterner review than the daily glance, and it takes fifteen minutes. Cards that have sat in “quoted” for a month are chased once or written off. Expiring items in the following quarter are checked for clashes — a cluster of deletion dates in the same week as a major delivery is rescheduled in her planning, doing the deletions properly a few days early where the clause permits, rather than rushed on the worst possible afternoon. That monthly pass is also when she notices the practice-level patterns no single job reveals: three retention periods negotiated shorter because she raised the subject at briefing, a new agency whose terms were markedly stricter and worth pricing accordingly, a licence renewal that will fall due during her planned holiday and must be handled before she leaves.

The End-of-Job Ritual: Deliver, Record, Delete on Purpose

Every profession has a closing ritual it performs well or badly, and translation’s is performed badly by default because the pressure of the next job arrives before the last job has been properly put away. Files sit in download folders. Queries remain answered only inside email threads. The glossary improvement she meant to record evaporates. And the source documents — the confidential heart of the job — linger on her machine for months beyond any need, not from carelessness but from the absence of a defined moment when their fate is decided. Lena’s end-of-job ritual in VaultBook exists to create that moment, and to make it short enough that she actually performs it on a tired Friday afternoon.

It begins within an hour of delivery, while everything is still warm. First, the final query answers are filed: any answer that arrived late, during review, sometimes in the client’s confirmation message itself, is entered into the query log in the standard format — question, answer, segment, consequence. Late answers are disproportionately valuable, because they are often the client’s most considered positions, given with the full text in view. Second, the glossary is updated from the job as delivered, not as drafted. The terms that count are the ones in the final version, including every change made in review — hers and the reviewer’s. Where review changed a term, the glossary note records the change and its reason, and version history preserves the earlier form beneath it, so the entry now tells the story of the term across the job: proposed, queried, approved, amended. Third, the style note absorbs any preference the job revealed that was not in it before. Ten minutes, typically. She describes this step to herself as being paid twice for the same job: once in the fee, once in knowledge that makes the next job for this client faster and better.

Invoice facts are noted at the same sitting, in the client entry, while the numbers are in front of her: purchase order or reference, word count basis as agreed, rate applied, delivery date, invoice number once issued. Freelance admin fails in the gaps between systems — the job in the CAT tool, the invoice in a spreadsheet, the agreement in an email — and queries about payment, which usually arrive six weeks later, used to mean spelunking through all three. Now the entry answers first: what was agreed, what was delivered, what was invoiced, when. When an agency’s payment run queries a rate, she replies in minutes with the reference they themselves issued. It is a small dignity, being able to answer your own business questions from your own records, and it compounds.

Then comes the part of the ritual that distinguishes a vault from a hoard: the disposition of the copies. For each job she records, in the client entry, what she holds and what will happen to it. The delivered target file: retained or not, under which provision. The source documents: the date from which the retention clock runs and the date deletion falls due, or an instruction to delete on delivery for the most sensitive matters. Working drafts, dictated audio and its transcripts, OCR text of scanned exhibits, the local translation memory export if she made one — each is named, because deletion performed on “the job folder, roughly” is deletion that cannot later be described honestly. The Kanban card moves to the retention-clock column, the expiry date sits in the sidebar, and the job leaves her active mind with its ending already scheduled.

When the deletion date arrives, deletion is an act, not a tidying impulse, and it is recorded like one. She performs it per the NDA or agency term that governs the job, then writes the record into the client entry: what was deleted, in categories precise enough to be meaningful — source files, target working copies, draft audio and transcripts, indexed attachments — the date of deletion, and the clause or instruction under which it was done. What survives deletion is chosen, not accidental, and the boundary is the professional core of the whole system. The client’s documents go. The client’s confidential content goes. What remains, encrypted in the entry, is her practice knowledge: the terminology decisions and their history, the style preferences, the query rulings stripped of the confidential text they arose from, the lessons about how this client reviews and what this domain demands. That residue is not the client’s property held past its time; it is a translator’s competence, made explicit. It is also, not coincidentally, what the client is actually buying when they return to her instead of briefing a stranger.

Version history gives this ritual its audit trail, and she has already had the occasion that proved its worth. A client contact, new in post, asked — reasonably, as part of a supplier review — what materials of theirs she currently held and how long she retained working copies. Practices that keep no records answer such questions with general assurances. Lena answered from the entries: for each recent job, the disposition recorded at closing, the deletion dates that had already fallen and been performed, the clause references, and the one job still inside its retention window, with its scheduled deletion date. The answer took her an afternoon to compile and evidently settled the matter entirely. Her own history in the vault — when entries were created, when glossary notes changed, when deletion records were written — corroborated the account without her having to assert her own reliability. The trail existed because the ritual had been performed job by job, not reconstructed for the occasion.

