Prompt Guard core workflows

Three recipes: scan a prompt pack, respond to a fail vs. a warn, and check cost by surface.

Overview

Prompt Guard's guardrail scan, eval loop, and cost log combine into a small number of recurring jobs before and after a batch generation run.

•
Workflow 1 — clear a prompt pack before a batch run
Point Guardrails at the prompt or prompt pack, run the scan, and read the result: a fail blocks the batch outright, a warn surfaces for a human look before you proceed.
•
Workflow 2 — investigate a regression after a prompt or rule change
After editing a prompt or a guardrail rule, run the eval loop's golden-dataset regression pass, and review anything it flags before the next batch goes out.
•
Workflow 3 — check what a pipeline actually costs
Open the cost log and review spend broken down by surface and model, rather than assuming a batch job's cost from memory.

A deterministic scan that catches leaked secrets, banned model defaults, and cross-brand leakage before generation runs.
A golden-dataset eval loop that re-runs automatically when a prompt or rule changes.
Sampled judge/human review on a set cadence, layered on top of the deterministic pass.
Per-surface, per-model cost logging for every generation.

Guardrails checks whether a prompt is safe and well-formed — it does not judge whether the generated output is actually good.
Judge-based evals are sampled, not run on every generation; most coverage comes from the deterministic golden-dataset pass.
This is built for content moving through Creytix's own generation pipelines, not a general-purpose firewall for third-party AI tools.