◆01
Policy-first accountability gate
A configurable policy runs before CiviQ engages at all answer, refuse, escalate, or hand to human-review and attaches its decision to every response.
Open source · Sovereign · Accountable
A grounded, sovereign citizen support agent councils can trust.
CiviQ answers residents only from the council's own published content, runs a policy gate before it engages, cites every answer, refuses rather than guesses, and records every decision on a small local model, with zero data egress.
Runs on a ~3B local model
Zero data egress
Apache-2.0
222 tests, green eval gate
A governed answer, end to end
Watch one resident question travel the whole accountable pipeline, policy gate, retrieval, refusal checks, citation, audit and come back as an answer you can defend.
The category
Not another council chatbot.
The chatbot is a commodity now, anyone can wire a model to a website this year. The value isn't the conversation. It's the accountability, governance, and intelligence layer beneath it: the part that decides whether to answer at all, proves where every word came from, and turns each question into a record you can defend.
That layer is the USP - and it's the part we own.
WHAT'S COMMODITY · WHAT'S THE USP
The problem
Information existing isn't the same as a resident getting helped.
Booking a bulk-waste collection should take less than two minutes. But, it took me twenty to thirty minutes, hunting through PDFs and website pages, second-guessing the whole way, nearly giving up to phone instead. The information existed. Getting a straight, trustworthy answer to it at nine o'clock on a Sunday did not.
The public-sector twist
The requirements here are different from a regular consumer chatbot. A public service has to:
- →Be correct, or honestly silent.
- →Be accountable for every answer it gives.
- →Be accessible to everyone, by default.
- →Keep data on the institution's own infrastructure.
None of that comes from just a bigger models.
What holds it up
Six pillars.
◆02
Refusal-first
It answers only when it has supporting content and clears its checks. Refusal correctness is measured, not assumed. Saying "I don't know" is a first-class outcome.
◆03
Glass-box evidence
Every answer ships its sources and policy checks; every decision is hash-chained in a tamper-evident audit.
◆04
Sovereign / zero-egress
Runs on your infrastructure, on a small local model or none at all. No resident data leaves the institution.
◆05
Omni-channel
One governed core behind channel adapters - web, embed, SMS, messaging, voice, each unit-tested against simulated provider payloads.
◆06
Citizen-experience intelligence (CiX)
Every question, and every refusal, becomes signal, a ranked backlog of what to fix, write, and automate next. The assistant is the sensor, the intelligence is the wedge.
The thesis I'm building on
Capability is a system property, not a model property.
That's the bet this whole project is built to test: that a small, sovereign model inside a strong accountability framework can hold its own on what actually matters to councils - correctness, governance, auditability, residency, and cost. It's a hypothesis I'm exploring in the open, not a finished claim and CiviQ is the experiment.
How it works
One governed request lifecycle.
Every resident question travels the same accountable path. It isn't a model improvising an answer, it's a sequence of deliberate stages, each one able to stop, refuse, or hand over before anything reaches the resident.
Two gates can halt a request cold rather than let it guess: one before generation, if there's no supporting content, and one after, if the answer can't clear its own checks. Whatever happens, the decision and its reason are recorded.
See the full technical architecture01
Policy gate gate
your policy decides whether to engage, and how
02
Skills
deterministic answers where appropriate, with provenance
03
Hybrid retrieval
finds relevant passages in your published content
04
Ground-or-refuse gate
stops here unless there is real supporting content
05
Grounded generation
drafts strictly from the retrieved passages
06
Self-check + confidence
challenges its own draft and scores its certainty
07
Refuse gate gate
discards the answer if the check or confidence falls short
08
Accessibility check
verifies the wording is plain and clear
09
Cited answer or named handover
delivers a cited answer, or hands over with context
10
Tamper-evident audit
records the decision and its reason, hash-chained
Accountability
Every answer is evidence. Every decision is a record.
Each response carries its sources, the basis for its confidence, and the policy checks it passed. Each decision is hash-chained with the reason it was made, into a signed, exportable log built for the Ombudsman, IG, and FOI. Not logging as an afterthought -accountability as the architecture.
Hash-chained ledger
#1042 answer · grounded+policy:answer · 9f3a…c1
#1043 refuse · low-confidence · prev 9f3a…c1 → 2b77…e4
#1044 handover · out-of-scope · prev 2b77…e4 → c10d…9a
each entry seals the one before it · break one and the chain breaks visibly
Sovereign & affordable
Your data never leaves the building and the bill doesn't grow with every question.
Because CiviQ runs a small model on your own infrastructure (or no model at all) with zero egress, cost is bound to infrastructure, not metered per query. It stays roughly flat as usage grows — the opposite of cloud-API pricing, which scales with every message. Sovereignty isn't a premium tier here; it's the default.