She is clear-eyed about what the ritual costs: perhaps twenty minutes a job, plus the deletion sittings the sidebar summons. Set against a career, she counts it as the cheapest professional insurance she owns — the difference between a practice that can demonstrate its confidentiality and one that can only declare it. And there is a quieter satisfaction she did not expect when she built the habit. A job closed properly feels finished. The entry stands complete: terms decided, questions answered, money recorded, copies accounted for, ending scheduled or performed. She closes the tab and starts the next piece of work with nothing trailing behind her — no orphaned files, no unrecorded answers, no date she is trusting herself to remember. For a one-person practice carrying other people’s secrets, that clean ending is not administration. It is the job, done whole.

She also protects access to the ritual itself as carefully as its contents: the vault’s password discipline, supported by its own password generator when she rotates credentials, means the audit trail she is so proud of cannot be read by anyone who picks up her laptop. A record of confidentiality that is itself unprotected would be a liability dressed as a virtue. Hers is encrypted entry by entry, which is why she can write deletion records, query answers and reviewer comments in full, frank detail — the candour that makes the records useful and the protection that makes the candour safe arriving, at last, in the same tool.

VaultBook Against the Note Apps Translators Actually Use

Comparison is easy to do badly. Every note app can list words, save a paragraph, and find it again later with search. If the comparison stopped there, Lena could pick almost any of them and get through the week. She could paste a glossary into any cloud notebook, type a style preference into any quick-capture box, and record a client’s answer to a query in any text field. The difference only appears when you ask where that practice memory lives while it waits for her, who else touches it on its way there and back, what happens to it when a device is offline, when an account is locked, when a vendor changes hands or changes its terms, and when a client with a non-disclosure agreement asks a direct question about handling.

There is a second test hidden inside the first, and translators feel it even when they do not name it. A practice memory must work at two speeds at once. In the middle of a job it must answer instantly — the approved term, the banned phrase, the name convention — without breaking the translator’s concentration on the sentence in front of her. Months later, in a completely different mood, it must answer forensically: not just what the decision was, but when it was taken, who confirmed it, what wording it replaced, and which documents it applied to. A sticky note can manage the first speed on a lucky day. A careless archive can sometimes manage the second after an hour of digging. A professional system has to manage both from the same record, because the instant answer and the forensic answer are the same fact viewed at different distances. Lena’s vault entries were built for that double duty: the term sits at the top where a working glance finds it, while the query history, the version trail and the linked style note sit in the same entry where a later audit finds them. When the comparisons below talk about search depth, labels, wiki links and version history, they are not listing ornaments. They are describing how the second speed is earned without slowing the first.

That is the test this article has applied from the start, because it is the test Lena’s work imposes. Her notes are not shopping lists. A glossary for a legal client carries the exact rendering the client approved for terms that carry liability if they drift. A query log records not just a word but a decision, the person who made it, the date it was made, and the sentence that prompted it. A style guide encodes tone choices a commercial client paid to have protected. None of that belongs to Lena in the casual sense; it belongs to the engagement, and she holds it on trust. Comparing VaultBook to Notion, Evernote, Microsoft OneNote, Google Keep and Apple Notes therefore means comparing custody models first, features second. Features matter, and each of those apps is stronger than VaultBook at something real, which this section will name plainly. Custody matters more for this reader.

Notion, and the workspace that is not hers alone

Notion is where many translators first build something that looks like a system. A page per client. Sub-pages for glossaries, for style, for queries. Toggles, databases, linked views. It feels organised, and in terms of surface organisation it is. Lena tried a version of that before she moved her practice memory local, and she understands the appeal because the structure is generous. You can shape a workspace around a client in an afternoon and invite someone into it in a minute.

The problem for her was never the shaping. It was the address. A Notion workspace lives on Notion’s servers, under Notion’s account system, reached through Notion’s application. Her client glossaries and query logs, once placed there, sit on vendor infrastructure for as long as they exist. That is the design, not a flaw in how she used it, and for many kinds of work it is an acceptable trade. For confidential per-client memory under NDA, the trade changes as soon as the vendor’s own processing chain widens beyond the vendor itself.

