- Human review required
- Full audit trail
- Federated glossaries
- Per-language quality
What's included
One case model.
Every language, every channel.
Why Docket with Chorus
Machine translation quality is not the bottleneck.
A system to govern it is.
A typical language access unit is small — around seven staff — covering fifty or more departments, each running its own intake, often a single shared form, with no shared glossaries across departments. Translations are inconsistent from one department to the next, interpreter availability at public counters is unpredictable, and staff struggle to quickly identify a visitor's language. The unit is simultaneously the service provider and the policy authority for the whole organisation.
Frontier models already clear the quality bar for major languages. What's missing is a system to route work, enforce a standard, capture what was decided, prove it happened, and let a small team supervise a large federated organisation. That's a case management problem.
What It Does
The four Docket with Chorus capabilities
Risk Tiering & Policy
Classifies every request into a risk tier -- Emergency, Vital, Program, Routine -- using deterministic policy rules first, then a model pass that must cite the specific triggering text. Requesters can raise a tier instantly; lowering one requires approval and is always recorded.
- What it does: classifies every request into a risk tier with a cited policy trigger
- What it does not do: never auto-approves Tier 0-2 content
- Reviewed by: the language access unit, via override analytics
Terminology & Style Governance
Manages a three-layer glossary -- organisation-wide canon, department layers, and time-bounded campaign layers -- resolved most-specific-wins, with one exception: a locked canon term always wins.
- What it does: resolves terminology across three glossary layers, most-specific-wins
- What it does not do: never merges or silently overwrites conflicting department terminology
- Reviewed by: the central language authority, in conflict arbitration
Multi-Engine Translation & Review
Runs two to three translation engines independently per segment, scores their agreement with a judge model, and uses that quality estimate plus tier to decide whether and how closely a human reviews it.
- What it does: scores engine agreement and routes only flagged segments to a reviewer
- What it does not do: never lets locked-term violations skip human review, regardless of tier
- Reviewed by: trained linguists and department bilingual staff, same screen
Audit, Provenance & Compliance
Maintains an append-only, hash-chained audit trail attached at the sentence level, and produces live compliance reporting -- coverage, gaps, turnaround, oversight rate -- that used to be compiled by hand, annually.
- What it does: records actor, timestamp, and the exact policy version in force
- What it does not do: never reports a single blended accuracy number across languages
- Reviewed by: agency leadership and compliance officers, on demand
How It Works
From intake to proof
Every request, document or live session follows the same four-stage path, whether it is a routine notice or an emergency alert.
Intake
Any channel -- file upload, pasted text, URL, photo -- with a live word or page count, one-click language selection, and a real-time tier preview shown before submission.
Tier & Route
Deterministic policy plus a cited model proposal set the tier, which determines review depth, engine configuration, and turnaround target.
Translate & Review
Multiple engines translate independently, a judge model flags what needs human eyes, and reviewers work a keyboard-first, flagged-only workbench with full visibility into what the machine considered.
Deliver & Prove
Delivery ships with a full provenance record and writes into the same audit trail and compliance dashboard used for every other case, document or live session.
The same lifecycle applies to real-time interpretation. A session moves through Scheduled, Live, Captured, In Review, and Approved -- using the same tier, glossary, and audit trail as a document. The listener experience requires no app install and no account: a language menu written in-language, live captions, and WCAG 2.1 AA accessibility.
Governance
Governance you can put in front of your board
Every claim in the system traces to a real action, a real reviewer, and a real point in time. The assistant interface can do nothing a person couldn't do themselves, and every override is visible, never silent.
Grounded, Never Autonomous
The assistant interface can do nothing a person couldn't do themselves. It inherits the signed-in user's exact permissions -- never a service account -- and asks for confirmation before any write action with a structured preview of the change. Every chat-initiated action is logged identically to a manual click.
Full Audit Trail
Every claim traces to a real action, a real reviewer, and a real point in time, down to the sentence. Questions like who approved this exact sentence, or which glossary version was in force, are answerable in under a minute, not a month.
Federated, Not Flattened
Each department keeps ownership of its own glossary layer; nothing is silently merged or overwritten. Only a centrally locked canon term overrides a department's choice, and every override is visible, not silent.
FAQ
Common questions
Everything you need to know before we start working together.
What is Docket with Chorus?
A government case management platform (Docket) with a language access layer (Chorus) configured on top of it -- one system of record for how an agency decides what gets translated, who reviews it, and how it proves compliance.
Is this a translation tool with an admin panel bolted on?
No -- it's the reverse. Docket is a horizontal case management platform built for permits, appeals, and other case types generally. Chorus is what gets added when an agency's case type happens to require multiple languages.
Does the system make final decisions on tier or approval automatically?
No. Deterministic policy and a cited model proposal suggest a tier, but a human always approves Tier 0-2 content. Auto-approval only exists at the lowest routine tier, above a threshold the agency sets, and can be turned off.
How is quality measured?
Per language, continuously, against the agency's own previously published reference translations -- never as a single blended score, and reported even where results are weaker for lower-resource languages.
Does a session (live interpretation) get treated differently from a translated document?
No. A session and a document are the same underlying case object with different lifecycle stages. They share the same tier logic, glossary, audit trail, and compliance reporting.
Who can access what?
Access is scoped to the person, not the screen. External validators, such as community-based organizations, can be granted case- or language-specific access with an automatic expiry -- they can review and comment but never edit glossaries or export data.