Governance & Audit
Control which models run, with what context — and prove it
Governance you can enforce, not just hope for. Wardary applies the allow-list and org context server-side, then records every prompt in an immutable trail you fully own.
Model allow-list
Decide exactly which models your org may use
Choose the providers and models on your allow-list and whether end users may pick within it. A request for a disallowed model is rejected at the server — not merely hidden in the UI.
- OpenAI, Anthropic, and Google in v1
- Per-model configuration and enablement
- allow_user_model_selection toggle: users pick, or the org decides
- The allow-list encodes each provider's data-processing posture
Prepend context
Apply standing org instructions to every prompt
Inject your organization's guardrails and standing context into every outbound prompt as a system-level message — versioned, with each audit record referencing the version that applied.
- System-level injection, not user text
- Versioned, with a visible_to_users toggle
- Each prompt log references the context version used
- Enforced server-side on every request
One immutable record for every prompt
No prompt is ever un-logged. Each one writes exactly one append-only record — even on blocked or error paths — with immutability enforced at the database, not in middleware.
- Requested vs. resolved provider and model
- Which redactions applied, and why
- How the model was chosen (user, rule, or default)
- Tokens in/out, estimated cost, and latency
Total cost
$1,284
Total tokens
85.2M
Avg / 1K
$0.015
Top model
Claude
Usage & cost
See spend by user and by model
Every audit record carries token counts and an estimated cost — kept even under bring-your-own-key — so you can report usage by user and by model over time, understand spend, and show leadership that AI use is both governed and accountable.
Hard budget enforcement is on the roadmap; v1 meters and reports so the data is already there when it lands.
Contextual routing, when you need it
A rule-based routing engine — match a prompt by topic, keyword, regex, or tag and send it to a designated model, with the option to enforce that destination over user choice — is designed and sequenced as a post-validation release. We're candid that routing and logging are commoditized; redaction and audit are the value, so they ship first.
Resolution order
enforced rule → user selection → org default
An enforce rule is a guardrail that can't be overridden from the UI, and no rule can route to a disallowed model.