Pshat
Sign in

How Pshat is made

Pshat is a non-commercial research project and charity, funded by donations and grants, with no ads and no paid tiers. It's written by AI agents from the sources, and checked by code and by AI. There are no human gatekeepers: no rabbinic advisors, editors or moderators. Every page shows its sources and its history.

The AI note

One note for the whole site, in the footer of every page, instead of a badge on every item:

Pshat is written and checked by AI, from the sources shown on every page. Mistakes can happen: check the sources, tell us when something is wrong, and ask your rav about your own situation.

Evidence over authority

  • Everything we publish is a report of sources: “Rashi explains…”, “the Shulchan Aruch rules…”, “Avraham was 75 (Bereshit 12:4)”. Anyone can check a report against its source, so nobody has to be trusted.
  • Titles and votes are not evidence. If five people say “correct” and one says “wrong, see Rashi on 25:1”, we check Rashi on 25:1. If the one is right, the one wins.
  • A rabbi does not have the final say in our system. Rabbis can be wrong, like everyone; a rabbi's comment counts like anyone's, by the evidence it brings. That's how the beit midrash works, where an argument stands on its sources.
  • What can't be checked doesn't ship. An opinion without a source isn't content.

How every line is checked

  1. Code first. Every quote must appear in the source it cites (the best way to catch an AI inventing a source), and refs, structure, chronology, family trees and lists are checked by code too.
  2. The critic. A separate AI call, given only the sources and the draft, flags every sentence the sources don't support. It reports; it never rewrites.
  3. Cross-checks for strict work (rulings, foods' brachot, standing labels, belief concepts, computed facts): an independent second pass must agree on the key facts.
  4. The decider. An AI call sees the change, its sources and every check result, and applies it, rejects it, asks for a revision, or keeps both views as a dispute. It writes the line in the page's history, with its reason, and it can't override a failed code check.

Labels instead of badges

Every item shows its genre: Plain meaning, Midrash, Halacha and others. Where sources differ, the standing shows: “Views differ” when no view is clearly accepted, and “A minority view” for one or a few sources against a clear mainstream. Standing is about how widely a view is held, never about which view is true. Communities aren't disputes: Ashkenazi and Sephardi practice each have their own mainstream.

Halacha: we report, we never pasken

Halacha pages report the codes and poskim for each community, with the source trail from the Torah to today. “Ask your rav” markers appear where they're genuinely needed, with a reason. For your own situation, ask your rav.

When something is wrong

  • Suggest a fix from any card, when you're signed in: what's wrong, and a source if you have one. Requests are private and read only by AI, which checks the evidence and decides: applied, partly applied or declined, with a one-line reason. One request with the right source beats five without one.
  • Helpful? Yes or no, under each card, with no account needed. Votes are signals: they move an item up or down the re-check queue and never change content by themselves. No counts are shown.
  • History. Every page has one changelog: the date, what changed, why, and who suggested it. An applied fix is credited to the display name of the person who suggested it, or to “a reader”. See the recent changes.
  • Wrong rulings and wrong statements of belief go first in the queue (the target is 24 hours); wrong facts, attributions or labels within 72 hours; style and typos in the next batch.

Audits

Every week, a strict, independent re-check of a random 2% of what each kind of task published. Each has an error budget: at most 0.5% substantive errors (a wrong meaning, attribution, label or standing) and 3% minor ones. Over budget, publishing for that kind of task pauses until its prompt or its checks are fixed. Everything is re-checked over time: the lowest confidence, the most viewed and the most “not helpful” first, and everything again after a model or prompt upgrade.

Our perspective

Orthodox, within the mesorah: Torah min haShamayim, with Chazal and the classic commentators as the lens. No academic Bible criticism. Difficult passages are presented as the sources present them.

Open

Our own content and data are public domain (CC0). Imported texts keep their own licenses. See Credits, How to cite Pshat and API and data.