# Tileward: know what your AI knows. Decide what it's allowed to say.

![Tileward](https://tileward.com/logo.png "Tileward")

Tileward organizes what an AI model knows into named **tiles**, so you can see its
knowledge, lock away the parts that are off-limits, and keep a record of every
decision it makes.

- Models about half the size, same answers
- Locked topics stay locked, even when questions are rephrased
- Every decision recorded

## Why Tileward

Three problems come up the moment you put an AI model to work in a real organization:

1. **"Please don't talk about that" isn't a policy.** The usual way to keep an AI
   away from a sensitive topic is to ask it nicely in its instructions. Rephrase the
   question cleverly and it often answers anyway, and there's no record of what it
   decided or why.
2. **Small models make things up.** Compact, affordable models are fast, but when
   they don't know something they confidently invent an answer instead of saying so.
3. **Big models are expensive to run.** Full-size models need serious hardware. Most
   teams would take one that's half the size if the answers stayed just as good.

## How it works: map it, watch it, govern it

Once a model's knowledge is sorted into named tiles, you can treat knowledge the way
you treat rooms in a building: see who goes where, and lock the doors that matter.

1. **Map.** Tileward sorts what the model knows into named tiles (finance, medicine,
   software, and so on). Knowledge stops being a blur and gets an address.
2. **Watch.** Every question is matched to its tile before the model answers. You can
   see, in plain terms, which part of the model's knowledge each question touched.
3. **Govern.** Lock a tile, and questions about it are politely declined, even
   reworded ones. Every declined question is recorded, so you can prove the rule held.

## The model family

Each build ships compressed, governed, and ready to run on your own hardware. Served models are
callable today through the API; more are on the way. The live list, with what each model is for
and how it is served, is on the landing page: https://tileward.com/#models

## What you get (measured numbers)

- **~60x faster at saying no.** The usual approach has the model write a whole reply
  just to decide whether to refuse. Tileward decides before a single word is written,
  using about 6x less computing per decision.
- **1.9x smaller models, same answers.** Tileward's compression makes a model roughly
  half the size and about 1.4x quicker to respond, with answers that are nearly
  indistinguishable.
- **No made-up answers about your documents.** Point Tileward at your own files and
  it answers from what they actually say, quoting its sources, instead of guessing.
- **100% of decisions on the record.** Every conversation, every comparison, and
  every declined question is saved. Dashboards roll it up per person and across the
  whole team.

## Straight talk

Locking a tile doesn't delete knowledge from the model. Genuinely erasing something
from an AI model is an unsolved problem, and we'd rather tell you that than pretend
otherwise. What Tileward gives you is a guard that recognizes the topic even in
disguise, declines to discuss it, and writes every decision down. For "don't surface
this topic," that is the strongest tool available today, and unlike a promise buried
in instructions, you can audit it.

## The console

One web console — a dashboard, the chat and arena, and the tools to govern them, all in one place:

- **Dashboard**: how well the locks are holding, and what they're saving you.
- **Chat**: talk to the model; locked topics are declined, everything else just works.
- **Documents**: upload files and get answers drawn from them, with sources shown (powered by Twinkle, our context-memory engine).
- **Compare / Arena**: the full-size model and the half-size one side by side, and Tileward's lock vs. the "just ask nicely" approach, scored on the same test.
- **Tiles**: browse the map of what the model knows, and lock, unlock, or switch the guard on and off.
- **Audit**: every conversation, test, and declined question, saved and replayable.

**Twinkle** — the context-memory MCP server that hands your AI only the relevant slice of a
conversation — is both a standalone product ([twinkle.tileward.com](https://twinkle.tileward.com))
and the Tileward platform module behind Documents and grounding.

**Your domain, your tiles. Yours alone.** The map ships with 62 governable tiles plus a catch-all
(63 in all), covering finance, healthcare, law, safety, and more. If the part of your business you need to govern is not on it,
we build tiles to match. Bring your topics, the language your own policy uses, and the questions
you need caught, and we fit them to the map and tune them against your traffic. A tile we build
for you is **private to your account**: it sits on your roster and nobody else's, its name is
never shown to another customer, and only your questions are ever routed against it. Tell us
what you need to govern at [tileward.com/#contact](https://tileward.com/#contact).

## The API

The console is one way in. The other is an API. If your code already talks to the OpenAI
API, change the base URL and the key and governance comes with it: bind a policy to a key,
and every call is checked before the model runs. Or call the guard on its own, as a
checkpoint in front of whatever model you already use.

- **$1 per million tokens.** An OpenAI-compatible endpoint for the governed model,
  counting input plus output. Prepaid, no subscription, no surprise bill. New accounts
  start with a free credit, so you can try it before you pay for it.
- **$10 per million tokens, the guard.** A higher rate per token, and a much smaller bill
  per decision: the guard reads the question and nothing else, so a typical check is about
  14 tokens, or **$0.00014**. No policy to re-read, no context to carry, no verdict to
  write. You are billed for what it read, and it stops reading at 128 tokens.
- **Off-policy requests never reach the model.** A locked question returns a governed
  refusal, with no tokens billed for an answer that was never written. Off-scope
  questions, prompt injection, and jailbreak attempts are declined by the same gate the
  console uses, and every decision is logged.

