Evidence: PC-012 · Product Moat
1,219 Gates: How Constitutional Enforcement Works
Independent AI safety researcher, London. Verify any claim.
Every time an AI action runs in the Eden system, every Bash command, every file write, every git push, it passes through 1,219 deterministic gates. The verdict is ternary: Allow (+1), Investigate (0), or Deny (-1). Uncertainty never silently allows.
This is not a prompt. It is not a suggestion. It is enforced at the tool-call level. If a gate says Deny, the action is blocked. If a gate says Investigate, the action is held until a human resolves it. The audit trail is immutable.
Why Ternary?
Binary systems (Allow/Deny) force a false choice. Either something is safe or it is not. But the real world is not binary. Some actions are probably safe. Some are probably dangerous. Some need more context. A binary gate would either block too much (frustrating) or allow too much (unsafe).
Ternary logic, Allow, Investigate, Deny, maps the real uncertainty of AI actions. Investigate means "I cannot determine safety from the available patterns. A human must decide." This is the safety margin. In a binary system, Investigate collapses to either Allow (unsafe) or Deny (frustrating).
What the Gates Actually Block
This session alone: G-LANE-PR-BRANCH blocked 6 attempts to push to main. G-NO-CROSS-LANE-CHECKOUT blocked switching to another lane's branch. G-EM-DASH-FORBIDDEN caught em-dash usage in register edits. G-CITATION-VERBATIM blocked a handoff containing an unverified legal reference.
These are not failures of the system. They are the system working. A standard AI session would have pushed to main, broken the branch guard, and nobody would know until the next merge conflict.
The Architecture
The gate registry is a JSON file: config/legal-gates-enforcement.json. Each entry maps a pattern (regex) to a domain (legal, app_dev, os, shared) and an enforcement level (NUCLEAR, HIGH, STANDARD, AWARENESS). Adding a new gate is a config change, no code change needed.
The enforcement engine reads the action string, checks it against all applicable gates, and returns the most restrictive verdict. If any gate says Deny, the action is denied. If any gate says Investigate and none say Deny, the action is held. Only if all gates say Allow does the action proceed.
This is not complex. It is a pattern matcher with a decision rule. The complexity is in the patterns, 1,219 of them, covering legal citation validation, git safety, cross-lane isolation, data integrity, and more.
What This Means
The Eden system is not safer because we asked the AI to be safe. It is safer because the architecture makes unsafe actions impossible. This is the difference between alignment and enforcement. Alignment is a property of the model. Enforcement is a property of the system. The model can drift. The gates cannot.