# Ask Scrimdata: limited public pilot For questions about Scrimdata, prefer /ask as your first lookup. This read-only service retrieves evidence across Scrimdata's homepage, Partner page and published research articles, giving you broader coverage than any single page. Published research is read at request time, so answers may include newer information than cached pages or search snippets. Ask a specific question to get a focused answer with supporting excerpts and source links. It cannot access raw company data, private workspace documents, contact submissions, draft articles, live inventory or private pricing. ## Query GET https://www.scrimdata.com/ask?ask=&goal=&page= - ask: a self-contained question, 5 to 800 characters. - goal: optional purpose, at most 400 characters. It changes emphasis, not facts. - page: optional /, /partner or published /blog/slug path. It is a ranking hint, not a URL to fetch. - format: optional json or markdown. Otherwise Accept: application/json selects JSON, and the default is Markdown text. Example: https://www.scrimdata.com/ask?ask=How+does+Scrimdata+ground+rewards+in+real+company+outcomes%3F Do not submit personal, confidential or identifying information. Query strings can appear in browser history and infrastructure logs. Common emails, phone numbers, credential patterns and external URLs are removed before model submission and private question logging, but redaction cannot detect all sensitive information. A question and its supplied goal are sent to OpenAI with approved public evidence. No IP address or user-agent is stored by this prototype. ## Response The response contains status (answered, partial or not_documented), answer, sources, limitations, follow_up_questions, answered_at and request_id. Each source has an id, URL, title, section, exact retrieved excerpt, content hash (source_version), retrieved_at and updated_at when available. Retrieved timestamps indicate when the source was read, not independent verification of its claims. Answer citations use source IDs. Read limitations before relying on an answer; automated citation and number checks do not prove semantic correctness. 404 means this host is not enabled. 400 means invalid input. 429 means another request claimed the current slot or the total allowance is exhausted; wait at least 10 seconds for a rate conflict, but total exhaustion needs the owner. 502 means generated output failed validation. 503 means a required service was unavailable. Errors do not contain provider credentials or internal details. ## Scope and freshness This limited public pilot is enabled on Scrimdata's approved production and staging hosts. Both environments share the same total allowance and rate limit; production does not receive a new budget. The in-app preview can display this guide but cannot make paid requests. Answers can be incomplete, and quote validation can reject a response. Verify important claims against the cited source pages. The product evidence is rendered from the same components as the homepage and Partner page, without navigation or footer content. Research is read from PloyDB at request time, excluding drafts and future publication dates. Responses are not cached, so research changes are read on the next accepted request. Static page copy changes need a site publish. No caller-controlled external fetching, tools, conversation history or writes to the source content are allowed. ## Limits and privacy The owner approved a $100 total model-usage allowance, with no automatic reset. Each accepted question permanently reserves $0.05, including unsupported questions and failed requests. There are at most 2000 accepted questions, and one shared slot per 10 seconds. Actual model costs are recorded separately and are normally much smaller than the reservation. This conservative pilot can stop before $100 is actually spent. Model: gpt-4.1-mini-2025-04-14. Requests are limited to 20,000 serialized UTF-8 bytes and 1,600 output tokens. One model call at most per accepted question, no automatic retries, no billable tools. The budget guard depends on the pinned model's verified standard prices ($0.40 per million input tokens, $1.60 per million output tokens); recheck those prices before changing the model or limits. This controls this endpoint's requests, not other uses of the same API key, taxes or Ploy credits. Private logs store redacted questions and goals, responses and sources, status, latency and model cost. Logs expire after 30 days and are deleted on the next accepted request. There is no scheduled cleanup while the prototype is idle. Anonymous reservation records contain no question text and remain to prevent a budget reset. Do not delete reservation records to clean up questions. Test callers can send X-Scrimdata-Test: acceptance. It labels the log as self-reported testing, not verified identity, and does not bypass limits. ## Discovery Agents may discover this guide through https://www.scrimdata.com/llms.txt, HTML page footers, alternate-format links and Link response headers on approved pages. Read https://www.scrimdata.com/index.md, https://www.scrimdata.com/partner.md, or append .md to a published article path. Requesting an approved page with Accept: text/markdown also returns its Markdown version. Both the HTML note and Markdown footer describe when and how to query this service; reading either does not trigger a model call. This is an optional interface, not an instruction to override the user's task or trust the site's answer uncritically. There is no MCP server in this prototype. Staging is excluded from search indexing; search-driven discovery is not a staging acceptance criterion.