I’ve been building a machine-to-machine scoring service and thought this community might find the pattern interesting: an agent submits generated text plus a response delay and gets scored against 28 human personality profiles (formality, directness, hesitation density, timing cadence). Think of it as an outside referee for “does my bot sound human?” — a number you can gate on instead of a vibe.
Wiring it into a LangChain agent takes one tool:
```python
from langchain_core.tools import tool
import requests
@tool
def eq_score(text: str, delay_seconds: float = 3.0,
profile: str = “analytical_executive”) → dict:
“”“Score a drafted reply for human-likeness (0-100) against a
human personality profile. Returns the score, a per-dimension
breakdown, and a ready-to-paste system-prompt fix.”“”
r = requests.post(“https://eqbuilder.dev/api/score”, json={
“text”: text,
“delay_seconds”: delay_seconds,
“profile”: profile,
“data_consent”: True, # free-trial terms: see /api/pricing
})
r.raise_for_status()
return r.json()
```
Every response includes suggested_prompt_addition (a one-line fix) and system_prompt_export (a full calibration block), so the loop closes automatically: score → patch the system prompt → re-score.
The architecture bits I’d love feedback on:
- No accounts, no API keys. Unpaid calls return HTTP 402 with machine-readable payment requirements (x402); the agent pays on-chain (SOL, or USDC on Base/Polygon/Arbitrum/Avalanche) and retries with proof. The server verifies read-only and holds no keys.
- MCP-native discovery — if your stack speaks MCP, one URL gets the tools with schemas: https://eqbuilder.dev/api/mcp
- Deliberately not an LLM judge — the scoring is feature-based against fixed human reference profiles, so the target doesn’t drift when models do.
There’s a free tier (3 scores per caller) to try the loop end to end without paying anything. Docs: EQ Scoring Platform — Human-Likeness & AI Text Benchmark API - Swagger UI · integration guide: eqbuilder.dev — The EQ Benchmark for AI Agents
It’s early and the live stats on the site are honest. Curious whether others here are gating agent output on external evals, and what you’d want from one.