◆Accountability
Prove it.
Evidence on every answer; a tamper-evident record of every decision and its reason.
Manifesto
Public institutions deserve an AI they can govern.
A statement of intent. CiviQ is accountable, sovereign, and open , built for correctness and trust, not engagement. The principle is general, but the product today is focused on local government: councils first. What follows is what we believe, what we refuse to build, and what is honestly true today.
The moment
A Sunday night, and a sofa.
It was nine o'clock on a Sunday when I tried to book a bulky-waste collection - an old sofa, a bed and a mattress. It should have taken two minutes. It took closer to thirty: clicking between PDFs and website pages, second-guessing every step, nearly giving up to phone a line that wasn't open.
Every fact I needed was published somewhere on the council's site. What didn't exist was a way to actually get helped - a straight, trustworthy answer, at the moment I needed it.
That gap, between information existing and a resident being helped, is the whole reason CiviQ exists.
The diagnosis
The obvious build is exactly wrong for the public sector.
Wire a frontier model to a council website and you inherit everything that model was built for and none of it fits a public service. The defaults are engagement-optimised, cloud-egressing, black-box, metered per message, and locked in.
A public institution's requirements run the other way. The sector doesn't need a more talkative model. It needs a more accountable system.
What the default optimises for — vs what the sector needs
- ✗Engagement → correctness
- ✗Cloud egress → data stays home
- ✗Black box → glass box
- ✗Per-message meter → flat, sovereign cost
- ✗Lock-in → open by default
What we believe
Ten tenets.
Correct or honestly silent.
Better to refuse than to guess. A wrong answer from a public service costs more than no answer.
The institution sets the policy, not the model.
What CiviQ may answer, refuse, or escalate is your decision - configured, versioned, and editable.
Show your work - glass-box by default.
Every answer carries its sources and checks. Nothing is asserted without evidence.
Data stays home.
Sovereignty is non-negotiable. Resident data does not leave the institution's infrastructure.
Never a dead end.
When CiviQ can't help, it hands over to a named human, with the conversation's context attached.
Memory is organisational, never personal.
It learns what content to improve - never who asked. Privacy by construction, not by promise.
Capability is a system property, not a model property.
Retrieval, grounding, refusal, self-checking and auditing are where capability lives.
Open by default.
Apache-2.0, self-hostable, no lock-in. You can read it, run it, fork it, and leave.
Measure, don't assert.
Refusal correctness, grounded-ness, and quality are evaluated - claims are backed by a green gate.
Accountability is a feature, not paperwork.
A tamper-evident record built for the Ombudsman, IG, and FOI is part of the product, not an add-on.
The redefinition
From "chatbot" to an accountable system.
"Chatbot" describes the surface and misses the substance. The work isn't the conversation; it's the governed system that decides whether to speak, proves where every answer came from, and records the decision. Calling that a chatbot undersells it the way calling a bank "an app with a balance" undersells a bank.
The category CiviQ is actually in: an accountable, sovereign citizen-support agent system for councils and government bodies.
The USP
Three layers no commodity chatbot combines.
The model is interchangeable, swap it out and the product still stands. The value is in the layers around it.
◆Governance
Control it.
A policy the institution sets and versions, running before the system engages at all.
◆Intelligence
Learn from it.
Every question and refusal becomes a ranked backlog of what to fix, write, and automate next.
Who it's for
Everyone the answer touches.
Residents
Especially the underserved and multilingual - a straight answer at 9pm on a Sunday, in plain language, or an honest handover.
Frontline staff
Fewer repetitive enquiries, and the ones that do reach a person arrive with context already attached.
IG, DPO & complaints
A signed, exportable, tamper-evident trail built for the Ombudsman, IG, and FOI from day one.
Service managers
A live read on demand and content gaps, what residents actually ask, and where the service falls short.
The institution
Sovereignty, open licensing, and flat cost, capability without dependence on a single vendor or cloud.
The anti-roadmap
What we refuse to build.
Some things we will not ship, not because we can't, but because they're wrong for a public institution. Each refusal is a commitment.
✗Commitment
No autonomous action on consequential journeys.
Benefits, housing, safeguarding - CiviQ informs and hands over. It does not act on a resident's behalf where the stakes are high.
✗Commitment
No per-citizen profiling.
It learns about content, never about people. No profiles, no targeting, no behavioural dossiers.
✗Commitment
No engagement optimisation.
Success is a resolved question, not time on page.
✗Commitment
No cloud egress by default.
Data stays on your infrastructure unless you explicitly choose otherwise.
✗Commitment
No black-box answers.
If it can't show its sources, it doesn't say it.
The thesis, in depth
Small sovereign model + strong framework > frontier black box.
On the axes councils actually buy on, the system beats the model.
"Capability is a system property" isn't a slogan; it's an engineering claim. The things a public service needs, finding the right passage, staying grounded in it, knowing when to stop, checking its own work, and recording why, are properties of retrieval, grounding, refusal gates, self-checking, and auditing. None of them get materially better by making the underlying model larger.
So a small, local, cheap model wrapped in a disciplined framework outperforms a frontier black box where it counts: correctness, governance, auditability, residency, and cost. The giant model wins benchmarks the sector isn't buying. The system wins the ones it is.
Honesty
What's true today.
CiviQ is at discovery stage. It is built and tested - 222 automated tests, and a refusal-and-accuracy eval gate that runs green on a 27-item set - but it is not yet running in live production, and there is no council pilot yet. The figures come from that evaluation set, not from real traffic. We're seeking a design-partner council to change that.
We say this plainly, as a matter of principle. A project about accountability has to be accountable about itself.
The bigger picture
An open, sovereign, accountable layer for public services.
This is public-interest technology. As AI moves into the institutions that serve everyone, the question isn't whether they'll use it - it's whether they'll govern it, or be governed by a vendor's defaults. Digital sovereignty means a public body can run capable AI on its own terms: its data, its policy, its infrastructure, its right to leave.
An open, sovereign, accountable layer is how that future stays public.
If you run a council, let's run a pilot.
CiviQ is open and self-hostable today. A design-partner council turns a discovery-stage build into a proven one and shapes what gets built next.