Deterministic Engineering Reasoning

Broks Forge has a reasoning layer — Brok and the Root Cause Explorer — and it contains no language model.

That surprises people, so this page explains the reasoning behind it. It is the most consequential architectural decision in the platform.

The problem with a generative answer

Suppose you ask an LLM-backed assistant, wired to your engineering data: "Has this failure happened before?"

It will answer. It will answer fluently, in your terminology, with the right shape. And it will answer just as fluently whether or not a precedent exists — because generating a plausible answer is what the model does. The failure mode is not that it is wrong; it is that a wrong answer is indistinguishable from a right one.

For an engineering partner, that is disqualifying. The whole value of asking "has this happened before?" is that a no is as actionable as a yes. An answer you have to verify by hand has saved you nothing.

What deterministic means here

Brok resolves your question to an engineering intent by explicit weighted phrase scoring, then composes the answer from database rows through a single, narrow data channel.

   Your question
        │
        ▼
   Intent resolution        deterministic phrase scoring,
   (25 real intents)        explicit weights, a threshold
        │
        ▼
   One read of the          agents, prompts, datasets, providers,
   engineering record       evaluations, runs, derived knowledge
        │
        ▼
   Answer composition       every sentence built from fields
                            that exist in that snapshot

The record snapshot is the only channel through which content can enter an answer. There is no second path — no free-text generation step, no template filled by a model.

This makes fabrication structurally impossible rather than discouraged. It is not a prompt instruction that says "do not hallucinate"; it is an architecture in which there is nothing to hallucinate with.

What you gain

Refusal is reliable. Ask something outside the record and you get an explicit "the engineering record cannot answer that" plus the questions it can answer. The refusal is a fact about the record, not a mood.

Answers are reproducible. The same question against the same record gives the same answer. You can cite an answer in a review.

Every claim is checkable. Each statement carries the records it was read from. Nothing is a summary you have to trust.

No data leaves. Answering a question requires no external API call, so your engineering record is not sent to a model provider to be reasoned about.

It is free and instant. No tokens, no rate limits, no latency budget for reasoning.

What you give up, honestly

This is a real trade, and pretending otherwise would be exactly the kind of claim this platform exists to avoid.

• Brok understands engineering intents, not arbitrary language. A question phrased far outside

the 25 supported intents resolves to unknown and is refused, even if the record could in principle answer it.

• No open-ended synthesis. It will not write you an essay comparing your architecture to

industry practice. It answers engineering questions about your record.

• Coverage grows by engineering. Supporting a new question means adding an intent and a handler,

not writing a better prompt.

We consider that trade clearly worth it for a partner whose answers are meant to be relied on. A tool that is right 90% of the time in an unpredictable 10% is worse than a tool with a smaller, known boundary.

Where inference is allowed — and labelled

Determinism does not mean the platform only states the obvious. It draws causal readings too. The difference is that they are marked.

Paired with a three-step confidence ladder — consistent with, likely, near-certain — instead of a fabricated percentage.

So when the Root Cause Explorer says a prompt promoted two hours before a failure is a related change, it labels that inferred and states plainly: proximity is not causation. The platform is allowed to reason; it is not allowed to pretend a reading is a fact.

Could a model be added later?

It could, in one specific place: mapping unusual phrasings onto the existing intents. That would widen the front door without touching the answer path, so answers would stay composed from real rows.

What will not change is the guarantee. Content in an answer comes from the engineering record, or it does not appear.

See also: Brok · Root Cause Explorer · Engineering Intelligence