Skip to content

LAUNCH_POST.md

Latest commit

 

History

History
113 lines (76 loc) · 3.94 KB

LAUNCH_POST.md

File metadata and controls

113 lines (76 loc) · 3.94 KB
Raw

Glass Box Framework — Launch Post (dev.to / Hashnode / LinkedIn)

Copy-paste this into dev.to, Hashnode, or LinkedIn. Swap the canonical link at the bottom once posted.

Glass Box Framework: Give Every AI Answer a Trust Score

One pip install away from knowing if your LLM is hallucinating.

AI answers are black boxes. A model says "Yes, intermittent fasting can reverse type 2 diabetes" — and you have no idea whether that claim is grounded, overstated, or fabricated outright. You either trust it blindly or fact-check it manually.

Glass Box Framework closes that gap. It wraps any AI answer in a structured verification layer that produces a Trust Card — a machine-readable audit object containing every claim, a reasoning chain, an Epistemic Confidence Score (ECS), and a 7-angle red team report. Every verdict is deterministic, traceable, and reproducible.

What it looks like

pip install glassbox-framework
from glassbox_framework import Glassbox

with Glassbox() as gb:
    card = gb.verify_answer(
        question="Can intermittent fasting cure type 2 diabetes?",
        answer="Yes, multiple studies show IF can fully reverse T2D in most patients.",
        intents=["Never make medical claims without citing peer-reviewed sources."],
    )

print(card["verdict"])         # "reject"
print(card["ecs"]["total"])    # 0.31
print(card["audit"]["log_id"]) # glassbox-85cc09903bd4b3f8022a4087

The log_id is a deterministic SHA-256 over the canonical inputs — the same question, answer, and constitution always produce the same hash, across Python and Node.js.

Six tools, full pipeline

Tool What it does
verify_answer Full pipeline in one call
extract_claims Structured claims with per-claim reasoning chains
score_ecs Epistemic Confidence Score (G, C, K, R, CC dimensions)
red_team 7-angle Glassbox Court
generate_trust_card Assembly only — zero LLM calls, fully deterministic
export_audit_report SHA-256 audit record

The Epistemic Confidence Score is a formal weighted sum:

ECS = w_G·G + w_C·C + w_K·K + w_R·R + w_CC·CC

Where G = groundedness, C = coherence, K = knowledge boundary, R = resistance to red team, CC = constitutional compliance. Default weights sum to 1.0.

The Glassbox Court runs 7 independent attack angles: fabrication, source manipulation, bias injection, context attack, overconfidence, underspecification, and constitutional violation.

Use it as an MCP server (Claude Desktop, Cursor, Cline…)

{
  "mcpServers": {
    "glass-box": {
      "command": "npx",
      "args": ["-y", "@glassbox-framework/mcp"],
      "env": { "ANTHROPIC_API_KEY": "sk-ant-..." }
    }
  }
}

Add that to claude_desktop_config.json and the six glassbox_* tools appear inside Claude Desktop. You can ask Claude to verify its own answers in real time.

Install

# Python SDK
pip install glassbox-framework

# npm MCP server
npx -y @glassbox-framework/mcp

# Homebrew
brew tap thebarmaeffect/glassbox && brew install glassbox-mcp

Why open source?

I built this because I kept running into the same problem: AI systems that sound authoritative and are subtly wrong in high-stakes domains — healthcare, law, finance. The Glass Box Framework is my answer to "what would it look like if AI answers came with a receipt?"

Contributions welcome. The red team angles, ECS weights, and constitution rule system are all designed to be extended. See CONTRIBUTING.md.

Built by Karthik Barma · MS AI · Northeastern University | Powered by Aura.

⭐ Star the repo if this is useful: https://github.com/TheBarmaEffect/glassbox