Release Notes

Release 3.0.0-rc.16 (October 2026), a release candidate of Silver Bullet 3 for the desktop, for macOS (Apple silicon and Intel), Windows (x64; on Windows 11 on Arm through Windows’ x64 emulation) and Linux (x64 and arm64). Every package is signed with the release key (fingerprint 559F-A83E-A382-315B) and comes with its software bill of materials (SBOM), vulnerability report and third-party notices.

Silver Bullet 3

Silver Bullet gives plain-language answers from your own files and live databases, on your computer. You ask a question in the chat; Silver Bullet reads it, works out the answer from your data and shows it as a headline, a chart and a table. It learns from every correction you make, without fine-tuning any AI model.

• One program, one person, one computer. Double-click it: your browser opens on Silver Bullet, already signed in. There is no window to keep open, only an icon in the menu bar (macOS) or the notification area (Windows, Linux). Silver Bullet answers on this computer only (127.0.0.1), never from the network.
• Your data stays with you. Files are loaded into a database inside Silver Bullet’s data folder. Databases are queried where they are, read-only, through several safety layers. The AI model (local by default) never sees your rows or results, nor the values of columns marked as personal data.
• A 14-day trial, then a licence key for this computer.

Silver Bullet 3 replaces Silver Bullet Suite 2. It is a new product, not an upgrade: Suite 2’s settings and data are not imported (see Moving from Suite 2).

Asking and answering

• A home page with question cards chosen from your data (six to eight, worded with your own column names and dates: What share was cancelled?, Best and worst month, What’s overdue?…) and Your brief: the main figure, what needs attention, what is growing, the top performer. Both are worked out in advance, so the page opens at once; each has Dig in and Why?.
• Four ways to ask, one engine: tap a card or a chip, Build it (measure, grouping, period, filter from drop-downs), type with type-ahead (columns with sample values, ready-made questions), or shortcuts (/top, /bottom, /trend, /compare, /share, /list, @column).
• The answer card: “How I read your question” chips (change the period, measure, grouping, filter, top N or chart with a tap, and the answer is worked out again without asking the AI model); a one-sentence headline; the chart chosen for the kind of question, or the one you ask for by name (bar, line, pie, doughnut, spider, histogram, scatter, bubble, waterfall…); a table view with totals; Indian formats (₹, lakh, crore, FY April to March); How I got this; Ask next; Read it to me.
• “No, fix it” in three taps: what looks wrong, alternatives worked out with real numbers, then the fixed answer.
• Saved answers (run again on the latest data), boards (pin answers as numbers, charts or tables, arrange, refresh, download as PDF), Copy table, Copy chart and Download picture.
• Insights: what changed and what moved it, the biggest movers, how concentrated the total is, runs of months and unusual months. Found by fixed checks on the figures, not by an AI model.
• Executive briefs: templates (Monthly sales review, Collections, Board pack), any answer added with Add to → A brief, a three-pane editor, and downloads as PDF, PowerPoint and Word with real, editable charts. The words are written from the figures, never by an AI model.
• Voice input is off until you turn it on (Admin → AI model).
• Errors show their code (SB3-…), explained in the Troubleshooting Guide.

Learning from corrections

See How Silver Bullet Learns. Questions are read into explicit choices first. A correction from “No, fix it” is in use for you at once; when you check it under Your corrections (after a safety replay of every verified question and recent questions, and a model check that asks the AI model again wherever a fix would change what it is told) it becomes a verified answer, reused for the same question in other words and shown with a Verified badge.

• Write the right query yourself (or have your DBA write it): Ask an expert (my DBA will write it) in “No, fix it” puts the question in Your corrections with an SQL editor beside the source’s tables. Run goes through the same safety check, read-only login and limits as every question; the result is shown side by side with Silver Bullet’s.
• The right lesson, not just one answer: Silver Bullet proposes what kind of fix a query is (a business measure, a rule, a saved view, the right reading, or the query itself), tries it on the question and checks it with the safety replay. Expert answers and saved views are checked every night; one that fails is marked Needs attention and its answers say Please check. Parts of a query you mark as changeable (the period, a value, a top N) let “the same for last month” reuse it safely.
• “Please check”: an answer the AI model read is flagged, instead of shown as confident, when it shows a sign of a silent mistake: a calculation asked for but one column added up, AND and OR mixed without parentheses, a share of exactly 0% or 100%, or a condition on a value the data does not have.

