One source / Full context / Human decision
Moderation needs
context, not just
detection.
NuanceDesk turns difficult content into evidence-backed moderation decisions—without removing the context that changes what the content means.
No credentials needed for the sample · built for trust & safety teams
Source · harborledger.example
The spokesperson declared that “our fighters will strike again and the harbour will burn.” The recording, which the Ledger has reviewed but will not republish, was released hours after the attack…
Awaiting human decision
Recommended: allow with label · reviewer confirms
A flag is not
a decision.
Detection tells you something might be wrong. Enforcement decides what happens to a post, an account, a community. Between those two moments sits the hardest work in trust & safety: reading the material, weighing the context, applying the policy, and writing down why. NuanceDesk is built for that middle — the contextual review layer between detection and enforcement.
- Detect
- Understand
- Review
- Decide
- Document
The highlighted middle is where NuanceDesk lives.
Interactive
Context changes the outcome.
Toggle between the isolated sentence a classifier sees and the full passage a reader sees. The material is fictional; the problem is not.
“Our fighters will strike again and the harbour will burn.”
A detector that sees only this sentence has every reason to flag it. Most systems stop here — and remove the wrong thing.
The Context Balance
Evidence on
both sides.
Instead of a single dangerous-looking score, every case shows two visible evidence columns: what raises the risk, and what changes its meaning. Reviewers see the argument, not just the number.
- ▲ Praise · recruitment · threats
- ● Journalism · condemnation · satire
▲ Risk signals
2 items
● Context signals
3 items
Built for the teams carrying the hardest queue.
Small platforms
The removal notices are just as hard when the team is four people.
Communities & forums
Volunteer moderators deserve better than gut feel and screenshots.
Marketplaces
Listings, sellers, and messages — with money moving underneath.
Gaming
Fast-moving chat culture where irony and threat blur.
Non-profits & research
Documenting harm without amplifying it.
Policy & escalation reviewers
The people who sign the decision need the full record.
Workflow
From source to defensible record.
- 01
Submit
Paste text, drop a URL, upload an image or PDF, or paste a transcript. Sources are hashed, redacted, and stored privately.
- 02
Analyse
A structured AI review maps evidence to your policy pack — risk signals, mitigating context, uncertainty, and a proportionality check.
- 03
Review
A human reads the Context Balance, answers the open questions, and accepts, modifies, or overrides the recommendation.
- 04
Export
Every case becomes a defensible record: decision, rationale, policy version, audit trail — as PDF, JSON, or an expiring share link.
Inside the workspace
A working tool, not a dashboard demo.
Context Balance — evidence on both sides of the axis
▲ Risk signals
2 items
● Context signals
3 items
Responsible AI
Principles the product enforces, not a page it links to.
- 01
Human final decision
The AI recommends; it never enforces. Every case closes with a named human reviewer and a written rationale.
- 02
Evidence before conclusions
Verdicts arrive attached to the excerpts that produced them — including the evidence that cuts the other way.
- 03
Explicit uncertainty
When decisive context is missing, the system says so and asks the reviewer a concrete question instead of guessing.
- 04
Proportionate action
Every recommendation passes a least-restrictive-action check, with the over-enforcement risk stated plainly.
- 05
Auditability
Analyses record their model, prompt version, and schema. Decisions record their policy version. Nothing is untraceable.
- 06
Privacy-aware handling
Contacts, links, and numbers are redacted at intake. Shared reports redact by default and expire on schedule.
Read the full Responsible AI statement.
Sensitive work, handled like it.
The material this tool exists for is exactly the material you least want leaking, lingering, or logged. Details on the security page.
- Private workspaces
- Strict per-workspace data isolation, enforced in every backend function.
- Role-based access
- Owner, admin, analyst, reviewer, and viewer — checked server-side, not just hidden in the UI.
- Controlled sharing
- Share links are hashed, expiring, revocable, and redact source content by default.
- Retention settings
- Configurable retention per workspace, with real deletion including stored files.
- Audit records
- Important mutations write an audit event with safe metadata only.
- Honest claims
- We state what is implemented. No certification badges we haven't earned.
Pilot plans
Priced for the teams doing the work.
Every plan starts with a workspace — no payment details required during the pilot. Billing is arranged with your team before any charge.
Free Pilot
£0for pilot teams
Evaluate the full workflow with your own policies and cases.
- 1 workspace · 3 members
- Fictional sample library included
- 50 live analyses / month
- Exports & expiring share links
Team
£149per workspace / month
For working trust & safety teams with a real queue.
- Unlimited members & roles
- 1,000 live analyses / month
- Policy Studio versioning
- Analytics & override tracking
- Priority support
Network
Let's talkfor networks & consortia
Multiple workspaces, shared policy templates, custom retention.
- Multi-workspace administration
- Custom analysis volume
- Co-branding for authorised partners
- Onboarding for review teams