About GC Simplified
An independent prototype exploring one idea: that most of the friction in government work is not the rules themselves, but the distance between how they are written and what a person is supposed to do next.
Important
GC Simplified is an independent tool, not affiliated with or endorsed by the Government of Canada, and it does not provide official advice. Plain language is an interpretation, and interpretations can be wrong. Always verify against the source and your organization’s own guidance before acting.
The two journeys
Deliberately separate, deliberately connected.
Step 1 · Translate
You read a sentence and cannot tell what it is asking. The translator explains the words, flags the jargon, and offers a plain, a new-employee and an informal reading of the same text.
Step 2 · Decode
You read the whole message and still cannot tell what it wants. The decoder quotes back the sentences that are actually requests, the dates that were named, and whether you're being chased — all of it counted from your text, none of it paraphrased. With an AI key configured it adds a reading of the intent on top.
Underneath both sits the reference they draw on: the glossary of acronyms and jargon, the said-vs-meant list, and a keyboard search over all of it (⌘K). Every entry is reachable on its own page, so nothing is locked inside a tool.
How the answers are produced
Three sources, always labelled, in descending order of trust: written by a person, produced by rules, generated by AI. You can always see which one you're looking at.
The trust ladder
- ✍️ Written by a person. The jargon glossary, the acronym entries and the curated translations. Nothing generated, nothing inferred.
- ⚙️ Rules, not AI. Word substitution, jargon detection, readability scoring, and the first pass of the decoder — which sentences are requests, what dates were named, whether you are being chased. Deterministic and blunt — the same input always gives the same output.
- ✨ AI-generated. Optional, opt-in, and never the default. Used where rules genuinely can’t cope: unfamiliar sentences, real French, and reading the intent behind a message. Always badged, always asks before sending anything.
Curated first, rules second
Common phrases have translations written by hand. When your text matches one, you get that. When it doesn’t, a rule-based rewriter substitutes plain equivalents and explains every glossary term it recognises — and the confidence label drops accordingly and says why.
Invented citations are caught on the server
A model can be instructed not to fabricate a citation. Only checking makes it true. Anything citation-shaped in an AI response — a named directive, a section number, an appendix — is matched against the text you supplied. If it wasn’t in your input, the result is discarded and you get the rule-based version instead. The model does not get to talk its way past the check.
The translator also runs backwards
“Make it sound official” takes a plain sentence and buries it under five layers, one at a time: longer words, institutional vocabulary, hedging, passive voice, then a preamble nobody reads. It is a joke, and it is labelled as one — but every layer is a real pattern found in government writing, and watching a clear sentence lose 90 readability points makes the argument for plain language faster than any style guide.
What this tool will never do
- Invent a policy requirement that isn't in the source text
- Fabricate a citation or a section number
- Create a link to an official page that doesn't exist
- Present an interpretation as if it were the source
- Hide uncertainty behind confident phrasing
- Quietly pick a side when two sources conflict — it flags the conflict instead
- Use Government of Canada logos, wordmarks or branding
Your data
- Nothing leaves your browser by default. Translation, analysis and search all run locally.
- The AI features are the exception, and they ask first. Pressing an ✨ AIbutton sends that text to DeepSeek, a third-party service outside this prototype, where it is subject to their terms and retention — not ours. Nothing is sent until you press it, the request goes through this app’s server so no key is exposed, and we store none of it. With no key configured the buttons simply don’t appear.
- Text you paste is never kept. What you put into the translator or the decoder lives in the page and is gone when you leave it. There is no account, no history and no server-side record of anything you looked up. The two exceptions are things you deliberately submit — a suggestion or a rating — and they are described below.
- Suggestions and ratings are sent, and nothing else is. Pressing submit on a suggestion, or rating an explanation, sends exactly what you typed into that form — and nothing alongside it. No name, no account, no address, nothing that links two submissions to the same person. A copy stays in your browser so you can see what you sent, and you can erase your own copy from the Insights page. Forms that reach a person are the point of having them; a form that quietly went nowhere would be the dishonest version.
- Page views are counted. Vercel Web Analytics records which pages are visited. It is cookieless and sets no cross-site identifier, but it is still a third party seeing traffic, and it would be strange to describe everything else here and leave that out.
- Don’t paste sensitive material anyway. Classified, protected, personal or otherwise sensitive government information does not belong in a prototype, including this one.
What's real and what isn't
- Nothing here states a threshold, a deadline or a clause number. Not as an oversight — as a rule. Those are exactly the facts that drift when a policy is amended, and exactly the facts people act on. A stale figure that looks authoritative is worse than no figure at all, so this tool explains what language means and stops there.
- The glossary is interpretation. Acronym expansions and jargon explanations reflect common usage, written by hand. They are plain-language readings, not official definitions — and the informal “what they probably mean” reading is a joke, labelled as one wherever it appears.
- The bureaucracy score is a count, not a verdict. Buzzwords, passive constructions and hedges are counted, never estimated, and every component can show you the exact strings it matched. Readability is Flesch reading ease, which is an approximation and is labelled as one everywhere it appears.
- No affiliation. Naming an organization is not a claim of any relationship with it. This tool is not connected to, endorsed by, or speaking for any organization mentioned in it.