Brok — the Engineering Partner

Brok answers engineering questions about your AI system, in plain English, from your own engineering record.

It is not a chatbot, and it is not ChatGPT in a dashboard. There is no language model behind it. Ask it something the record cannot answer and it will say so.

What it does

You ask "Why did last night's deployment fail?" and Brok reads the evaluation record, the failed runs, the provider attribution and the artifacts involved, then answers with:

• A verdict — one line saying what is true, with its epistemic status and confidence.
• Reasoning — a chain of statements, each declaring whether it is derived or inferred, and

what it was read from.

• Impact — what this holds open, in engineering terms.
• Evidence — the actual records it read, clickable.
• Recommendations — each with its reasoning, its confidence, and a **next action into a real

surface** of the product.

• Follow-ups — engineering-specific questions worth asking next.

Every answer follows that shape. It is the constitutional narrative, and it is enforced by the data model rather than by convention.

The conversation carries context

The thing that makes Brok feel like a partner rather than a search box: you never restate the subject.

   You:  Why was Support Prompt promoted?
   Brok: ... (subject resolved: Support Prompt)

   You:  Show me the evidence.
   Brok: Read as a question about Support Prompt, carried from
         "Why was Support Prompt promoted?" ...

   You:  Compare it with v1.
   You:  Open the graph.
   You:  Should I promote it?
   You:  What was the reasoning?

Six turns, one subject named once. And notice the second answer: the inherited subject is declared, not assumed silently. An inherited context is an auditable claim, so if Brok carried the wrong subject forward, you can see it immediately.

The questions it answers

Brok resolves 25 engineering intents. These are genuinely different workflows, not rephrasings.

Anything else resolves to unknown and produces an honest refusal plus the questions the record can answer.

"Has this happened before?"

Worth calling out, because it is where Brok stops answering and starts investigating.

Ask it about a failure and Brok searches the evaluation record for earlier failures sharing an agent, prompt or dataset with this one. If it finds a precedent, it reads the failed runs of both and compares their recorded causes, then pulls in what the team decided afterwards and the engineering memory behind that decision.

"Yes — this has happened before. Checkout Quality #1 failed 19 days ago against the same agent and dataset. Both failures recorded the same cause: 'Connection refused'. After that failure, the team recorded: 'Moved the endpoint behind the internal gateway.'"

An identical recorded cause escalates the verdict from attention to risk, because a recurrence is a different engineering problem from a coincidence. If there is no precedent, it says so plainly and offers the diagnosis path instead.

The eight Engineering Briefs

Brok also writes standing readings of the record. Each follows: what happened → why → evidence → impact → recommendation → next action.

Daily · Deployment · Incident · Prompt · Evaluation · Dataset · Knowledge · Architecture

Actions, not advice

Every recommendation ends in a real surface. Brok never invents a destination — it hands you back into the platform at the exact place the answer came from:

Open the Forge Graph · Open the failure graph (already narrowed to the break) · View the execution graph · Compare revisions · Open AI Git · Open Engineering Intelligence · Open Evolution · Open Knowledge · Open the Registry · Open the evaluation · Open Analytics · Open Insights · Start an investigation

The workspace

Brok is a workspace you travel to, not a widget that follows you. Reachable from anywhere with Ctrl+. or the header, and from every artifact, the Registry, Knowledge, the Forge Graph, the Execution Graph, Evolution, AI Git and the dashboard — each carrying the context you were in.

The page is a conversation plus synchronized panels: engineering context, evidence, artifacts, evaluations, decisions, knowledge, AI Git revisions, engineering memory, and the Forge Graph. Clicking a referenced record or a graph node moves the whole workspace onto it.

While it works, it shows the investigation it is performing — Reading the engineering record → Resolving the subject → Searching for precedents → Composing the answer — never "Thinking…".

Why it has no language model

See Deterministic Engineering Reasoning for the full argument. The short version: an LLM answers "has this happened before?" fluently whether or not a precedent exists. A deterministic engine over real rows either finds it or says there is none — and that is the only kind of answer worth building a workflow on.

See also: Root Cause Explorer · Engineering Intelligence