November 2026 made that chain visible in a way an abstract privacy policy never had. Notion published an advisory, updated on 10 November 2026, about an incident involving Fireworks AI, the third-party inference provider used for some Notion AI features. The final dataset, confirmed on 8 November, included Notion AI chat transcripts for Insights, and Notion told affected customers — fewer than one percent of workspaces — to treat their data as exposed. Lena was not writing this article to dramatise that event, and she noted the proportion carefully. Fewer than one percent is a small share. The point for a translator is not the share. It is the mechanism. When an AI feature sends workspace content to an external inference vendor for processing, the number of places confidential text can be reached grows. Transcripts that a user may have thought of as a conversation inside one app were present in a dataset accessed in an incident at a company the user may never have heard of before the advisory arrived.

VaultBook removes that whole chain for the material Lena places inside it, not by promising to choose better vendors, but by having no vendor chain to choose from. The application makes zero network requests. There is no sync server to hold her glossary, no inference endpoint to receive a query log when she searches it, no account backend that must be available for her to open a client entry on a train or in a room with the router switched off. Her semantic question-and-answer, her full-text search, her optical character recognition on a scanned instruction sheet, her transcription of a recorded client call — all of those run locally in the one file she opens. When she asks the vault a question about a client’s earlier decision, the text that answers it does not travel somewhere else to be understood. It is read where it lies.

Per-entry encryption sharpens the same difference. In Notion, protection depends on the account and the platform around the data. In VaultBook, Lena encrypts the sensitive entries themselves with AES-256-GCM, derived through PBKDF2 and executed through Web Crypto, so a client glossary can be locked as an object, not merely placed behind a login. The login model and the encryption model answer different threats, and for client work the second one is the one she can explain to a client without qualifiers.

The concession is real and she makes it without irritation. Notion is genuinely stronger for live collaboration. When an agency team wants to share a workspace, comment on a page in real time, assign a task inside a database, build from a rich template library, and watch each other’s edits land, Notion does that gracefully and VaultBook does not do it at all. If Lena’s week were built around a shared team workspace rather than a private, segregated practice memory, she would choose the tool built for sharing. Her week is not built that way. The segregation is the requirement, and a tool optimised for bringing people into the same page is optimised against the thing she most needs: keeping each client’s material separate, silent, and hers to guard.

Evernote, and the notebook behind an account

Evernote earned the loyalty of a generation of knowledge workers by being early, capable, and everywhere. Translators used it for terminology clippings, for reference articles, for the screenshot of a parallel text that might matter in six months. Its web clipper remains a small marvel. The model underneath, however, is a cloud notebook. Notes live on Evernote’s servers under the user’s account, synchronised to devices that sign in to that account. Access, continuity and recovery all run through that account identity.

For Lena, the dependence shows up in quiet moments rather than dramatic ones. A device that has not signed in does not have the notebook. An account problem becomes a notes problem. A change in how the service is offered becomes a decision she must respond to whether or not her clients’ deadlines care. The notes themselves are not files she holds in a folder she controls; they are records in a service she visits. When the material is a restaurant list, that distinction is trivial. When the material is a confidential commercial glossary that a client expects her to retain only for the life of the engagement and then delete, the distinction becomes a governance question. She needs to be able to say, truthfully, where the copy is, how many copies exist, and what happened when she removed it. A cloud-synced notebook makes each of those answers longer, because synchronisation is designed to spread copies helpfully across a service and its devices.

VaultBook answers with custody on her disk. The vault is one file. She knows its name, its folder, and which backup copies she made of it and when. She opens it without signing in to anything, because there is no account gate between her and her own practice memory. If she moves to a new laptop, she copies the file and opens it there. If she finishes a job and her retention obligation ends, she deletes the client entries and can describe exactly what she deleted because the system of record for that memory was never dispersed through a vendor’s sync graph. The local Whisper transcription matters here too: the dictation a project manager left about a sensitive product name can be turned into searchable text on her machine, instead of being uploaded to a transcription service as a condition of becoming useful.

None of that erases what Evernote does better. Its web clipper and its long-mature cross-device sync are stronger for capturing web research effortlessly. A translator trawling parallel texts, regulatory pages, and terminology forums all day will feel that strength honestly. Clipping a page with its formatting and source preserved, and finding it on another device moments later, is polished in Evernote in a way a local single-file tool does not attempt to imitate. Lena still does research on the open web, of course. She simply refuses to let the ease of capture decide the home for the confidential residue of a job. Research clippings can live in a capture tool. Approved client decisions live in the vault, where capture convenience is not allowed to outrank custody.

