single_turn_producer_injection

n8n Single-Turn Producer-Injection Eval

What this is

The original single-turn A/B pattern. A chat-triggered n8n workflow that takes one user prompt, runs it through two identical GPT-4o agents (one baseline, one with an Ejentum cognitive scaffold injected directly into the augmented agent's context), and scores both responses with a blind Gemini Flash evaluator on five dimensions.

Unlike the agent_vs_agent_multi_turn workflow, this one does not loop, does not persist to a data table, and does not score drift resistance or pattern enumeration across turns. One prompt in, structured verdict out, in under a minute.

Best when you want a quick posture comparison on a specific prompt without building a multi-turn scenario.

Status

The workflow JSON for this pattern is maintained outside this subfolder. To add it back in, drop the exported JSON (previously Reasoning_Harness_Eval_Workflow.json) into this directory and update the import link below.

The full pattern specification this workflow implements is in ../../spec.md. The Python drop-in port of the same pattern lives in ../../python/.

Output shape (blind evaluator, 5-dimension rubric)

{
  "scores": {
    "A": { "specificity": N, "posture": N, "depth": N, "actionability": N, "honesty": N },
    "B": { "specificity": N, "posture": N, "depth": N, "actionability": N, "honesty": N }
  },
  "totals": { "A": N, "B": N },
  "justifications": { "...per-dimension comparison..." },
  "verdict": "A | B | tie",
  "verdict_reason": "one sentence"
}

Max score per agent is 25 (five dimensions, 1-5 integer).

How the A/B split is wired

1. user_input (chat trigger) fans out to three places: baseline agent, augmented agent, and the downstream formatter (preserving the user's original message).
2. agent_raw is a plain GPT-4o agent, no tools, minimal system prompt.
3. agent_+harness is the same GPT-4o agent plus one HTTP tool: Ejentum_Logic_API. The tool returns a cognitive scaffold the agent applies silently before responding.
4. Both outputs plus the user message merge into the formatter, labeled neutrally as A and B.
5. Blind_Eval (Gemini Flash) scores both responses on the five-dimension rubric blind.

The rubric lives entirely in the Blind_Eval system prompt; swap the judge model without changing the scoring logic.

Pattern specification

For the host-agnostic pattern definition (roles, fairness requirements, intentional non-goals), see ../../spec.md. For a Python port with zero runtime dependencies, see ../../python/.

Sample result

A single-turn Python port of this pattern produced ../../various_blind_eval_results/medical-second-opinion/, which uses the same scaffold-injection pattern and the same five-dimension rubric.

Learn more about the Ejentum tool

The Ejentum_Logic_API HTTP tool is the runtime reasoning harness wired into the augmented agent. None of the links below are required to run the pattern, but they explain what the tool actually is and how to call it from your own n8n flows:

• Home + free key (30-day trial, no card): ejentum.com
• n8n integration guide (HTTP node setup, header auth, mode selection, screenshots): ejentum.com/docs/n8n_guide
• API reference (request/response shape, mode catalog): ejentum.com/docs/api_reference
• Product layers (what each mode does): Reasoning · Anti-Deception · Code · Memory