🛡️ The Third Author Governance (SCANOSS Earnie Steal)

Your code got a third author:

1. The Human Engineer (accountable, peer-reviewed).
2. Open Source (packages, snippets, license obligations).
3. AI Agents (faster than review, never learned your policy).

Earnie by SCANOSS (October 2026) introduced the paradigm of "One policy, three checkpoints" and "Four risks, one record". ThumbGate steals the architectural format and operationalizes it as an in-line, CPU-local, zero-token PreToolUse execution firewall.

🚦 One Policy, Three Checkpoints

1. Checkpoint 1: MCP / PreToolUse Gate (<0.5ms on CPU)
   • Intercepts proposed tool calls (Bash, terminal, run_command).
   • Blocks copyleft package installs (e.g. fast-charts [AGPL-3.0]) before execution.
   • Injects vetted, permissive alternatives (e.g. chart.js [MIT], recharts [MIT]) directly into the agent's context.
   • Blocks unpinned supply chain downloads (curl ... | bash).
2. Checkpoint 2: Git Pre-Commit Hook
   • Evaluates staged files against cryptography, licensing, and secret policies before commit.
   • Emits deterministic finding IDs (TG-LIC-XXXX, TG-PQC-XXXX).
   • Blocks commits on critical violations; developer runs npx thumbgate explain <FINDING-ID> to see evidence.
3. Checkpoint 3: PR Gate / CI Evaluator
   • Evaluates pull request diffs against organizational governance rulesets.
   • Emits definitive verdicts: PASS, WARN, REQUIRE_APPROVAL, BLOCK.

📦 Four Risks, One Record (The BOM Quadrant)

🛠️ CLI Commands & Usage

# Run full Third Author audit across all 3 checkpoints and 4 risks
npx thumbgate third-author

# Run audit in machine-readable JSON
npx thumbgate third-author --json

# Inspect any finding with raw evidence and permissive alternatives
npx thumbgate explain <FINDING-ID>

# Run PreToolUse gate check manually
echo '{"tool_name": "bash", "tool_input": {"command": "npm i fast-charts@4.2.1"}}' | npx thumbgate gate-check

🚫 Hard Boundaries (Honesty & Compliance)

• Never vendor SCANOSS SaaS: Do not call scanoss.com APIs, install scanoss-py, or send code fingerprints over the network.
• Local-First Ground Truth: All evaluations run on CPU in <0.5ms using local pattern registries and AST detectors.
• Fail Closed: Any tool call attempting to introduce unvetted viral copyleft or broken classical cryptography into production is blocked before execution.