Microsoft OneNote, and the ecosystem notebook

OneNote has a different kind of strength: it belongs. Inside a Microsoft 365 workplace it sits beside Word, Outlook, Teams and OneDrive, and a translator working in-house or embedded with a corporate language department will find notebooks already provisioned, already shared through the organisation’s accounts, already governed by the organisation’s policies. As a freeform canvas it is generous. You can type anywhere, ink over a page, drag a scanned paragraph beside a handwritten scrawl, and build the sprawling, visual notebook that a terminology project sometimes wants to be.

The custody model follows the ecosystem. OneNote syncs through OneDrive and a Microsoft account. A notebook’s home is that synced cloud structure, and the local application works against it. For a corporate employee that may be exactly right, because the corporation is the data controller and its ecosystem is the approved place. For a freelancer who is a separate controller for each client’s material, borrowing the ecosystem’s home for her private practice memory mixes roles in a way that becomes hard to unwind. Her memory starts to live inside an account graph built for someone else’s organisation.

There is a practical worry layered over the governance one, and it comes from translators’ own reports about synced setups. When a sync engine constantly touches the folder where a working database lives, corruption risk follows the touching. Translators know this story from translation memory databases and termbases damaged by synced-folder behaviour — a file half-written, a conflict copy mistaken for the master, a database opened from a placeholder that had not fully arrived. Lena does not need to prove that OneNote itself will corrupt anything to take the design lesson seriously. Any system in which a background sync process has opinions about her working files adds a moving part between her and the file at the exact moment she can least afford ambiguity. She keeps her CAT tool’s data out of casually synced folders for the same reason.

VaultBook’s answer is almost stubbornly plain. The vault is a local file. No sync engine touches it unless Lena herself copies it, deliberately, to a backup location, at a moment she chooses, when the file is closed and whole. There is no placeholder state, no conflict copy with a device name appended, no background reconciliation deciding which version of a glossary deserves to survive. If two copies differ, it is because she copied at two different times, and the version history inside the vault gives her an internal record far more useful than a sync conflict. She would rather own the copying than rent the convenience.

The concession follows naturally. OneNote is stronger inside Microsoft 365 workplaces, where its integration is the point, and it is stronger for freeform handwriting and inking. A translator who annotates pages by hand, sketches a layout relationship, or works on a tablet with a pen as the main instrument will find OneNote’s canvas more natural than VaultBook’s structured entries, labels and wiki pages. Lena types her practice memory; she does not draw it. For the colleague whose notebook is genuinely a visual, inked workspace, the honest recommendation differs, and pretending otherwise would make the rest of this comparison worthless.

Google Keep, and the note that was only meant to be quick

Keep is the app Lena reaches for least in argument because it makes the smallest promise. It is quick notes, in the cloud, tied to a Google account. A colour, a checkbox list, a reminder. For a two-line capture on a phone it is admirably fast, and its labels are fine for the scale it intends. The trouble begins only when a tool built for ephemera is stretched into a practice memory, which translators do more often than they admit because the quick note is already on the phone when the client calls.

A per-client confidential glossary is not a quick note. It needs structure that survives growth: a label system that can keep one client’s entries cleanly apart from another’s without relying on colour memory, a wiki-like link between a term and the query that settled it, search deep enough to find a decision buried in a long entry from last spring, and protection proportionate to the content. Keep offers no per-note encryption. Its notes live in the Google account’s cloud, and its organising tools stay light by design. A term decision recorded there is a line in a stream. Six months and four clients later, the stream is the system, and the system cannot answer the question a careful client would ask: show me everything you hold for us, show me that it is protected as a unit, and show me that you can remove it as a unit.

VaultBook was built for exactly that unit thinking. Lena’s client material lives as entries, each carrying its labels, each linkable through the wiki to the style guide and query log that give it meaning. The entries that need locking are encrypted per entry with AES-256-GCM, not merely hidden behind the device’s unlock. Search runs across the whole practice memory, attachments are indexed so a scanned client instruction can be found by the words inside it, and the calendar with its Due and Expiring sidebar keeps the temporal obligations — the certificate that lapses, the framework agreement that renews, the retention deadline that approaches — visible beside the content they govern. Keep could hold the reminder. It cannot hold the record the reminder is about, at the standard the record requires.

The concession is speed and ubiquity. Keep is faster for a two-line capture on a phone, and it is already everywhere Android already is. No setup, no file to open, no decision about where a fleeting thought belongs. Lena does not sneer at that; fleeting thoughts are real, and a tool that catches them before they evaporate earns its place on a home screen. She draws the line at promotion. The moment a capture contains a client’s approved term, a confidential instruction, or a decision she will be accountable for later, it is transcribed into the vault the same day and removed from the quick stream. Convenience may catch the thought. It may not keep the record.

