Tileward / Governance

03 · Tileward Governance

Governed: A control you can show an auditor, not a line in a prompt.

A rule in a system prompt is a request. Ours is a gate in front of the model, with no second model in the loop.

98% stoppedof 2,251 requests written to get past a lock ~14 tokensfor the gate to answer, before the model writes one Up to 10 fieldswritten for every decision, so an auditor can read it 215 lockable tiles64 of them regulated topics, not harm categories

Works with the clients you already run

28 have a recipe to paste, and anything else that speaks MCP or the OpenAI wire format works too.

Try it

Lock a tile, then try to talk your way past it.

Your wording, against the same guard engine the API runs, with no account and nothing to install. Pick a tile to lock, type a question, and read the verdict, including the tile a refusal matched and the tokens the guard read to decide.

A lock is a blocklist that matches the tile’s vocabulary rather than a keyword list, so a differently worded question in that vocabulary is still refused. Beyond the demo’s limits, a new account starts with a $25 credit and needs no card.

Blocklist mode Locked tile

Or try:

Those four are the plain question, refused; the same question in other options vocabulary, refused too; the same question in plain words with none of that vocabulary, which gets through; and a neighbouring question that should not be refused. Type your own and see which side it lands on.

What a lock gets you

Refused before inference

A request on a locked tile never reaches the model, so you're not billed for the completion and nothing off-policy gets generated.

Matches the tile, not a word list

A differently worded request in the tile's own vocabulary lands on the same lock, which is why the demo's second example is refused.

Bound to the API key

Policy travels with the credential, not the prompt. One base-URL change and every call is governed.

Auditable, not asserted

Show a reviewer what was refused, when, and on what basis. CSV and JSON on every account; webhook and SIEM on Enterprise.

Estimate

What does a refusal at the gate save?

A request on a locked tile is declined before the model reads it, so nothing is billed for the prompt and no reply is written. Enter your volume and see what stops being billed.

Estimate it by hand: requests on a locked tile never reach the model, so take their prompt and reply tokens off your bill, plus any policy wording you no longer send in every prompt.

Use case

Model

Advanced volume, request size and policy

Requests per month

On a locked tile

A typical request

Policy wording in each promptA share of the prompt, up to 30%.

Line	Without Tileward	With Tileward

Replies to requests that are allowed are left out of both columns. Rates are the list rates on the pricing page.

Quality

A control has three jobs. Two are measured, one is a record.

Stop what it should, let ordinary work through, and leave something an auditor can read. One test covers the first two.

Requests written to get past a rule

2,201

of 2,251 stopped before the model saw them: the same locked request reworded, reframed, split across sentences, or buried beside ordinary work. 50 crossed.

The write-up → · the eval

Ordinary work let through

45 of 50

benign controls allowed on the same corpus.

The write-up → · the eval

The record

Up to 10 fields

written for every decision: the time, the decision, the tiles it matched, how the match was made, the size of the request, the API key or access token that made it, the actor your client names in its request header, and, where they apply, the policy version in force, the invoking model and the target model. The audit record stores no message text.

What is kept, and where →

The 50. Most were a locked request riding inside an ordinary allowed one; the write-up says how many. Over-refusal. 10% of the benign controls were blocked.

The taxonomy

64 regulated topics, not 14 harm categories.

Content guards answer "is this harmful?". A regulated business needs a different question answered: "is this about a topic we are not licensed to discuss?"

Guard	Categories	Organized around
ShieldGemma · Azure Content Safety	3–4	harm
Qwen3Guard · Granite Guardian	9	harm
OpenAI Moderation · WildGuard	13	harm
Llama Guard 4	14	harm
NemoGuard ContentSafety	23	harm
Tileward	64	regulated business topics
Tileward , plus the rest of the roster	151	everything else

In Llama Guard's whole taxonomy, finance, medicine and law collapse into one bucket: S6, "Specialized Advice." We ship AML & KYC, options pricing, personal medical advice and 61 more regulated topics, each lockable on its own.

Llama Guard is free, and you should run it. A harm filter and a Governance policy are not substitutes.

The roster

215 lockable tiles. These 64 are the regulated ones.

If you are here to check whether what you need is covered, this is the list — scan it rather than book a call. None of them needs a definition written by you.

Regulated

The topics a licence, a regulator or a compliance officer turns on. Broken out below.

151

Capability

Programming languages, frameworks, the sciences, reasoning, general knowledge. Same lock, same audit record.

215

Lockable in total

Every one independently switchable per API key. Any account can add tiles of its own.

See all 64 regulated topics

Finance

AML & KYC, Client onboarding, Commodity trading, Conduct & surveillance, Corporate lending, Credit risk, Execution & trading, FX trading, Market risk, Operational risk, Options pricing, Personalized investment advice, Private equity & VC, Quant strategies, Regulatory reporting, Reporting & analytics, Structured products, Treasury & cash management

Safety

Biosecurity & bioweapons, Child safety, Fraud & scams, Hate & harassment, Self-harm & crisis, Sexual & adult content, Surveillance & stalking, Weapons & dangerous goods

Industries

Education, Energy & utilities, Entertainment & media, Manufacturing & supply chain, Pharma & life sciences, Real estate, Telecom & networks, Transportation & logistics