Jev vs citation services
Citation services (search APIs, academic indexes, browser agents) find or format references. Jev judges a pair (claim, source span) with a typed question. It will not invent a DOI.
Unofficial comparison. Official method: citation check cookbook — one Choice on whether the quote’s context supports the claim; confidence flags human review. We do not restate their accuracy figures. Not affiliated; no keys.
Cite vs generate boundary
| Step | Owner |
|---|---|
| Propose a quote / URL | Retriever, user, or LLM |
| Normalize the span | Your parser |
| “Does this source support this claim?” | Jev Choice or Noul |
| Write the footnote string | Template / LLM |
| Show it to the user | Only if code says support ≥ τ and confidence ≥ floor |
If you skip the judge, you have a generator with decorations. If you skip the generator, you have a support label — often enough for RAG filters.
Failure modes
- Claim and source not both in
state— the model cannot fetch the web. - Dates and numbers in the claim — extract in code; Jev is weak at numeric identity (jaggedness).
- Low-confidence “supports” presented as verified — treat as handoff-shaped review.
- Schema-safe Choice ≠ the source is true in the world.
When not to use Jev
Do not ask Jev to mint a citation. That is generation: inventing a title, DOI, page, or URL. TypeSafe’s jaggedness note for jev-1.13 lists generation as a failure mode — do not chain Choices to write the footnote string.
Do not send Jev a claim without the span. Crossref, Semantic Scholar, your index, or a browser agent must produce the candidate first. Jev has no browse tool.
Do not treat a support label as a truth label. A span can support a false claim that the source itself asserts. Schema-safe ≠ correct.
Skip Jev when the job is only formatting (“APA this DOI”) or only discovery (“find papers about X”). Those are citation-service jobs.
When citation services win
Indexes and search APIs win at recall of candidates: they know the corpus, mint identifiers, and resolve metadata. Jev does not store a bibliography and will not return a DOI that was not in state.
A citation manager wins at style, dedup, and export. Keep BibTeX and CSL out of the judge.
Pin jev-1.13.0 after you fit τ and the confidence floor. jev-latest and the marketing line jev-1.13 can move. TypeSafe’s published list price for jev-1.13 is $0.042 per million input tokens (vendor claim — confirm on the models page); output tokens are free on that same page.
Pairing-specific failure modes
- Date/number identity. “Q3 revenue was $4.2M” vs a table cell — extract the number in code; ask Jev only whether the surrounding sentence is about that metric.
- Truncated span. A quote that omits the hedge (“…in mice”) will look supportive. Normalize enough context that a human could also judge.
- Alias drift. Floors fitted on
jev-1.13.0silently move if you shipjev-latest. - Presented-as-verified. Shipping a footnote whenever the Choice key is
supports, ignoringconfidence, is how fake citations survive a “we added Jev” review.
Production sketch
- Retriever or user proposes
{claim, url, span}. - Parser stores the span; reject empty or over-long dumps (jaggedness: large irrelevant state).
- POST one Choice: support / not-support / unclear. Pin
jev-1.13.0. Logresponse.model. - Code, not the model, applies both floors:
# Compose rule: support and confidence are two axes. Fit both on labeled pairs.
SUPPORT = {"supports"}
if key in SUPPORT and conf >= CONF_FLOOR:
render_footnote(template.format(url=url, span=span))
else:
queue_review(claim, span, key, conf)- The footnote string is a template. An LLM may polish punctuation; it must not invent the identifier.
Hub: Comparisons. Siblings: RAG pipelines, retrieval-only search.
FAQ
Can Jev browse? Not by itself. Feed text.
Is this plagiarism detection? No. It is support / not-support / review on spans you already have.
Does a high-confidence “supports” mean the paper is true? No. It means the span you fed is consistent with the claim. World-truth stays outside the schema.
Should I generate the citation with Jev? No. Mint identifiers with a citation service; judge support with Jev.
Disclaimer
This is an independent unofficial site and is not affiliated with TypeSafe AI; official documentation is available at https://docs.typesafe.ai. Never treat jev.pro as TypeSafe official documentation. We do not sell, issue, or proxy API keys.