Your data sources

• Add files in four steps (Choose files, Check data, Review, Add), with one screen per decision, Undo and Remember these choices: Excel, CSV, TSV, Parquet and text. Dates with two-digit years, day-first and month-first dates, Indian number grouping, rupee signs and European decimal commas are read correctly or asked about; postal codes stay text; row counters and columns without a heading are left out by default.
• Add a database in five steps (Pick type, Connect with a five-check connection test, Choose tables, Name things, Done): PostgreSQL, MySQL, MariaDB, SQL Server and Oracle, each shown with its own logo. Create a read-only login on the form gives the SQL to run yourself, or send to whoever manages the database, filled in from the form, with Copy SQL, Save as a file and the address of this computer the database must accept.
• The data dictionary: plain-sentence rules, business measures (including one total divided by another, Amount ÷ Qty), the main measure, grouping and date (guessed when a file is loaded, marked guessed, and changed there), what each column is, Try a question and Publish.

Your AI model

• The default is Ollama on this computer with qwen3-coder-next, and nomic-embed-text to match questions with verified answers by meaning. A cloud model (Google Gemini and others) can be chosen under Admin → AI model, with its key.
• No AI model is a choice too: no model is ever called and nothing is sent anywhere. Silver Bullet then answers what its built-in reader reads (the home cards and questions said like them), and says so politely for any other question (SB3-4004).
• A model that cannot be used is shown before you ask, with three ways out: set up a cloud model, install Ollama, or use no AI model.
• With Google Gemini, Gemini’s own embedding model matches questions with verified answers (39 of 39 paraphrases matched in our measurement).

The admin area

Admin on the desktop: Data sources, Data dictionary, Your corrections, AI model, Health and logs (with the support bundle and automatic backups: every 24 hours, the newest 7 kept, Back up now), Components, Licence, Network and firewall, and Activity & trust. Search settings in the admin header finds any setting. About Silver Bullet in the account menu shows the version and how it is licensed.

New in 3.0.0-rc.16

30 October 2026. The same release is 3.0.0-intl-rc.9.

A security review before 3.0.0. An independent review found 23 issues; all are fixed, each with tests that try the original attack and variants of it (Security and Privacy guide; the list of tests is in docs/SECURITY-TESTS.md). What you may notice:

• Licence texts 1.1 are asked for once again. The EULA and the Terms now say that packages are signed with Govirang’s release key only, by design (EULA clause 10A), and that a licence moved to another computer stops working on the old one 7 days after the move. The Privacy Notice describes the daily check of a key bought online exactly as the program makes it.
• Signed packages, checked before anything is replaced. SilverBullet verify-package checks a package’s signature, its files, its edition, platform and release line, and that it is not older than what is installed; the install command and the upgrade use it. A package that holds a link, or files added after it was signed, is refused.
• Licences and the trial: a key works only on the computer it was issued for; a licence moved to another computer stops on the old one; a computer whose own id cannot be read is not licensed (SB3-6011, see the Troubleshooting Guide). The trial is per computer account.
• The desktop answers only the browser tab it signed in, and the sign-in link no longer appears in the list of running programs.
• Personal data: masking holds whatever names a query gives its tables or columns, board tiles are masked for every viewer, and a short value (or anything worked out from a personal column) is hidden whole. A database’s error message shown to anyone but an administrator no longer contains values from the data.
• Files: a file that unpacks into far more than its size, a Parquet file of more than 4,096 columns, very long numbers and oversized form fields are refused at once. A slow question about files no longer holds up other people: each person, and all background work together, has a share of the processors and memory, and a question that waits too long is told “busy”.
• Secrets stay out of logs: passwords, keys and personal values are removed from the log, the syslog stream and the support bundle in many more forms, and an AI provider’s error never keeps any spelling of the API key. A stored password or key is never sent to a new address.
• Backups and restore never follow links, and a restore checks the backup it copied, not the original.

Fixed on Windows: backups (and so upgrades, which back up first) failed on Windows with “being used by another process”. The trial record of an account in the Administrators group was not read.

Charts name both axes, with an axis line (design rule 7). The Administration Guide has a new Limits section: the most of each thing, what has no limit, and how to change each. Board PDFs print the minus sign and the change arrows.

