judged.systems Review: A Judgment Engine for Support Tickets

What Is judged.systems?

It is a judgment endpoint for support operations. One ticket per call, typed answers, deterministic results. Instead of generating a chatbot response, the engine runs a pack of observational questions — is this urgent, does it request a refund, does it carry policy risk — and returns a decision your thresholds define: accept, review, or reject. Because the helpdesk remains the system of record, the service slots into an existing pipeline rather than replacing it.

How a Judgment Runs

The flow is small and explicit. You send a ticket with subject and customer message; the engine redacts emails, phone numbers, and payment-like numbers before anything is stored or shown to the model. One evaluation call answers every question on the same ticket, and code — not the model — applies the review rules to produce the result. Any rule that fires, or a model failure, stores review; everything else stores accept. A person can later label a review, and that label never retroactively changes the recorded evaluation.

• Redaction happens before storage and before the model
• One call answers the whole question pack
• Code decides the result, rules stay in your pack
• Model failures default to review, not silent accept

REST, MCP, and Webhooks

The service ships four ways to judge a ticket. REST is POST /api/judge with an API key and returns the stored evaluation; MCP connects an agent client over OAuth so an assistant can judge a published pack; an inbound webhook accepts a ticket with 202 and later POSTs the signed completion; and the dashboard playground judges a single ticket against a draft or published pack. API keys are scoped by permission, and an Idempotency-Key is supported for retries.

• POST /api/judge returns the stored evaluation
• MCP with OAuth for agent clients
• Inbound webhook with signed completion POST
• Playground for trying a pack before wiring it up

Packs, Versions, and Thresholds

A pack is the brain you control. New packs start from a triage draft covering department, urgency, refund requested, policy risk, and escalation, and you replace it with your own questions. Drafts are edited freely; publishing freezes an immutable version that REST, MCP, webhooks, and simulations all judge against, so a live pipeline never quietly changes under you. The next publish creates a new live version. Thresholds live in the pack too: a floor, a range, or a middle band decides what reaches a person.

• Choice, score, and boolean question types
• Published versions are immutable and live
• Thresholds define the review band
• Simulations can judge datasets against a published pack

What judged.systems Adds That Other AI Support Tools Miss

Most AI support products answer the customer; judged.systems decides what the team should do. Classification is framed as evidence: a choice plus probabilities, a score on a rubric, or the probability a statement is true. The model returns evidence and your thresholds make the call, which is a different trust model from a chat agent's confident paragraph. Redaction before storage, append-only evaluations, and threshold-visible results give a compliance story that reply-generation tools rarely match, because the vendor holds a deterministic, auditable record of every decision.

• Decisions with evidence instead of generated replies
• Evidence from the model, thresholds from you
• Redaction and append-only storage for audit paths
• Typed results that an approval workflow can consume

Security, Redaction, and Retention

Redaction is structural: sensitive identifiers are replaced before the ticket is stored and before the model ever sees it. Evaluations are append-only rows holding the pack version, the redacted ticket, the answers, the model id, usage, and latency, so you can replay what happened rather than trusting a dashboard. Authentication separates browser sessions, API keys, and MCP OAuth, with each route checking the permission it needs. For support pipelines subject to data-handling rules, that clarity is the headline feature.

• Sensitive data replaced before storage and inference
• Append-only evaluation records with model and usage metadata
• Session, API key, and MCP OAuth authentication separated
• Per-route permissions checked explicitly

Alternatives to judged.systems

The current directory has no competing judgment-API entry: adjacent tools like [ZenCall AI](/tools/zencall-ai) answer calls or messages rather than deciding triage. Externally, AI agents such as Fin by Intercom lean toward resolved-reply automation, while old-school routing rules in Zendesk and other helpdesks make deterministic but brittle category guesses. The service sits between those two: model-backed judgment with your thresholds and an auditable trail. Builders can also wire it through MCP into the same agent stack they already run.

• ZenCall AI — adjacent support automation, not a judgment API
• Fin by Intercom — reply generation versus decision output
• Helpdesk routing rules — deterministic but rigid
• MCP integration fits existing agent toolchains

Pricing & Plans

Free at launch with no published paid tier as of October 2026. The product is live on Product Hunt; docs cover REST, MCP, webhooks, and the dashboard playground.

Best For

Recommended use cases and scenarios where judged.systems shines.

Frequently Asked Questions

Common questions about judged.systems, answered.