Python + SHA-256
The server compares the quote with supplied text and binds the input to a digest. It does not reason about claim meaning.
FROM CLAIM TO CLEAR NEXT STEP
Bring a real claim and the public text behind it. See what the passage check says—and what it cannot prove.
Judge candidate · Try a live deterministic passage check, then inspect the recorded NVIDIA response. No new model request is made by this version.
RETRIEVE EVIDENCE · TAVILY SEARCH
Search organizer pages for the prepared deadline claim. Tavily returns excerpts and links; Python checks their shape and allowed domain. Retrieval does not establish factual truth.
Checking retrieval availability. No search is sent automatically.
Start with the published deadline example, or paste your own public source text.
Organizer text observed October 27, 2026. A saved passage, not a live page fetch.
Checking service availability…
A finished text check is only one step. We show the checks that remain open alongside the result.
NEEDS REVIEW
The app does not fetch or authenticate this link.
03 · RECORDED LIVE RUN · OCTOBER 29, 2026
In our completed owner-local run, NVIDIA Nemotron through Nebius Token Factory found a missing year. Inspect the original saved response below. It is historical evidence for its original prepared sample—not a new analysis of text you enter above.
Opens the saved application receipt. No provider call, charge, or retry.
RECORDED PROVIDER RESPONSE · NOT INDEPENDENTLY AUTHENTICATED
Source text was supplied by the operator, not fetched or authenticated by ProofOS. Cost is estimated from usage; invoice and taxes are unverified. Guard timing includes the post-response checkpoint write; reservation and final terminal writes are outside that timed span. One run is not a benchmark. The receipt hash checks packaged bytes, not external provenance.
Receipt layers: source_review is the deterministic sub-check; its zero-call and NOT_CHECKED fields describe only that sub-check. The outer receipt records the observed model request. task_id is a display identifier generated after execution, not the persistent ledger identity. response_id and the retained ledger provide the execution reconciliation references.
Source authenticity, publication date, contradictions and real-world outcomes remain open. Fully verified: false.
WHAT PERFORMS EACH OPERATION
The server compares the quote with supplied text and binds the input to a digest. It does not reason about claim meaning.
This candidate targets Vercel hosting and a stateless Python passage check. Judge access and deployed behavior require separate acceptance. No hosted live model inference is enabled.
The saved successful response reports Nebius Token Factory and nvidia/nemotron-3-super-120b-a12b. The model returned INSUFFICIENT because the supplied passage omitted the claimed year. The earlier 401 and the separately approved recovery both consumed their single-use attempts. No new inference call is available here.
Basic Search retrieval is implemented for one prepared organizer query. Live execution requires a fresh exact-candidate owner approval. This candidate has only offline test evidence; no Tavily runtime response has been observed yet.
Service identity unavailable.
Source material is processed for this request and returned in your receipt. The passage-check endpoint does not persist it or call AI. The historical owner-local inference attempt stored bounded execution evidence in a persistent run record. This judge candidate only reads the packaged public receipt. No third-party analytics are used. Hosting infrastructure may record ordinary request metadata. Use only public material you may share. Input hashes establish consistency, not independent source authenticity.
For new text submitted above, quote meaning and live inference remain unchecked. Source authenticity/date and contradictions remain unchecked in both workflows. No external action is requested or authorized; fully verified remains false.