Upgrading from an earlier release candidate: the upgrade now checks the release line of each package, which earlier release candidates cannot do for this one. Install 3.0.0-rc.16 with the install command or from the zip; later upgrades work as before.

New in 3.0.0-rc.15

29 October 2026. The same release is 3.0.0-intl-rc.8.

Getting Silver Bullet (Installation Guide, Getting Silver Bullet):

• Download it from silverbullet.govirang.com, with no account. The page picks the package for your computer and shows each package’s SHA-256 and the release key.
• Or install it with one command, which downloads the newest package, checks its SHA-256 and installs it (no administrator rights): curl -fsSL https://silverbullet.govirang.com/install.sh | sh on macOS and Linux, irm https://silverbullet.govirang.com/install.ps1 | iex in PowerShell on Windows. Run it again to update. On Windows it asks you to quit Silver Bullet first if it is running.
• The guides, the licence agreement, the terms and the privacy notice are also on the website, under Guides.

Support and company details:

• Licence and payment e-mail goes to silverbullet-license-support@govirang.com: the Licence page’s key request, Buy a licence, About and the legal texts. Help with the product is community support, as before; there is no telephone support.
• The Terms of Service, EULA and Privacy Notice now give Govirang Microspaces Private Limited’s CIN (U58200MH2024PTC424124) and registered office (Thane, Maharashtra). Nothing else in them changed, so they keep version 1.0 and are not asked for again.

Studio fixes:

• An empty board no longer scores 15 of 15. With no tile there is nothing to check: the badge says Design check: add a tile, and the Design check tab says so until the first tile is added.
• The side panel keeps its width while a tile is built (it used to widen with the builder’s longer choices and push the board aside).
• In a narrow window, the starters are on the first screen. Who is it for? is one row of buttons with the chosen one’s description below, instead of a tall list that pushed the starter boards out of view.
• In a narrow window, the top bar no longer overlaps itself. It had more buttons than room, so the data picker ran into Recent. Saved answers, Studio, Insights and Help are now in the account menu in a narrow window (under 700 pixels wide, such as a browser window snapped to a third of the screen), with Boards and Admin.

New in 3.0.0-rc.14

28 October 2026. The same release is 3.0.0-intl-rc.7.

Before you start: the licence agreement, terms and privacy notice.

• The first start now shows the End User Licence Agreement, the Terms of Service and the Privacy Notice, and nothing works until they are accepted (a tick and Agree and continue), before the licence page and the welcome. They are asked for again only when one of them changes. The acceptance is kept on your computer. About links to each document and shows the date you agreed, and says what support covers: licence and payment questions by e-mail; help with the product from the community.

Buy a licence from inside the app (Administration Guide, Buying a licence):

• No account to create. Buy a licence on the Licence page (and on the trial banner and the Licence needed page) asks Monthly or Yearly, shows the price for your region and this computer’s code, and opens the payment page of Dodo Payments, which takes the payment, charges the tax and sends the invoice. A QR code lets you pay from a phone.
• The key arrives by itself. Once the payment goes through, the key for this computer is installed within a minute and e-mailed to you. A monthly subscription’s next key is fetched each month by itself. Both can be switched off; a key can always be pasted in by hand, as before.
• Prices per month (yearly = 10 months): India ₹250 + GST; the Americas, the Middle East and North Africa and the rest of the world US$7 + tax; Europe €8 + tax; the UK £8 + tax; Singapore S$10 + tax; Australia A$10 + tax.
• Lost key, new computer: Get my key and Move to this computer on the Licence page, with the order number and e-mail.
• No refunds once the key is delivered. A key that never arrives: e-mail the proof of payment; once the payment is confirmed, the key is e-mailed. Support covers licences and payments, by e-mail.
• What is sent: the computer’s code, the app’s version and, for a subscription, the current key, to the licensing service only (listed on the Network page). Nothing about your data.
• New error codes SB3-6006 to SB3-6010.
• Buying opens once the licensing service is live. Until then Buy a licence says how to get a key by e-mail.

Help inside the app:

• Every guide is built in, with a search across all of them, and opens in a new tab: ? in the header, Help in the account menu, Guides in About. It works without the internet and before a licence.
• Pages link to their part of the guide, and an error’s SB3 code links to its entry in the Troubleshooting Guide (What this means).
• The licence agreement, terms and privacy notice are in it too.