Apple Notes, and the beautiful walled library

Apple Notes is the strongest of the five on pure feel for a large group of users. It opens instantly, syncs through iCloud without visible effort, handles attachments and scans gracefully, and sits inside an ecosystem where the phone, tablet and laptop behave like one continuous desk. Locked notes exist, and for an all-Apple translator the library can feel both private and effortless. Lena has colleagues who keep immaculate practice notes there and she does not doubt their immaculacy.

Her objection is the wall around the garden and the home inside it. The library’s home is iCloud, under an Apple account, in an Apple-shaped structure whose export and portability are limited compared with holding a file. A translator whose working life crosses ecosystems feels that wall at specific joints: the CAT tool running on Windows, the agency platform that expects a particular desktop setup, the second machine that is not a Mac, the moment a client or an auditor wants the records in a form that travels. Moving a practice memory out of a service-shaped library is a project. Moving Lena’s vault is a copy operation, because the vault is one HTML file that runs wherever she opens it. The file does not ask what logo is on the laptop. It asks for a browser and her passphrase for the locked entries.

That portability is not a party trick. Over a freelance career, machines change faster than obligations do. A glossary approved in 2023 may be queried in 2026 on different hardware, under a different operating system, in a different country, during a week when the original device is in repair. Independence from an account and a server is what makes the memory’s survival a matter of her backup discipline rather than a vendor’s continuity, a subscription state, or an ecosystem’s migration path. She can open the same file, with the same entries, labels, wiki links and version history, and continue the professional conversation with the client as if the hardware had never changed — because for the memory itself, it has not.

The concession is integration, honestly given. Apple Notes is seamless and strong for all-Apple users, with excellent device integration: dictate on the phone, scan with the tablet, continue on the laptop, and the seams barely show. If a translator lives entirely inside that ecosystem, does not cross to Windows for translation work, and is comfortable with iCloud as the library’s home, the experience will be smoother than a deliberately independent file can be. Lena crosses the boundary every working day. Her CAT environment decides part of her hardware life for her, and a practice memory that only feels at home on one side of that boundary would be in the wrong place half the time.

Stepping back from the five, the pattern is consistent. Each leader wins on the axis it was built to win: collaboration, capture, ecosystem integration, instant quick notes, device harmony. VaultBook wins, for this reader, on the axes an NDA practice actually audits: custody with no upload, per-entry encryption of the confidential units, offline reliability with zero network requests, one-file portability across the machines a translator really uses, and no account or server standing between her and her own record. That is not a claim that the others are poor tools. It is a claim about fit, argued on mechanisms that can be checked, for a scenario in which the notes are other people’s secrets held on trust.

A Crunch Week, Run From the Vault

Theory is tidy. A crunch week is not, and Lena judges any system by the worst week it has survived rather than the calmest demonstration it can give. The week that settled her confidence began on a Monday with three deadlines already overlapping before the rush job arrived, which is to say it began like most of the weeks translators remember.

Monday opened with inventory, not translation. A commercial contract amendment for a long-standing legal client was due Wednesday. A set of marketing pages for a retail client, German into English, was due Thursday, with a style guide that forbade half the renderings a less careful translator would reach for first. A human resources handbook update, English into German, sat due Friday with a terminology history long enough that consistency mattered more than speed. Lena’s first act was the board. In VaultBook’s Kanban she moved the three jobs into their columns, checked the calendar, and let the Due and Expiring sidebar do the work it is quietly best at: the retail client’s framework note reminded her that an approved glossary review was pending, and a certificate related to the legal practice area was approaching its expiry window, not this week but soon enough that ignoring the warning would have been a choice. Nothing on that board translated a single sentence. Everything on it prevented the classic crunch-week failure, which is not slow translation but a forgotten obligation surfacing on deadline day.

She worked the contract amendment first because Wednesday was closest and legal text punishes haste most severely. The vault’s per-client entries gave her the approved renderings without a hunt through old email, and the query log for that client showed a decision from the previous quarter about a term the amendment used three times. She applied it, noted the application, and moved on. That is a small moment, and crunch weeks are built from small moments either saved or lost. Multiplied across a day, the difference between opening one file and excavating four systems is the difference between translating and archaeology.

Tuesday brought the rush job, because rush jobs have an instinct for arriving after the plan is set. An agency needed a certified translation of a short set of personal and corporate documents, urgently, with the formalities exactly right. Certified work adds a layer that has nothing to do with linguistic difficulty and everything to do with procedure: the statement, the format, the record of what was translated and when. Lena accepted it only after checking the board again — the vault makes the cost of a yes visible, because the other three commitments are sitting there in columns rather than in her optimistic memory. She slotted the certified job for Tuesday afternoon, pushed the handbook work into Wednesday morning’s early hours in her plan, and wrote the acceptance note to the agency from facts rather than hope. The record of the certified job went into its own entry that day, source documents indexed as attachments, so that months later, if anyone asked what exactly had been certified, the vault could answer without relying on a downloads folder and a prayer.

