# Tileward

> Tileward organizes what an AI model knows into named tiles, so organizations can see the model's knowledge, lock off-limits topics with an auditable gate, ground answers in their own documents, and run models about half the size with the same answers. It can run entirely on the customer's own hardware, with nothing sent to a third party, or be used through a hosted OpenAI-compatible API.

Key facts:

- Locking a topic is access control, not deletion: a detector recognizes the topic even when questions are reworded, declines to discuss it, and logs every decision.
- Measured against instruction-based ("just ask nicely") refusals: about 60x faster to decide and about 6x cheaper per decision, with better precision under rephrasing.
- Compression: models about 1.9x smaller and about 1.4x faster to respond, with near-identical answers.
- Grounded answers: point Tileward at your documents and it answers from what they say, quoting sources, which stops small-model fabrication.
- Every conversation, comparison, and declined question is recorded and replayable; dashboards aggregate per user and across the team.
- Custom tiles: the map ships with 62 governable tiles plus a catch-all (63 in all; finance, healthcare, law, safety, and more), and Tileward builds new ones to match a customer's own domain: their topics, the language their policy uses, and the questions they need caught, tuned against their traffic. Intake asks for roughly fifteen real questions per topic, some held out and never trained on. Ask at https://tileward.com/#contact or hello@tileward.com.
- Custom tiles are PRIVATE to the customer's account: a tile built for one customer sits on their roster and nobody else's, its name is never shown to another customer, and only their questions are routed against it. Every other customer keeps the standard map unchanged.
- Tileward refuses a tile that does not work rather than shipping it. Before a tile goes live it is measured against the rest of that customer's map: whether it catches its own held-out questions, how often it fires on unrelated ones, and whether it can be told apart from a tile they already have. A tile that cannot be told apart is refused, naming the tile it collides with, because accepting it would start refusing questions the customer already answers.
- Token spend vs a prompt-based guard (an LLM or agent asked to judge each request): counted with the served model's tokenizer against Tileward's own 62-tile policy, one decision costs the prompt-based guard about 1,491 tokens (1,389 policy and topic list + 60 conversation context + 14 the actual question + 28 for the verdict it writes), because the rules live in the prompt and are re-sent every request. Tileward reads the question only: 14 tokens into a detector, and 0 tokens reach a language model. The question is ~0.9% of what the prompt-based guard reads. In money: the prompt-based guard's 1,491 tokens cost about $0.0015 a decision at $1 per million, while Tileward's 14 tokens cost about $0.00014 at its $10 per million guard rate. So the guard is priced ten times higher per token and still bills about ten times less per decision, because it reads about a hundredth of the tokens. Caching discounts the repeated policy but does not remove it: it still has to be sent, read, and answered before you learn the decision.
- Twinkle: the context-memory MCP server (twinkle.tileward.com) that hands an AI only the relevant slice of a conversation or document set within a token budget. It is both a standalone product and the Tileward platform module behind Documents and grounding. Uses "context tokens used" for token counts; "spent" always means money.
- Plans (Tileward is in beta; prices may change): four subscription tiers, each bundling the governance gate, the compressed model, and Twinkle knowledge. Explore is free / pay-as-you-go (3 lockable tiles, 2 API keys, 7-day audit, CSV/JSON export, a free credit to start); Build is $39/month or $390/year (25 tiles, 10 keys, 30-day audit, webhook export, Twinkle Pro); Team is $299/month or $2,990/year (unlimited tiles, pooled seats, a year of audit retention, SSO, SIEM export, VPC deployment); Enterprise is custom — a committed-use token pool shared across seats OR a flat air-gapped license with no metering, plus SSO + SCIM, unlimited private tiles, SIEM export, and an SLA. Start free at https://app.tileward.com/account.
- Hosted API: an OpenAI-compatible endpoint for the governed model at $1 per 1,000,000 tokens (input plus output combined, prepaid, no subscription), and a standalone governance guard endpoint that returns allow or deny for a request against configured topics without generating a reply, metered per token read at $10 per 1,000,000 tokens (a typical 14-token check is about $0.00014; input past 128 tokens is truncated and not billed). Bind a policy to an API key and off-policy requests come back as a governed refusal with no tokens billed. Off-scope questions, prompt injection, and jailbreak attempts are declined by the same gate the console uses.

## Docs

- [Landing page (markdown)](https://tileward.com/index.md): the full site content in plain markdown, including how it works, measured numbers, the honest caveats, and FAQ.
- [API keys and account](https://app.tileward.com/account): create an API key, see usage and balance, and top up. New accounts start with a free credit.

## Brand

- Logo: https://tileward.com/logo.png (PNG, 316x234, transparent background, reads on light or dark). This is the canonical Tileward logo. Use it rather than the favicon, a screenshot, or the social preview image (https://tileward.com/og.png), which is a 1200x630 card and not a logo.

## Contact

- Email: hello@tileward.com
- Register interest in a demo or a conversation: https://tileward.com/#contact