Studio fixes:

• Fix all never removes a tile. A copy of a tile (rule 14) is removed only by its own fix; two tiles that show the same question as a different chart or a different top N are no longer called copies.
• A board has at most 10 pages; the design check uses the board’s own settings; Others stays in a table cut to 50 rows and is left out when one of the shown figures is hidden from you.
• A chart tile can be any chart. With Show it as: Chart, a Chart list offers the charts the chat’s Chart chip offers: Usual chart (the default), bar, line, pie, doughnut, spider, histogram, scatter and bubble. The chart is kept with the tile and drawn so in the preview, on the board, and in the PDF, PowerPoint and Word files. When an answer cannot be drawn that way, the preview says why and keeps its usual chart; the design check reads the chart drawn, and Done for you no longer praises a chart the check flags.
• The data is a dropdown on Studio’s first page, as in the Build tab. With more than one source it says how many there are; an administrator is also told how many are switched off, or where to add more.
• Open sits at the foot of every starter card, in line across a row.
• A board with no tiles reads the data picked (it waited for a tile before); its empty board says how to add one from Studio. Period, Compare with and Who is it for? sit above their dropdowns, so long choices are not cut short; the fields’ hint is no longer in capitals; The grid has 1 empty cells now reads The grid on page 1 has empty cells to fill (1).

Others in a Top or Bottom answer (User Guide, Build it):

• The rest is shown, not left out. A top (or bottom) few of a total or a number of rows now ends with one more bar and one more table row, Others: everything outside the few, added up. Ask for the top 3 regions and you see the three, then Others for the rest.
• The Total is the whole. The bars add up to the whole, and the table’s Total is the total of everything, not only of the few shown. Before this release the Total of a Top answer added up only the rows on screen, so a Top answer’s Total is now larger than before. The figures of the few themselves are unchanged.
• Always last, always grey. Others comes last whatever its size, even when it is bigger than every bar, and it is never the bar in colour: the headline is never about it.
• Everywhere the answer goes: the chat, Copy table, Studio tiles, boards, and their PDF, PowerPoint and Word files. A board’s PDF shows at most 8 bars in a tile; when there are more, Others keeps the last place.
• When there is no Others:
  • the figure is an average, a highest or lowest value, a number of different values, a figure per something, or a business measure from the data dictionary (Amount ÷ Qty): the rest of an average is not the whole less the part, so these show only the few;
  • fewer rows came back than you asked for (the top 5 of 4 regions): nothing is left outside them;
  • the rest adds up to zero, or cannot be shown to you (personal data);
  • the answer is drawn as a line: a line does not run through “everything else”. A share (and a top few asked for as a pie) still shows the rest as one slice, as before.
• Safe as every answer. The rest is worked out by the same checked, read-only path as the answer itself, as the same person, with the same rules and masks. It is one more small query to the source when the few are full.
• Studio’s design check does not count Others as one more bar: a top 15 with its Others still meets rule 15 (Pick the right charts).
• The Test Kit’s correct answers are unchanged. Four of its questions (Q06, Q07, Q21 and Q35, all a top few of a total or a count) now show Others in their chart and table; their headlines and rows are the same.

The guides:

• Getting Started has a fifth step in its five-minute tour, Open a ready-made dashboard in Studio, and lists Studio among the buttons along the top.
• The Troubleshooting Guide has a Studio section: no Studio at the top, not enough to make a starter, a preview that cannot be worked out, Not filtered by … on a tile, a design check score that went down, and board files that will not download.
• SB3-1016 (Too many of these in the last hour) now says what it covers: a board’s PDF, PowerPoint and Word files share one limit of 60 an hour per person. It used to name the PDF only.
• The guides’ descriptions (on the website and in the README) now list Studio and the 15 dashboard rules.

New in 3.0.0-rc.13

Studio: build a dashboard yourself, and it follows the 15 dashboard rules for you (User Guide, Studio and Designing a good dashboard).