Wednesday was the day the week tried to break. Mid-morning, the legal client changed a key term. Not a minor synonym — a central defined term in the amendment, with a new approved rendering, effective immediately, because their own counsel had shifted position. In a scattered system this is the moment a translator’s stomach drops: the term occurs throughout a nearly finished text, earlier documents used the old rendering correctly for their time, and somewhere in a notebook is a note about why the old rendering had been chosen. Lena’s sequence was unglamorous and fast. She found every occurrence in her translation environment, applied the new rendering to the live job, and then opened the vault. Version history showed her the glossary entry as it had stood before, with the earlier approval recorded, so there was no confusion between what had been right on Monday and what was right now. She updated the entry, recorded the client’s instruction, its time, and its source in the query log, and added a wiki link between the term entry and that log note so the reason would travel with the term forever. When she delivered on Wednesday evening, she could explain the change in one paragraph, citing the instruction rather than her memory. The client did not need the paragraph. Having it ready changed how the delivery felt on both sides.

Thursday tested the other promise: no connectivity. Lena had to travel for a family obligation that could not move, and the retail marketing pages were due that day. She worked on the train, in the specific modern misery of a connection that appears, lies, and vanishes between stations. It made no difference to the vault. The file was on her laptop, the entries opened, the style guide’s prohibitions were searchable, the glossary was whole, and the local tools did not care whether the carriage had signal. She finished the pages offline, checked them against the client’s style entry a second time precisely because the rush to a deadline is when a forbidden rendering slips in, and sent the delivery when a real connection returned. Compare that, quietly, with any workflow whose memory lives in a cloud workspace: the train journey becomes a negotiation with sync states and cached copies, and the translator is left unsure whether the glossary she is reading is the current one or the last one the train’s first tunnel allowed her to fetch. Lena was sure. The file she had at departure was the file, entire.

Friday held two tests at once. The handbook update was due, and the agency from Tuesday’s certified job sent a query about a name rendering in a different, ongoing project — the kind of question that is trivial if the earlier decision is recorded and embarrassing if it is not. Lena answered it in minutes, from the vault’s query log, with the date of the original decision and the project manager who had confirmed it. Minutes matter here beyond politeness. An agency that gets a fast, sourced answer marks the translator as the one who knows, and that reputation, compounded over years, is a freelance practice’s real marketing. Only after sending that answer did she turn to the handbook’s final pass, where the vault’s semantic search earned its keep: she asked, in plain language, for earlier decisions about a cluster of training-related terms, and the answers surfaced entries she had written months apart, in different words, for the same client. Consistency across a long document is rarely a talent problem. It is a retrieval problem wearing a talent costume. She delivered the handbook on Friday evening, moved all four cards across the board, and performed the small closing entries for the week: what had changed, what had been decided, what would expire next.

One more detail from that week deserves recording because it shows the system working at the level of minutes rather than days. On Thursday evening, after the train journey and the retail delivery, the legal client sent a short follow-up about the amended contract: their counsel wanted confirmation that the new defined term had been applied consistently, including in a recital that referred back to an earlier agreement. Lena did not reopen the whole job in a panic. She checked the vault entry first, where the instruction, its time and its scope were already written in her own words from Wednesday, and then checked the delivered text against that recorded scope. The recital fell outside the amendment but referred to the earlier agreement’s terminology, and her note made clear why leaving its wording aligned with the original agreement was the defensible choice rather than an oversight. She replied that evening, explaining the distinction in a few calm sentences. Without the midweek note, that reply would have waited until Friday, required a reconstruction of Wednesday’s reasoning under Friday’s fatigue, and might still have come out hesitant. With it, the answer was a retrieval. Crunch weeks create dozens of these small evidential moments, and they are rarely dramatic enough to be remembered individually. Together, they are the difference between a translator who sounds certain because she is guessing well and one who sounds certain because her records are doing their job underneath her.

Notice what the vault never did that week. It never translated. The CAT tool remained the system of record for translation memory and the live bilingual work, exactly as it should. What VaultBook held was everything around the translation that determines whether the translation is right, defensible, and on time: the decisions, the glossaries, the prohibitions, the deadlines, the records. A crunch week does not usually fail at the level of language. It fails at the level of memory under pressure. That is the level the vault is for.

