When AI misstates a rule, citizens act on the wrong answer.
An AI assistant tells a traveller, a worker, a student the wrong eligibility, procedure or fee — and they act on it. Often the correct rule exists, is official and is free to read, but no machine can extract it. We find why, and fix it — natively, in the languages you serve.
Bought through empanelment, tender or purchase order — no payment form.
Asked the way your audience asks
Diagnose
one root cause
Correct
from your facts
Prove
a crawler fetched it
The problem
The rule is correct. It's just unreachable.
Citizens no longer read the notification or the circular end to end — they ask an AI assistant and act on the single answer it returns. When the official page publishes rules only as scanned PDFs or JavaScript-only portals, retrieval falls through to an adjacent surface still carrying old text. Nobody did anything wrong — a scanner was used instead of a text export, once — and the consequence lands on a citizen with a deadline. The whole workspace speaks one language: accuracy and evidence — whether a citizen's answer is correct, and the proof that it is.
In the wild
Where a wrong answer reaches a citizen
A traveller checks a visa fee
An AI assistant quotes a fee from a schedule that changed at the last revision, and the traveller plans to pay the old amount before a deadline. We compare it to the fee in force, make the current figure and its effective date machine-readable at the source, and confirm the AI assistant can reach it.
A student tests scholarship eligibility
A model states an income ceiling or age limit that no longer matches the criteria, and the student decides not to apply. We diagnose why the correct rule couldn't be extracted, publish it natively from your own record, and re-measure until the answer matches.
A worker asks about a deadline
An AI assistant describes an obsolete filing window or a superseded document requirement, and the worker misses the real cut-off. We trace the stale source, correct the deadline against the verified record, and keep the evidence trail.
A pensioner checks a benefit
An AI assistant states a pension rate or a life-certificate procedure revised at the last notification, and a senior plans around the wrong figure or submits the wrong form. We compare it to the current entitlement, make the correct figure machine-readable, and confirm the answer is right.
Who currently heads the department?
Someone asks who leads the department and the answer names a predecessor who moved on years ago. We show which engines are behind, publish the current office-holder as extractable, dated text, and re-measure until the answer reflects who's in post.
A rule answered in the wrong language
A citizen asks about a rule in a regional language and the AI assistant answers wrongly, because the authoritative wording exists only in English. We diagnose it as a language gap, prepare the native-language version from your record, and confirm the answer is right in the language it was asked in.
An impersonating site feeding the answer
An AI assistant answers from a page that impersonates the department, or repeats an unofficial post, and the citizen acts on it. We name the exact source, publish the authoritative record for the engine to cite instead, and file a report-request — every step kept in the audit trail.
How it works for you
Find the one cause. Fix it from your facts. Prove it landed.
- 1
Measure
We ask the questions citizens actually ask, in every contracted language, and capture each answer verbatim with its sources, the model version and the timestamp.
- 2
Diagnose
Every wrong answer resolves to one cause — a scanned PDF with no text layer, a fact only in English, a page no crawler could fetch. Computed, not guessed.
- 3
Correct
A ready-to-publish correction on your own domain, built only from your verified record, in each contracted language — natively written, not machine-translated.
- 4
Prove
We confirm a genuine AI crawler fetched it, then re-measure — the answer, its evidence, and the movement, side by side.
The substance
How the loop works, in detail
Ask the questions citizens actually ask
We don't test slogans — we ask the plain questions a citizen puts to an AI assistant about eligibility, fees, procedures and deadlines, in every language you are contracted to serve. The same questions go to the AI assistants your citizens actually use, so the measurement reflects real usage, not an ideal query.
A verdict against the verified official record
Each answer is judged only against the current official record you confirm as authoritative — never opinion, never another source. It's marked correct, outdated or incorrect, with the AI assistant's exact text and the rule it should have returned side by side. The only question: is the answer a citizen received accurate as of the day it was given?