• A Studio tab next to Boards: the controls on the left (the data, its fields, the I want to… verbs, and settings for the whole board), the dashboard on the right.
• Starter dashboards made from your data: Sales overview, Products, Customers, Regions and This month, worded with your own names, for a Business owner, Sales head, Finance or Operations. Each fits one screen and starts at Design check: 15 of 15.
• A live preview: while you pick, the tile shows on the board, worked out but not kept, with Done for you: the rules applied to it, in plain words.
• The design check: the 15 rules published by SQLBI, in our words. A rule the board breaks is shown on its tile with a one-click fix (move to page 2, tidy up, sorted bars instead of a crowded pie, add the change, the top 5 instead of a long table, remove a copy…), or Fix all.
• The whole board: one period every tile follows, a comparison with the period before or the same period last year, and page names.
• Export as PDF, PowerPoint or Word, or Make a brief in Insights from the dashboard.
• Ways in: Open in Studio under any answer and on any board; Build a dashboard in Studio beside Build it. Build it itself is unchanged.

Charts everywhere (the chat, boards, briefs and their files) now follow the rules too: the bar the headline is about is in colour and the rest in grey; a line’s unusual points are ringed; a line or waterfall whose axis does not start at zero shows a break on the axis; and number tiles show their change. The chart colours are tested for colour-blindness; the dark theme’s “last period” colour is a shade darker to stay apart from “this period”.

New in 3.0.0-rc.12

Build your question is rebuilt for people who are not analysts (User Guide, Build it):

• Pick what you want to do: Total, Average, Count, Top, Bottom, List rows, Share, Trend or Compare, then fill in a sentence of drop-downs. A number can be the total, the average, the highest or the lowest; a count can be of rows or of different values (different customers).
• Where…: up to three conditions — is, is not, more than, at least, less than, at most, between, contains, starts with, and for dates on, before, after. A text column suggests its values as you type.
• No dead ends. Data without dates offers no period (it used to offer “last quarter” and answer with an error), and Trend and Compare say “needs dates”. Only the periods your data reaches are offered, starting from the latest financial year with data. A price, rate, rating or percentage starts as an average, not a total. A column with a value for nearly every row (an ID) is greyed out under split by: use Top or List rows. Show me the answer says what is missing until the sentence is complete.
• Start from: five ready-made questions from your data, one tap each.
• Edit in builder under every answer: its question back as the sentence, to change one word and ask again.
• More periods: the last 7, 30 and 90 days, and this and last calendar year.

A built question still never waits for the AI model.

New in 3.0.0-rc.11

• Pin to a new board from the chat works on the desktop. Pin to board → a new board was refused with “this is for a team: it is in the server edition”. The board is now made, and the chart is on it.

New in 3.0.0-rc.10

A security release:

• A database account that can change data is always refused. The setting that allowed one (Allow an account that can change data) is gone: Silver Bullet only reads, and Create a read-only login on the data source’s page writes the script for your database. A source that used the setting stops answering until it uses a read-only login; its message says what to do.
• Opening the app on a Mac the first time: the guides now give the way that works on macOS 15 Sequoia and later (System Settings → Privacy & Security → Open Anyway).
• More checks against hostile input: the SQL check and the reading of uploaded files are tested with millions of generated inputs, and the benchmark functions built into the file engine are proven unreachable.

New in 3.0.0-rc.9

Nothing used is hidden any more, and nothing removed is lost by accident:

• See and undo. Verified answers, saved views, retired rules and dismissed corrections can be restored; hidden insights come back when they change (and Show hidden shows them now); Remembered file choices are listed on Data sources, each with Forget.
• Asked first. Deleting a measure, removing a saved answer, a board tile or a brief finding, retiring a rule, dismissing a correction, removing an AI key, discarding an import: each asks before it happens. Delete source lists everything that goes with the source.
• All conversations (from Recent): every conversation, searched by its title or its questions, with Rename, Delete and Delete all.
• Backups on Health: change how often and how many are kept, every backup listed with Download and Delete; the copies kept before each restore are listed too, and old ones tidied away by themselves.
• Smaller things: rename saved answers and board tiles, show a tile as a number, chart or table; delete a page of a brief and change its sources; take back a Yes; hide one suggestion on the home page; the financial year’s first month and the time new databases refresh are set on screen; Getting started again in the menu; Activity & trust lists every download.
• Your licence. Request a licence key, beside your computer’s code, writes the e-mail for you with the code in it. The trial ends at midnight after its last day. A computer’s clock once set ahead no longer locks out a valid key. A key pasted with words around it (“Key:”, quotes) is read; a key that ends sooner than the one in use asks first. After the trial ends you can still download your newest backup and make a support file; scheduled refreshes pause until a key is entered.
• Safer upgrades and data. The first start of a new version that changes your data takes a backup first. Replacing a file’s data can be undone (Undo the last file load on Data sources). Your corrections stay yours when you delete the conversation they came from. The largest upload is set on Health and logs.
• Clearer words. Error messages say what to do next; empty pages say how to fill them; the desktop no longer speaks of a team or a server. On a phone, board tiles, suggestions and the review tabs fit.