Get an API key at [app.tileward.com/account](https://app.tileward.com/account).

## Plans

Tileward is in beta, so prices may change. Four tiers, each with the governance gate, the
compressed model, and Twinkle knowledge built in. Start free at
[app.tileward.com/account](https://app.tileward.com/account) and upgrade when you're ready.

- **Explore — free**: pay-as-you-go for real usage. 3 lockable tiles, 2 API keys, a 7-day audit
  window, CSV/JSON export, and a free credit to start.
- **Build — $39/mo** (or $390/yr, 2 months free): for a solo developer shipping one governed app.
  25 lockable tiles, 10 API keys, a 30-day audit window, webhook export, and Twinkle Pro.
- **Team — $299/mo** (or $2,990/yr): shared governance and seats. Unlimited lockable tiles, pooled
  seats, a year of audit retention, SSO, SIEM export, and VPC deployment.
- **Enterprise — custom**: a committed-use token pool shared across seats, or a flat air-gapped
  license with no metering. SSO + SCIM, unlimited private tiles, SIEM export, and an SLA.

## Token spend: the question is 1% of what a prompt-based guard reads

The usual way to govern an agent is to describe the rules to a model and ask it to judge
each request. The model does not remember your rules, so you resend them every time. One
decision, counted with the real tokenizer against our own 62-tile policy:

| Per decision | Prompt-based guard | Tileward |
| --- | --- | --- |
| Policy and topic list | 1,389 tokens | 0 |
| Conversation context | 60 tokens | 0 |
| The actual question | 14 tokens | 14 tokens (embedded) |
| Verdict the model writes | 28 tokens | 0 |
| **Total** | **1,491 tokens** | **14 tokens, 0 of them to a model** |
| **Cost** | **~$0.0015** (at $1 per million) | **~$0.00014** (at $10 per million) |

The question is **0.9%** of what the prompt-based guard reads. The other 99% is your own
policy, re-sent on every request.

**Ten times the price per token, ten times cheaper per decision.** The guard costs more per
token than the model does, and still bills roughly a tenth of what it costs to ask a model
the same question, because it reads about a hundredth of the tokens. Charging by the token
keeps it honest in both directions: a longer question costs proportionally more, and you are
billed for what the guard actually read rather than a flat fee per request.

**Straight talk:** providers that cache a repeated system prompt will discount those 1,389
tokens, and a policy smaller than our 62-tile one costs less to describe. But the rules
still have to be sent, still have to be read, and the model still has to write a verdict
before you learn whether the request was allowed. Tileward never sends them at all, because
the rules live in the gate instead of the prompt. That is also why it decides before a
single word is generated.

## FAQ

**Do I need to be technical to use it?** No. The console is point-and-click. Setting
up the system is a job for your technical team, but using and governing it isn't.

**Can you add tiles for my industry?** Yes, that's a conversation we want to have. The 62
governable tiles that ship cover common ground, but the topics you actually need to govern are usually
specific to your business: your products, your processes, the questions your regulator cares
about. Tell us what those are and we'll build tiles for them, seed them with the language
your policy uses, and tune them against real questions until they hold. Bring us around
fifteen real examples per topic, including some we don't train on, because those are how we
prove the tile catches your traffic.

**Are the tiles you build for me private?** Yes. A tile we build for you belongs to your
account: it sits on your roster and nobody else's, its name is never shown to another
customer, and only your questions are ever routed against it. Other customers keep the
standard map, unchanged.

**What if my tile doesn't work?** We tell you, and we don't ship it. Before a tile goes live
we measure it against the rest of your map: whether it catches its own questions, how often it
fires on unrelated ones, and whether it can be told apart from a tile you already have. If it
can't, we say which tile it collides with and why, rather than selling you something that
quietly starts refusing questions you already answer. A tile that can't be measured isn't a
tile, it's a guess.

**What happens when someone asks about a locked topic?** They get a polite,
clearly-marked refusal, and the attempt is recorded with a timestamp. Reworded and
disguised versions are caught too.

**Does locking a topic make the model worse at everything else?** No. The lock only
intercepts questions that match the locked tile; unrelated questions pass through
unchanged.

**Is the locked knowledge deleted?** No: it's access control, not deletion. The
knowledge stays inside the model; Tileward makes sure it doesn't come out, and keeps
the receipts.

**Where does my data live?** On your own machines. Models run locally, documents stay
in your corpus folder, history is stored in your own database.

**How is this different from writing rules into the AI's instructions?** Instructions
are a request; Tileward is a checkpoint. In head-to-head tests the instruction
approach was slower, cost about six times more per decision, over-blocked, and could
be talked around by rephrasing.

## Contact

Register interest in a demo or a conversation at https://tileward.com/#contact, or
email [hello@tileward.com](mailto:hello@tileward.com).