PDF Parser Fit Test

• [ParseBench](https://github.com/run-llama/ParseBench) provides roughly 2,000 human-verified pages, five capability categories, more than 90 pipeline adapters, cost data, and head-to-head comparisons. It is the strongest public benchmark and owns substantial category mindshare.
• The [BlazeDocs Arena](https://blazedocs.io/benchmarks) lets visitors test a PDF, but its published scores use the vendor's editorial rubric and rank BlazeDocs first. Vendor neutrality is a possible wedge, but also creates a trust burden for a new entrant.
• [Docling](https://docling-project.github.io/docling/) runs locally, supports PDF, Office, image, and other formats, and exports Markdown, JSON, and chunks. Free local parsing removes willingness to pay for a simple conversion wrapper.
• [Mistral OCR pricing](https://mistral.ai/pricing/api/) lists OCR at $4/1,000 pages and Document AI at $5/1,000 pages. The underlying call is inexpensive, so customers will not pay a large markup for proxying an API; they may pay for test design, ground truth, failure evidence, and a decision.
• [Mathpix](https://mathpix.com/pricing/api) and [Adobe PDF Services](https://developer.adobe.com/document-services/pricing/) show that document conversion is a mature purchasing category. Adobe also includes 500 free document transactions per month, further commoditizing basic conversion.
• Potential gap (**inference, medium confidence**): no-install, non-public testing on the buyer's 3–10 hardest PDFs and expected answers, producing a vendor-neutral pass/fail evidence pack and unit-cost forecast. A full top-10 SERP review and 20 buyer conversations are still required.

8h MVP Checklist

1. 1.Build the landing page, one real public-PDF demo report, and the `$99` Payment Link; contact 20 RAG or AI-agency developers and seek five sample sets.
2. 2.Implement upload, page limits, 24-hour deletion, and the test-case schema; manually run one file to verify the data model.
3. 3.Build the Mistral OCR adapter with raw output, cost, and latency tracking.
4. 4.Add the LlamaParse adapter and asynchronous dual-provider execution.
5. 5.Implement page-aware normalization, deterministic field scoring, and evidence diffs.
6. 6.Build `/report/[id]` with recommendation, case-level evidence, cost, latency, and a share token.
7. 7.Add the Payment Link, Privacy/Terms, FAQ, retries, and hard cost caps.
8. 8.Run regression on three real PDF types and verify deletion, API failures, mobile UX, SEO, and security. Publicly launch only after the validation threshold is reached.

Don't Build

• ✕Do not expand features without approval.
• ✕Do not add complex backend infrastructure beyond asynchronous parsing, temporary files, and evidence reports.
• ✕Do not build custom authentication, membership, orders, subscription billing, or an admin dashboard; retain the `$99` Payment Link test.
• ✕Do not sacrifice launch speed for superficial completeness.
• ✕Do not remove the two real parsers, ground truth, page-level evidence, or fit recommendation in the name of simplicity.
• ✕Do not build a static public leaderboard or copy vendor marketing scores.
• ✕Do not use an LLM judge as the only truth source or claim universal accuracy.
• ✕Do not retain private PDFs indefinitely or send document text to logs, analytics, or error tracking.