New in 3.0.0-rc.8

• A new look: the Silver Bullet logo is now two chevrons on Deep Mint, in the app, the browser tab, the app icon, the menu bar and the tray.
• A welcome on first start: step 1 sets up the AI model (a cloud model and its key, a model on this computer, or none for now), step 2 adds your data (a file or a database). Either can be left for later.
• Create a read-only login: adding a database writes the exact statements for a read-only account, for you to run or to send — Save as a file, with this computer’s address filled in.
• A cloud model is sent no value from your data that the question does not name.
• Industry and region packs (Tech Preview): standard measures for banking and lending, insurance, manufacturing, healthcare, retail and distribution, and education, and the way money and dates are written in seven regions.
• Banking and lending (BFSI) pack: a loan book is recognised by its columns (loan account, principal outstanding, days past due…) and gets its standard measures by itself — loan book, GNPA % (NPL ratio outside India), 30+ DPD %, gross NPA, NNPA %, provision coverage, disbursements, ticket size, collection efficiency, write-offs and recovery rate, as the columns allow — each tried on your data first and defined in words in every answer that uses it. Undo on the Data dictionary takes it back; your own measures and meanings always win.
• Admin → Packs, marked Tech Preview: the region, whether packs apply by themselves, and every industry pack with an on/off switch and the data sources it is used on.
• Two-step sign-in (optional): user menu → Two-step sign-in. Scan a QR code with Google Authenticator (or any authenticator app) and opening Silver Bullet then also asks for the app’s 6-digit code; ten one-time recovery codes for a lost phone; reset-admin turns it off when both are lost.
• Education pack (schools, colleges, coaching): students, fees due and collected, fee collection %, fees outstanding, concession %, attendance % and students below 75% attendance, average marks %, pass rate, student–teacher ratio, enquiry-to-admission conversion and dropout rate, as the columns allow. Student IDs, names, dates of birth and parents’ phones are personal by default.
• Retail and distribution pack: store bills, store summaries, stock and a distributor’s receivables get net sales, bills, average bill value, units per bill, average selling price, gross margin, discount %, returns %, sales per sq ft, conversion, sell-through, customers, receivables, overdue % and collection %, as the columns allow. It applies itself only to clearly retail or distribution data, never to a plain invoice table.
• Healthcare pack (hospitals and clinics): admissions, discharges, average length of stay, bed occupancy, revenue per occupied bed day (ARPOB), per admission and per visit, OPD footfall, patients, 30-day readmission and mortality rates and claim disallowance rate, as the columns allow. Patient ID, name, date of birth, phone, address and diagnosis are personal by default.
• Manufacturing pack (pharma batches included): a production log, downtime or maintenance log, dispatches or batch record gets production, rejection rate and PPM, first pass yield, scrap rate, plan attainment, capacity utilisation, availability, downtime, breakdowns and mean time to repair, energy per unit, on-time delivery and batch yield, as its columns allow — “Line 2 had the highest rejection rate: 5.7%”.
• Packs say what they cannot work out, and why: OEE, MTBF, roll rates, vintages, persistency and the renewal ratio are answered with the reason and the nearest figure the data has, never guessed.
• Insurance (BFSI) pack: a policy register, claims register or portfolio summary gets gross written and net earned premium, claims incurred and paid, loss, expense, combined and commission ratios, claim count, severity, claim settlement ratio, claim frequency, policy count and lapse rate, as its columns allow — “Health had the highest loss ratio: 90.8%”, with the definition in How I got this.
• Regions: money, dates and the financial year are written the way your country writes them: India, United States, United Kingdom, Europe (euro), Canada, Australia, or International (English). A new installation picks its region from the computer’s time zone and language; an existing one keeps India. Change in the welcome or under Health → Region and formats.
• Problems on Health: what is wrong (the AI model, a data source, room on the disk, backups, the licence) in plain words, each with a button that fixes it: Set up the model, Check the connection, Free space, Back up now, Open the licence.
• Send to support on Health: what the support file holds is listed before you save it, then Open the support page or E-mail support.
• Restore in the app (desktop): Health → Backups lists the backups kept; Restore backs up everything first (so it can be undone), restores, starts again and signs the page in again by itself. The restore command stays for when Silver Bullet will not start.
• Database sources show their own logo; long joined names stay inside the answer card; chart axes say lakh or crore.
• Security updates: Go 1.26.10 (crypto/x509, GO-2026-6809) and Excelize at its fix for GO-2026-6695.