The seven root-cause diagnosis
When an answer is wrong we find the one reason the correct rule couldn't be reached — not a list of guesses. Common causes: a rule published only as a scanned image, content in one language while citizens ask in several, a JavaScript-only portal no machine reads, a page that blocks retrieval, a rule with no effective date, a superseded page left live, or a figure buried in an attachment. Naming the single cause is what makes the fix durable.
Native-language corrections from your own record
Corrections are drawn from your official record and published natively in each contracted language — never machine-translated, never invented. Where you hold the authoritative wording, we make it extractable, dated and reachable; where it's missing, we prepare it from your record for your officers to approve. Your office remains the author and approver of every word.
Fetch-proof and re-measurement with an audit trail
Publishing a correction isn't proof it worked. After a fix we look for server-log evidence that a named AI crawler retrieved the corrected page, re-ask the original questions, and record whether the answer is now accurate — every step retained as a dated audit trail you can show internally or to oversight.
How a wrong answer resolves to one cause, computed not guessed
The workspace doesn't hand an officer a list of possibilities to sift — it computes the cause. It fetches your official page as a plain machine would, no scripts, and checks whether the current fact is even present: if absent, it's unreadable to a crawler and must be restructured; if present but a non-official source was cited, your record must be fetched and dated; if your page is right and other engines answer correctly, this one engine is just behind and will refresh. One true cause, so the correction is durable, not a recurring patch.
Corrections written as structured data an AI assistant can extract
A correction is only useful if a machine can lift the fact out of it — so the workspace doesn't produce prose and hope a crawler notices. Each is generated in the formats AI assistants consume — structured data, a clean HTML fact block, an llms.txt entry — carrying the value and its effective date. Your officers approve the wording, drawn from your own record, and what publishes is text both a citizen and a model can read.
Capabilities
Built for how you work
Language accuracy gap
The accuracy of AI answers per contracted language, on the home screen — the number that makes the case for the work, invisible everywhere else.
Reputation Radar — misinformation & impersonation watch
Continuous monitoring across news, social and video for misinformation, impersonation and fabricated notices in your name — read, classified and cross-checked against your verified record, with an immediate alert on anything critical (including deepfakes) so a false circular is caught before it spreads.
Every number is evidence
Each figure drills to the captured answer, its cited sources, the exact model version and the timestamp. An unverifiable number is worse than no number.
Two-stage approval + audit trail
A desk officer prepares; a division head approves. The append-only trail is a deliverable, built to be read during an audit.
Accessibility & conformance
Government web conformance and accessibility standards for your jurisdiction — WCAG 2.1 AA, Section 508, GDS and the like — are built in from the first commit, not retrofitted after a pilot.
Machine-readable corrections
Each correction is generated as structured data, a plain-text HTML fact block and an llms.txt entry — the formats an AI assistant can extract a rule from, not a scanned notice it cannot read. Dated with an effective-from, so the current rule is unambiguous.
Fetch-proof from your own logs
After a correction publishes we look for server-log evidence that a named AI crawler actually retrieved the corrected page. Publishing is not proof it was read; the fetch record is.
Misinformation & impersonation defence
When an AI assistant repeats misinformation about a rule, or answers from a page that impersonates your department, we identify the exact source feeding the answer and give your officers two levers: publish the authoritative record so the engine has an official source to cite instead, and a report-request draft for the impersonating or misinforming page. Still only accuracy and evidence — never ranking, and there is no competitor.
Every figure here links straight to the captured answer, its sources, the model version and the timestamp — because for a public body an unverifiable number is worse than no number. The one question the workspace asks is whether the answer a citizen receives is accurate, and whether the official record can prove it. It never ranks the department against anyone; there is no competitor and no share-of-voice, only whether a citizen's answer is correct.
Data sovereignty
Hosted on your platform. Owned by you. Run by our engine.
The official record, every captured answer, the evidence and the audit trail can stay entirely inside your boundary. We run the measure, diagnose, correct and prove service against your data where it sits — you own every record, and nothing leaves your control without your say-so.
