Method
Most moderation tooling optimises the moment of detection. NuanceDesk optimises the moment of judgement — the minutes a human spends deciding what a flag actually means. The method below is built into every analysis the system produces.
The analysis begins by extracting the specific excerpts that matter — the quoted threat, the attribution, the condemnation, the coded phrase — each with a location, a confidence, and an explanation.
Only after the evidence is on the table does the system state a verdict. Reviewers can trace every conclusion back to the excerpt that produced it, and click any excerpt to see it highlighted in the original source.
For every case, the system is required to look for risk-reducing context with the same seriousness it applies to risk signals: journalism, condemnation, counterspeech, academic distance, satire, quotation for criticism.
The result is the Context Balance — two visible evidence columns around a central axis. A case with a frightening quotation and strong journalistic framing looks exactly like what it is: reporting, not threat.
When decisive context is missing — an ambiguous symbol, an unverifiable publisher, a fragment without its thread — the analysis says so, names the missing fact, and asks the reviewer a concrete question.
'Needs context' is a first-class verdict, not a failure state. A case closed with a clear open question is a better record than a low-confidence enforcement.
Every analysis maps its evidence to the specific rules of the workspace's selected policy pack, with match strength and reasoning. Rules carry their own mitigating-context definitions, so the newsworthiness exception is applied by design rather than by memory.
Published policy versions are immutable. A decision made under v1.0 stays auditable under v1.0, no matter how the policy evolves afterwards.
The recommended action must be the least restrictive action the evidence supports, and every analysis includes a written proportionality check plus a stated over-enforcement risk.
Removal is one tool among eight — allow, label, limit, remove, preserve-and-escalate, request context, no action, archive duplicate — because real queues need more than a delete button.
The AI never executes enforcement. A named human reviewer accepts, modifies, or overrides the recommendation, writes a rationale, and finalises the decision. Overrides require a written reason — and override patterns feed back into team analytics.
This is not a compliance fig leaf. It is the product's core loop: the system organises the evidence; a person makes the call.
Every finalised case carries its evidence, its policy version, its reviewer, and its rationale. When a decision is questioned — by a user, a regulator, or your own team six months later — the answer is one export away.