In 3.0.0-rc.7

• Activity & trust: every question you asked and every request sent to an AI model, exactly as sent, beside what was kept. A lens on every answer (Nothing sent, or What was sent · 2.1 KB, no data); an overview with where requests went; flags for personal values typed into questions; a tamper-evident record kept a year (request text 90 days), with signed zips and Check a zip. Measured with 750,000 entries.
• Insights and executive briefs (described above).
• Network and firewall (Admin): the databases, AI model and other services this installation must reach, the host names a web proxy must allow, and what is never needed (no inbound rule at all on the desktop; no licence server, update check or telemetry). Check from here tries each connection; Download CSV and Copy as text give the request for your network team.
• A security re-audit (see Security below).

Since the first release candidate

The changes that matter on the desktop, from 3.0.0-rc.1 to rc.6, in short:

• Windows packages, tested on Windows 11.
• A 14-day trial and a licence per computer: the Licence page shows This computer’s code with Copy code (also Copy licence code in the tray menu); reinstalling keeps the code. Copy sign-in link in the tray menu, for when the browser did not open.
• The desktop is one person on one computer: no people to manage, share links or sign-ins; your corrections are checked and approved by you under Your corrections.
• Charts asked for by name (pie, doughnut, spider, histogram, scatter, bubble) and a Chart chip; text in the main area drawn at 85% on laptop and desktop screens.
• A loaded sales export fills the home page with no setting by hand (dates, main measure, date and grouping guessed); home cards chosen from the data; Suggested for you (You often ask, and up to four AI-suggested questions, each checked against the data first).
• Expert queries and saved views, No AI model, and Gemini’s embedding model (all described above).
• The Components page with the SBOM, the vulnerability report and signed advisories; SQLite 3.53.4, fixing CVE-2025-3277 and five high-severity issues.
• A loaded file’s memory is given back at once (a 190 MB CSV: 2.5 GB down to under 200 MB); a busy cloud model is asked again twice before an error; the read-only login SQL is on the add-database form.

Measured results

The Test Kit at the first release candidate

The Test Kit (Vardhora Hospitality Supplies, INR and USD; 51 questions, 41 for the database, 29 for the files; see the Test Kit) through Silver Bullet’s answer engine with the local model qwen3-coder-next on Ollama, every question typed as a person would, on five test databases and the files, twice, on the 3.0.0-rc.1 release build, on 4 October 2026. The answer engine has not changed since rc.6, so the kit was not run again for rc.7.

Suite 2, with the same model, got 38 of 41 (93%) on PostgreSQL and 23 of 29 (79%) on the files. Every question asking to change data was refused.

The wrong answers (the same in both runs), and whether the answer said Please check:

No right answer was marked Please check (0 of 219 in each run). An answer the model wrote SQL for says that Silver Bullet could not show its reading fully: check those, and correct them with No, fix it.

Later measurements: with expert queries for the wrong answers, 203 of 205 database answers right (from 187), none made wrong. On Google Gemini 3.5 Flash-Lite, 199 of 205 without any teaching; again on 21 October 2026, sent only the question’s own values (a cloud model never sees a list of your data’s values), 200 of 205 on the databases and 27 of 29 on the files: no loss of accuracy. With No AI model, 10 of 205 on the databases and 1 of 29 on the files (the kit’s questions are written for a model).

The learning loop

On the Test Kit (five databases and the files, run twice): wrong answers corrected through “No, fix it” and checked, and right answers marked as verified, then asked again in other words. In each run nine wrong answers were corrected, 71 of 71 later questions (originals and paraphrases) were right, and no right answer became wrong: the five databases went from 192 (run 2: 193) to 202 of 205; the files stayed at 26 of 29 (their wrong answers were not corrected in this loop). The checks stopped 13 of 24 marks that would have changed other answers, 8 of them found only by the model check. Details in How Silver Bullet Learns.

