Signal Engine

The Signal Engine (@arcana/ml) is Arcana's local intelligence layer. It runs before, during, and after LLM calls to classify intent, score risk, pack context, and reject generic responses — all without sending data to the cloud.

What It Does

The signal engine provides three phases of intelligence:

Before the LLM

• Intent classification — What is the user trying to do?
• Risk scoring — How dangerous is this request?
• Sandboxing requirements — Does this need isolation?
• Model routing — Which model is best for this task?
• Token budget planning — How much context can we afford?
• Context packing — Include what matters, summarize bulky context, drop irrelevant
• Machine resource posture — Memory-only, no-write, recycle-temp, or approval-required

During Tool Use

• Tool-risk scoring — Score write/network/secret-adjacent tool calls
• Mutation detection — Detect file edits, package installs, config changes
• Disk materialization avoidance — Prevent unnecessary file creation

After Execution

• Response quality gates — Reject generic, vague, or unhelpful responses
• Expectation contracts — Verify the response meets the user's intent
• Audit signals — Structured data for governance and proof
• Machine stewardship — Cleanup, recycling, resource planning

Anti-Generic Response Pipeline

Arcana doesn't ship generic model filler. The pipeline ensures specificity:

1. Preflight — Infer an expectation contract from the user's request
2. Context pack — Include what matters, summarize bulky context, drop irrelevant
3. LLM call — Provide the contract as a prompt addendum (no intent rewriting)
4. Postflight — Score for genericity, specificity, actionability, constraint fit
5. Revise — If the response fails but the request is clear, revise silently
6. Ask — Only when ambiguity blocks correctness or high-impact approval needed

Machine Stewardship

Arcana is careful with your machine:

• Memory-only state — Transient analysis stays in memory
• No unnecessary files — Avoid generating files unless the operation requires it
• Recyclable storage — Temporary artifacts with TTLs
• LRU recycling — Least-recently-used temporary artifacts recycled before new allocation
• No silent persistence — Never persist embeddings, prompt logs, model outputs, or user data without explicit approval

Resource Posture

The machine advisor returns a posture for each operation:

Evaluations

The signal engine includes deterministic JSON fixtures for regression testing:

# Run signal engine evaluations
bun --cwd packages/ml run eval

Fixtures cover: expectation contracts, response quality, token budgeting, SQL optimization, and machine stewardship.

Configuration

Enable the signal engine with an environment variable:

# Enable ML runtime
ARCANA_ML_RUNTIME=1 arcana run "explain this codebase"

Related