When Things Go Wrong: Theft, Crash and the Client Who Asks Questions

A system for confidential material should be described by its bad days, not its good ones. Lena plans for three, because each has either happened to her or happened close enough to her that treating it as hypothetical would be vanity.

The first is theft. A laptop bag taken from a car, a machine lifted from a co-working table during a coffee run, a burglary at a home office — the variations matter less than the common core, which is that hardware containing client material is suddenly in someone else’s hands. This is where per-entry encryption stops being a specification and becomes the whole story. Lena’s most sensitive client entries are encrypted individually, each protected by AES-256-GCM with keys derived through PBKDF2 in Web Crypto. A thief who powers on the machine faces the device’s own protections; a thief who somehow reaches the vault file faces a second, independent fact: the confidential entries are ciphertext without her passphrase. The distinction between a file sitting behind a login and a file whose sensitive contents are mathematically sealed is the distinction she would want explained to her if the roles were reversed and a translator held her company’s documents. She would still have obligations after a theft — notify affected clients promptly, describe what was held, cooperate with their assessment — and encryption does not abolish those duties. What it abolishes is the worst sentence in that conversation. She would not have to say the material was readable. She could say what was taken, in what form, and why that form matters, and the conversation could proceed as an incident to manage rather than a breach to confess. She keeps the passphrase out of the machine’s own saved items, because an encryption system unlocked by a credential stored beside it is theatre, and she knows the difference.

The second bad day is quieter: the crash. Translation environments fail, termbases corrupt, a software update arrives the night before a deadline and rearranges something essential. It has happened to Lena in the specific form translators dread, a termbase that would not open on the morning it was needed, and the recovery taught her the value of redundancy that is not a copy of the same fragile thing. Her CAT tool remains the system of record for translation memory, and its proper backups are her responsibility there. But the vault gave her a second, independent path back to working order. The client glossaries she maintains as vault entries — the approved terms, the contexts, the prohibitions, the decisions behind them — could be read, searched and exported into the rebuilding work immediately, because they lived in a different file, in a different application, untouched by the CAT failure. A glossary rebuilt from her practice memory is not identical to a termbase restored from its own backup, and she would not claim it is; the vault version carries the reasoning and the client confirmations, which is sometimes more valuable in a dispute about a term than the bare pairing itself. The deeper lesson shaped her habits: the memory that explains the work should never live only inside the tool that does the work, because tools fail in correlated ways at the worst times. Separation is a backup strategy in its own right.

The third bad day arrives as a polite email. A client, often a larger one with its own compliance function, asks what personal and confidential data Lena holds about their organisation, where it is held, how it is protected, and how long she keeps it. An auditor’s version of the same question is less polite and more structured, but the substance is identical. This is the moment a scattered practice dreads, because the honest answer requires an excavation: some material in email, some in a note app, some in folders named by date, some in the CAT environment, and the retention story is whatever the excavation finds. Lena’s answer is built as she works rather than assembled in a panic. Her per-client entries define what she holds. Her end-of-job deletion ritual defines how long she holds it: when an engagement closes and its retention period runs, the client entries are deleted, and the deletion is a deliberate act in the same system that held them, not a hope that a sync service eventually forgets everywhere at once. Her calendar and the Due and Expiring warnings keep retention dates from passing unnoticed. So the reply to the client is short, factual and checkable: these categories, in this form, encrypted this way, retained until this point, then deleted. Clients with compliance functions do not need a translator to be a large company. They need her answers to be as organised as their questions. The vault is how a one-person practice gives a one-page answer with a straight face.

There is a fourth, less cinematic failure she plans for as well, because it is the one most likely to be self-inflicted: her own mistake. A glossary entry edited in haste and saved wrongly. A label applied to the wrong client’s entry during a tired evening. A paragraph of working notes pasted over a paragraph she meant to keep. Cloud systems advertise version recovery as a safety net for exactly these moments, and Lena would be dishonest to pretend a local tool is excused from the same standard. This is where version history inside the vault earns its place on ordinary days rather than in disasters. When she caught, during the handbook job, that she had overwritten a client’s earlier instruction about training terminology while recording a newer one, she did not have to choose between the two or reconstruct the older wording from memory. The history held both, in order, and she restored the earlier instruction as a dated layer beneath the newer decision, so the record showed evolution instead of pretending the first instruction had never existed. That distinction matters to clients more than translators sometimes realise. A client who refines an instruction does not want a supplier whose records silently rewrite the past; they want the change visible, dated and attributable, so their own internal reviewers can follow it. A memory system that preserves its own corrections is, in that sense, behaving like a professional witness rather than a notebook.