Speed and safety

• Speed (an Apple silicon Mac, the test databases on the same Mac): a typed question answered with the local model in 4.4 seconds at the median, 13 seconds at the 90th percentile, 24 seconds at most; changing a chip 2–280 milliseconds (no AI model); the home page’s brief 1–5 milliseconds (no database is queried).
• Safety: the read-only safety tests of every database connector passed on PostgreSQL 18, MySQL 9.7, MySQL 8.4, SQL Server 2022 and Oracle 26ai.

Security

• Tested in-house: every route for every kind of caller, secrets at rest, outbound connections, what is sent to the AI model, prompt injection through data, personal-data masking on five databases and files, and fuzz testing of the parsers; re-audited for rc.7. Fixed among others: personal-data masking that some queries could get round, two database functions that could run a query given as text, an Excel file that could take gigabytes of memory.
• Spreadsheets: copied and downloaded tables never carry a formula (cells starting with =, +, - or @ are written as text).
• No code from the internet: the files engine never downloads extensions. Silver Bullet contacts no licence server and sends no telemetry.
• Components and vulnerabilities: each release lists what Silver Bullet is built from (CycloneDX 1.6 and SPDX 2.3) and the known vulnerabilities as of the build, with severity, whether the code is reachable and whether a fix exists. Admin → Components shows both without the internet, and Load an advisory checks a signed advisory published after a release. rc.7: 4 findings, none open (each not affected, with the reason, or accepted).

Moving from Suite 2

• Fresh start, no import. Add the data sources again (the same read-only database logins work), load the files again, and set up the data dictionary. Suite 2’s query rules and verified answers are not carried over.
• Suite 2 is not touched and can keep running beside Silver Bullet 3 (different port, folder and trial).
• Each computer needs its own Silver Bullet 3 licence key.

Compatibility

• macOS 11 or later (Apple silicon and Intel), Windows 10 and 11 (x64), and Linux with glibc 2.31 or later (x64 and arm64: Ubuntu 20.04+, Debian 11+, RHEL 9 and relatives).
• Tested: macOS on Apple silicon and Intel; Linux x64 and arm64; Windows 11 on Arm through Windows’ x64 emulation (no console window, the browser opens signed in, the trial, files and questions, a second start hands over, stops by itself with everything saved, nothing lost after End task; the tray icon in earlier release candidates).
• Databases: PostgreSQL 14–18, MySQL 8.4 and 9.7, MariaDB 10.11–12.3 (10.5 and 10.6 with a warning), SQL Server 2017–2025, Oracle 19c, 21c and 26ai. The Installation Guide has the full matrix with what was tested where.
• Browsers: current Chrome, Edge, Firefox and Safari.
• Upgrading: replace the program with the newer version; your data folder is kept (tested from rc.7 to rc.8).

Known limits

• Not tested yet: a physical x64 Windows PC; Kerberos and Microsoft Entra ID sign-ins to SQL Server, and SQL Server’s Windows sign-in; Oracle 19c on a real server (21c and 26ai were tested). Of the cloud AI models, only Google Gemini was measured.
• Model variance: the same question can be read differently after a change to the dictionary, the rules or the release, and sometimes from one run to the next. Answers the model writes SQL for are the weakest: check them. Please check flags some signs of a wrong answer, not all: an answer without it is not guaranteed right. Measure on your own data before relying on it (the Test Kit shows how).
• The model check covers recent questions: those asked of the source in the last 90 days (at most 200), of which at most 40 are asked within 5 minutes. It needs an AI model; without one, only the replay runs (and says so).
• Matching by meaning has limits. A verified answer is reused only when a question names the same specifics. The check for opposites (paid / unpaid, on time / late, with / without, included / excluded) knows a fixed list of word patterns, so another opposite (“new customers” against “returning customers”) is caught only when another detail differs. A paraphrase further than the threshold goes to the model.
• Indian scripts in board PDFs: PDFs print Latin text and ₹; Devanagari and other Indian scripts are not printed yet.
• Not in this release: combining two sources in one question (for example Excel targets against Oracle actuals), the Snowflake plugin, e-mailing boards, languages other than English.
• Signed by Govirang, not by Apple or Microsoft: every package carries Govirang’s release-key signature only, by design (EULA clause 10A). The Mac app has an ad hoc signature (not notarised) and the Windows program has no Microsoft-trusted signature, so both systems warn the first time you start it (README-FIRST.txt in the package says what to do).