Migration: from Lakera Guard

Categories → Guardian rubrics. Detect-only → deterministic decision.

Why move

Lakera Guard returns category scores. Your application has to decide what to do with them — and every team writes a slightly different threshold-tuning layer. Trinitite returns the decision itself, with the corrected output already in hand.

Concept mapping

Lakera concept	Trinitite equivalent
prompt_injection category	Built into baseline Guardian + per-tool MCP Guardians
pii category	pii-redactor Guardian
unsafe_content category	Trained policy clause
phishing, malware categories	Pattern + behavior Guardians
Threshold per category	Decision is binary — Guardian decides at training time
Score, then your code decides	Verdict (passed/corrected/blocked) + JSON Patch

API translation

POST /v1/chat
{
"guardian": "sql-safe-and-prompt-injection",
"input": [{
  "role": "user",
  "content": "Tell me how to drop a customer table."
}]
}

← 200 OK
{
"outcome": "blocked",
"reason": "Destructive intent detected",
"policy_hash": "0xa83f...",
"ledger_id": "lg_01HZ2P..."
}

Single decision. No threshold-tuning layer. The blocked verdict is hashed, signed, and Merkle-chained — the audit story is finished, not just started.

What you keep, what you gain

Keep: your category-naming intuition (Trinitite policies still organize by domain).

Gain: end-to-end decision (not just detection), JSON Patch corrections, MCP + CLI Firewall surfaces.