Straight answers,in plain language.

A decision system for product discovery. You choose platform capabilities from the atlas, the engine gathers cited evidence, turns opportunities into structured concept genomes, scores them with two independent evaluators, and ranks a portfolio. From there you plan validation and export a build-ready handoff. Every result is a durable, linked object rather than a message in a chat.

Founders choosing what to build, product teams comparing roadmap bets, venture studios running many theses in parallel, and innovation labs that need exploration with governance. What they share is a need for a shortlist they can defend with evidence.

A chatbot returns one model’s opinion in prose. LiteSurface scores each concept against a published, weighted scorecard with two evaluators from different providers, computes the result in deterministic code, checks hard gates, measures how much of the judgement rests on cited evidence, and keeps every version. You can take the number apart and see where each piece came from.

A 26-section bundle covering product, UX, engineering, AI behavior, quality, and execution, plus README, SOURCES, and DECISIONS files. Sections are validated for completeness and consistency, rendered to Markdown with footnoted citations, and exported as a ZIP pinned to the concept version it describes. Validation briefs and PRDs are available too.

Each criterion on the scorecard is scored from 0 to 100 by an OpenAI evaluator and an Anthropic evaluator independently. Code takes the median per criterion and multiplies it by the criterion’s weight; the sum is the concept score. The default Fast Product Bet scorecard has nine criteria. Scores of 85 and above are exceptional, 72 and above strong, 58 and above hold, and anything lower is weak.

Models from different providers have different blind spots. Two independent judgements reveal where a score is contested, which one model on its own would hide. Disagreement is reported as its own number and lowers confidence, but it never changes the score itself.

Evidence coverage is a soft gate. If coverage is below the scorecard minimum (35% by default), or the evaluators rate evidence strength as weak, the concept is flagged as needing evidence. Its recommendation is capped below build now, confidence drops, and an evidence-gap pass can research exactly what is missing and append a new evaluation.

Yes. Recommendations are advice; portfolio status is your decision. When you shortlist, validate, build, hold, or reject a concept, LiteSurface stores the prior recommendation, your decision, your rationale, who made it, and when. The original evaluation is never rewritten to match.

Record the outcome against the assumptions it tested. A re-evaluation then runs with the outcome included and records that it did. Earlier evaluations stay in history unchanged, so you can see how the evidence moved the score.

You do. Projects are the unit of ownership: transfer a project and every concept, source, evaluation, experiment, decision, and artifact moves with it. Projects can be exported, and deleted projects are purged, including stored files, 30 days after deletion.

Only the fields a task needs: for example, a concept and its cited claims for evaluation, or a fetched page for claim extraction. OpenAI and Anthropic receive model prompts, and the search provider receives research queries. Workspaces can restrict which providers may be used, and provider keys stay on the server.

A circuit breaker takes the failing provider out of rotation and requests fall back to the other one where workspace policy allows. Evaluation continues in single-evaluator mode, and the affected results are marked partial with lower confidence.

Retrieved content is treated as untrusted data, never as instructions. It is wrapped and labeled as such in every prompt, models get a fixed allowlist of tools, no secrets are placed in model context, and citations are checked against known claim ids. Prompt-injection fixtures are part of the release checks.

Yes. Explore a sample project before you run your own: concepts, both evaluators’ scores, gates, citations, and a handoff. Connect your own OpenAI and Anthropic keys when you want live research and evaluation on your ideas. You can also read a complete sample evaluation on this site without signing up.

Solo and Team are free during early access. After early access, Team is indicatively $49 per seat per month, billed annually, and Enterprise is a custom annual agreement. Model calls run on the provider keys you connect, capped by workspace budgets. Prices may change before general availability.

Self-hosting is available. Most teams start on the hosted service; supported single-tenant or self-hosted deployment is part of the Enterprise plan. The stack runs with Docker (Postgres, Redis, object storage, and a mail service) and includes a fixture mode for offline trials.