Managed, in-region
We host and operate the service in a sovereign region inside your jurisdiction. Data never leaves the country, and your officers get the full loop with none of the operational burden.
Private tenant in your cloud
Deployed into your own government cloud account or VPC, on your storage and your keys. We operate the engine; the data plane is yours, end to end.
On-premise / air-gapped
Runs fully inside your network with local model serving and zero egress — for environments where nothing may cross the perimeter.
You own all the data — the verified record, every captured answer, the evidence and the audit trail — with full export, hard delete, and a signed data-ownership and processing agreement. We are the engine; the record is always yours.
Inside the team
Who uses it inside a public body
Public-communications officer
Sees which citizen questions are being answered inaccurately by AI assistants, in which languages, and turns each finding into a correction sourced from the official record.
Spokesperson's office
Holds the verbatim answer, the source, the model version and the timestamp for any misstatement, so a public response rests on evidence rather than recollection.
Digital and web team
Receives the specific technical reason a rule was unreachable, whether a scanned document, a script-only page or a blocked path, and the exact change that makes it extractable.
Desk officer and division head
The desk officer prepares each correction from the record and the division head approves it before publication, keeping authorship and sign-off inside the responsible division.
Language and translation cell
Confirms that the native-language wording used in a correction matches the authoritative version the department holds, so a regional-language answer is accurate in its own right and never a machine translation of the English page.
Audit and oversight liaison
Draws on the append-only trail — every captured answer, its sources, the model version, the timestamp, and the correction that followed — to answer an internal review or an oversight query with evidence rather than recollection.
The takeaway
What your office gets
FAQ
Questions, answered
How do you fit our procurement or empanelment process?
We are engaged as a measurement and content-correction service against your official record, with a defined scope, deliverables and data terms suitable for a public-procurement or empanelment file. We can provide the documentation, references and conformance statements your process requires before any work begins.
How is our data handled?
We work only from the public official record you designate as authoritative and the answers AI assistants return about it; we do not process citizen personal data as part of the loop. Captured answers, sources and logs are retained as your audit trail under the retention and access terms agreed in the contract.
Which languages do you cover?
We test and correct in every language you are contracted to serve, using the questions citizens actually ask in each one. Corrections are published natively in those languages from your own record and are never machine-translated.
Do you edit or influence the AI models?
No. We do not train, tune or alter any AI assistant, and we make no attempt to influence how a model ranks or selects sources. We fix the official record so it is accurate, dated and machine-readable, then confirm the AI assistant can reach it.
Do corrections meet our accessibility and conformance requirements?
Corrections are prepared to be extractable and readable by both people and machines, and we align published text with the accessibility and conformance standards your body follows. Your team reviews and approves each correction against those standards before it goes live.
Do you rank us against other departments or measure share-of-voice?
No, and by design we cannot. There is no competitor, no ranking, and no share-of-voice anywhere in the workspace. The only question it asks is whether the answer a citizen receives about your rules is accurate against your own official record, and whether that accuracy can be proven. The vocabulary is accuracy and evidence, never marketing.
How do you decide why an answer was wrong — is it a guess?
It is computed, not guessed. We fetch your official page the way a plain crawler would, with no scripts, and check whether the current fact is actually present in that text. From that we resolve one cause: the fact is unreadable to a machine and must be restructured, the fact is missing in the language it was asked in, an unofficial or superseded source was reached instead of your record, or your page is correct and a single engine is simply behind. Each cause carries a specific fix, so the correction holds rather than recurring.
What exactly is captured as evidence for each answer?
For every answer we record the verbatim text the AI assistant returned, the sources it cited, the name and version of the model, and the timestamp, alongside the current rule from your record it should have returned. Nothing in the workspace is a bare number; every figure links to that captured evidence, because for a public body an unverifiable number is worse than no number.
When AI misstates a rule, citizens act on the wrong answer.
Bought through empanelment, tender or purchase order — no payment form.
Not sure which fits? See every kind of reputation we work on.