Drive failure closes the list, and it carries the honest limit this article owes its reader. VaultBook has no cloud standing behind it, which means there is no vendor quietly keeping a spare copy of Lena’s vault in a data centre. If her drive dies and she has made no copies, the vault dies with the drive, and all the custody arguments in the world will not resurrect it. Her protection is a discipline, not a service: the vault is one file, so backing it up is copying one file, and she copies it to two separate places on a rhythm tied to her working week, with one copy kept away from the laptop’s usual location. The one-file design makes the discipline easy to keep — there is no database folder to assemble, no export ritual to postpone — but it does not perform the discipline for her. She regards that trade as fair and even healthy. A cloud backup she never thinks about would also be a cloud copy she would have to explain to clients. Her copies are hers, in places she can name, encrypted in the same sealed form as the original where the sensitive entries are concerned. The risk is not hidden from her. It is assigned to her, visibly, and she manages it the way she manages terminology: on a schedule, in writing, before it becomes urgent.

What VaultBook Will Not Do for a Translator

Honesty about limits is not a courtesy in a comparison like this one. It is the evidence that the strengths claimed elsewhere were measured rather than imagined. So, plainly: VaultBook is not a CAT tool. It does not hold translation memories in the working sense, does not leverage previous segments against a new source text, does not perform the segment-by-segment bilingual editing that is the centre of Lena’s production day, and it should never be asked to. Her CAT environment remains the system of record for that work, with its own termbases, its own quality checks, and its own project files.

It offers no machine translation. It will not draft a first pass of a commercial text, will not post-edit, will not connect her to an engine. It has no client portal: a client cannot log in, upload a job, approve a term in place, or watch a project’s progress through it, and any translator whose clients demand a portal will need that portal from another system entirely. It provides no live collaboration, no shared cursors, no simultaneous editing, no commenting thread in which an agency team debates a rendering in real time. Colleagues who need those things — and some genuinely do — should use the tools built for them, as the concessions earlier in this article said without hedging.

What remains, once those absences are listed, is not a remainder. It is a definition. VaultBook is the private practice memory that sits beside the production systems: the encrypted, searchable, offline record of what each client decided, prefers, forbids and approved, kept in one portable file under the translator’s own custody. Judging it for not translating would be like judging a termbase for not invoicing. The categories are adjacent, not competing, and a practice is stronger when each system is allowed to be excellent at its own layer instead of being stretched, apologetically, across all of them.

The Practice Memory Is the Business

Strip a freelance translator’s practice down to what cannot be replaced quickly and surprisingly little of it is equipment. A new laptop can be bought in a day. Software can be reinstalled in an afternoon. Even language skill, in the raw sense, exists in other people a client could hire by Friday. What cannot be bought, borrowed or rebuilt on demand is the accumulated per-client knowledge: which rendering the legal client approved and why, which phrase the retail client banned after a campaign went wrong, which project manager confirmed the name convention and on what date, which handbook terms carry a history that makes consistency a matter of trust rather than style. That memory, compounded over years, is why clients return to a particular translator instead of re-tendering every job to the cheapest bidder. It is, in the most literal sense, the business.

Holding that asset carelessly is therefore not a housekeeping lapse. It is a professional failure with a client attached to it. To keep other companies’ confidential decisions in a cloud stream of quick notes, or scattered across inboxes and downloads folders, or inside a workspace whose contents can be routed to vendors the client has never heard of, is to treat the most valuable and most sensitive thing in the practice as if it were the least important. Lena’s conclusion, reached through the scattered years and sharpened by one November advisory about transcripts in someone else’s dataset, is that the standard is simple to state: the practice memory should be encrypted where it rests, available where she works including the places networks do not reach, searchable deeply enough to answer in minutes, and portable enough to survive every machine and ecosystem change a career will contain. It should be hers — not lent back to her by an account, not hosted for her by a server, not contingent on a vendor’s roadmap.

VaultBook meets that standard in the form she can maintain on her busiest week: one file, opened locally, zero network requests, per-entry AES-256-GCM for the material that needs sealing, labels and wiki and search and version history for the material that needs understanding, a calendar that watches her obligations approach, and a deletion ritual that lets her keep her word when a client’s retention clock runs out. None of that makes her a better linguist. It makes her the kind of professional a careful client can trust with the next confidential job, and the one after that — which is to say, it protects the exact asset her livelihood is built on. The practice memory is the business. Keep it like it matters, because to the clients who entrusted it to her, it does.