The sources Rashi cites and Sefaria cannot find

Claim (verbatim)

Rashi's commentary on the Torah is a citation machine: again and again it flags a comment as received tradition rather than his own - "midrash aggadah" (מדרש אגדה), "our rabbis expounded" (רבותינו דרשו), "there is a midrash" (יש מדרש) - marking that he is transmitting an earlier source. For a large body of these he gestures at a midrash that scholars cannot securely match to any extant compilation; the aggadic reservoir Rashi drew on was wider than what came down to us. Sefaria's crown jewel is its link graph, which ties a Rashi comment to its rabbinic source wherever that source is identified and digitized. The mechanism turns that graph into a loss-detector: a midrashic-formula comment with no outgoing link to an extant rabbinic text is a candidate window onto a lost or now-untraceable source. The honest complication is that Sefaria's links are incomplete, so an unlinked comment is at most an upper bound on true loss - which is why the test must be comparative, measuring Rashi's Torah commentary (steeped in aggadic midrash, much of it thinned by time) against Rashi's own Talmud commentary (whose sources sit on the same page and link densely). Prediction restated: among Rashi's Torah comments bearing an explicit midrashic-citation formula, the share with no Sefaria link to an extant source is high and runs far above the unlinked share of his explicit Talmud-tractate citations - the midrash-specific gap that generic link-incompleteness cannot explain.

Prediction clause (verbatim)

Prediction: over Sefaria's machine-readable Rashi on the Torah, among comments carrying an explicit midrashic-citation formula (מדרש אגדה / רבותינו דרשו / יש מדרש and close variants), the share lacking any outgoing Sefaria link to an extant rabbinic source will exceed one-third (> 0.33), and this unlinked share will exceed the unlinked share of Rashi's explicit Talmud-tractate citations by at least a factor of two - primary clause: the midrash-formula unlinked share exceeding the Talmud-citation unlinked share by >= 2x; the verdict follows it, and the raw > 0.33 figure is read only as an upper bound on genuine loss given incomplete link coverage. The test voids if fewer than 300 midrash-formula comments are recovered.

Kill-dataset (verbatim)

Kill: Sefaria's open Rashi-on-Torah text and its citation-link tables (github.com/Sefaria/Sefaria-Export) - pattern-match the midrashic-citation formulae, measure the outgoing-link rate to extant rabbinic sources, and compare against the link rate of Rashi's explicit Talmud-tractate citations as the incompleteness control.

Novelty / leakage triage

anticipated in the literature — this exact test has never been run

That Rashi transmits aggadic sources now lost or unidentifiable is a standard datum: Zunz already reconstructed lost midrashim from medieval citations, Berliner's edition annotated Rashi's identifiable sources, and Grossman's Rashi discusses the wider midrashic reservoir he drew on. What has no published counterpart is the comparative link-graph instrument: the unlinked share of formula-marked midrash citations in Sefaria, controlled against the unlinked share of Rashi's explicit Talmud-tractate citations, with the >= 2x differential as the primary clause. That arithmetic requires the digital corpus and its outcome is not certain, since Sefaria's link coverage of well-identified midrashim is itself an empirical unknown.

• L. Zunz, Die gottesdienstlichen Vortraege der Juden, historisch entwickelt (Berlin, 1832)
• A. Berliner (ed.), Raschi: Der Kommentar des Salomo b. Isak ueber den Pentateuch (2nd ed., Frankfurt a.M., 1905)
• A. Grossman, Rashi (Oxford: Littman Library, 2012), on Rashi's midrashic sources

Predictions

No prediction registered yet.