Detailed sample cost modelling is in the repo for those who want it.
What you don't pay for
One core, every channel
Wherever a resident reaches you, the same governed core answers.
Each channel is a thin adapter, built and unit-tested against simulated provider payloads — ready to wire to a live carrier, not yet running against one. Whichever surface a resident uses, the request meets the same policy, grounding, refusal logic, and audit; the accountable core never changes.
One governed core — not a chatbot bolted onto each surface.
Channel adapters · simulated-payload tested
Web
Embed / iframe
SMS
WhatsApp
Telegram
Short-code / USSD
Voice / IVR
CiX intelligence · the wedge
Every question and every refusal - is a backlog of what to fix next.
A refusal isn't a failure; it's a flag that your content has a gap. CiviQ clusters demand and gaps over its own audit log, scores the citizen experience, and ranks what to write or automate next. The assistant is the sensor; the intelligence is the wedge, where CiviQ stops being a chat surface and starts improving the service itself.
Honest comparison
How it compares.
| Capability | Typical cloud chatbot | Frontier-LLM cloud | CiviQ |
|---|---|---|---|
| Answers from your content, cited | Sometimes; rarely cited | If retrieval is added; citations optional | Always grounded + cited |
| Hallucination control | Minimal | Reduced, not eliminated | Refusal-first + self-check |
| Honest refusal | Rarely | Inconsistent | Core, and measured |
| Data residency | Vendor cloud | US / vendor cloud | On-prem, zero egress |
| Running-cost behaviour | Per-seat / per-query | Per-token, scales with use | Infrastructure-bound, flat |
| Lock-in | High | High | Apache-2.0, portable |
| Audit trail | Limited | Provider logs | Tamper-evident, signed |
| Named human handover | Rare | Bolt-on | Built-in, with context |
| Governed policy before answering | None | None by default | Policy gate runs first |
Where the moat could travel
Built for councils. The governed core is what could extend.
CiviQ is built for councils, and government bodies - that's the focus, and where it's getting tested. Because the moat is the governed core (policy, grounding, refusal, audit), the same approach could extend to other public institutions that answer to the same bar. Institution profiles are config-driven, so a new context is configuration rather than a fork, but councils are the only one it's built and proven for today.
◆Built & proven here first
Local government
Waste, council tax, planning, parking, missed collections - the everyday, high-volume questions that flood the contact centre. This is CiviQ's home turf: where the policy gate, grounding, refusal, and audit are built and proven before anything else.
Universities
Admissions, fees, accommodation, visas - high-stakes answers that must come from the prospectus and policy, never guessed.
NHS trusts
Appointments, services, and guidance - where an honest refusal matters more than a confident wrong answer.
Housing providers
Repairs, tenancies, complaints - with a named handover the moment a case needs a person.
Police & public safety
Non-emergency guidance from force content, under strict scope and an up-front policy gate.
Universities, NHS trusts, housing and police are illustrative of where the same approach could be applied, not contexts CiviQ is built or tested for yet.
Honest status
What's built and what's honestly still ahead.
Discovery-stage and open-source - not live-deployed, with no council pilot yet. The eval gate runs green: 15/15 answerable, 7/7 correct refusals, 3/3 out-of-scope, 0 false refusals, 0 confident-wrong, on a 27-item set, not live traffic. Next in our pipeline is to expand the eval pipeline to 1000 and test it, I will keep updating here the true numbers.
Built & verified
- ✓222 automated tests
- ✓Green accuracy / refusal eval gate
- ✓Tamper-evident audit log
- ✓ GOV.UK design system based control panel
- ✓One-command install
- ✓DPIA pack, threat model & cert-readiness map
Still ahead, stated plainly
- ○A live design-partner council, with real-traffic numbers
- ○An external WCAG 2.2 AA audit
- ○A signed DPIA
- ○Cyber Essentials / penetration test
- ○A head-to-head small-vs-frontier benchmark
Get started
Run it on your own infrastructure, or talk through a pilot.
CiviQ is open-source and self-hostable, clone it, point it at the council's published content, and it runs on your own infrastructure. Or get in touch to talk through a discovery-stage pilot.
Get involved
It's open source. Help shape it.
CiviQ is Apache-2.0 and open to contribution. If the approach resonates, the simplest first step is to star and watch the repo to follow along - and to tell me where it's wrong.
Where help moves it forward
- →A council willing to run a discovery-stage pilot
- →Contributors on the channel adapters
- →An accessibility audit against WCAG 2.2 AA
- →Content-ingestion adapters for more sources
- →Honest feedback on the approach