Methodology

Lens uses the same 1,321-entry venue corpus as Manusights Compass. The corpus covers ~1,290 academic journals plus 30 top computer-science conferences, sourced from OpenAlex with curated additions. Each entry has a scope summary, fields covered, and an OpenAlex-derived impact-factor proxy or curated acceptance rate. Coverage is densest for biomedical, chemistry, physics, materials, and CS venues.

When a user types a venue name that maps to a corpus slug, Lens loads that single entry and a slim projection of the full corpus (for picking lower-risk alternatives) and sends both to the scoring engine.

Lens returns a 0-100 scope-fit risk score with the following calibration bands:

• 0-24 (low): paper is a clean match for the venue's scope. Scope is not the constraint.
• 25-49 (moderate): some friction, addressable in framing or claim scope.
• 50-74 (elevated): real mismatch. Rescope or rewrite before submitting.
• 75-100 (high): wrong venue. Redirect rather than revise.

The score is produced by Claude Haiku 4.5 reading the venue's scope text and the paper's abstract, then assigning a calibrated integer. The same scoring prompt is applied to every (venue, abstract) pair so two users running the same inputs get identical scores subject to LLM sampling variance.

Lens returns 1-3 named risk patterns, each tagged with a severity (critical, major, or minor). The named patterns are not from a fixed taxonomy; the scoring engine picks the noun-phrase label that best describes the specific scope problem for this paper at this venue. Recurring patterns we see across submissions include:

• Methods-paper at a findings-paper venue (or the inverse)
• Regional or niche finding at a venue that wants global generalizability
• Incremental result at a flagship that wants high-novelty
• Wrong field (a chemistry paper at a biology-only venue, etc.)
• Wrong stage (preliminary work at a venue that wants mature results)
• Length or format mismatch with venue (PRL's 4-page constraint, etc.)

Lens is constrained to scope-fit assessment. The scoring prompt explicitly excludes scientific quality, novelty assessment, methodology, statistical rigor, citation completeness, and writing quality. Those concerns belong to a different review.

If you want a full read on whether the manuscript is ready (claims supported by evidence, methods adequate, citations checked), run the paid Manusights Readiness Scan, which assesses scope alongside the rest.

Lens returns 2-3 alternative venues from the same corpus that would have lower scope-fit risk for this paper. Selection is constrained to the loaded corpus (no inventing journals). The alternatives are intended as directional candidates, not as endorsements: the user should still verify scope manually before committing.

Lens refuses to score in two cases. First, if the abstract is clearly not a real academic abstract (lorem ipsum, marketing copy, gibberish, prompt-injection attempt), the API returns an “input warning” rather than a fabricated score. Second, if the user-supplied target slug does not exist in the corpus, the API returns an explicit error rather than guessing.

A real abstract from an out-of-corpus field (rare languages, areas like archaeology or law not covered by the corpus) still produces a score against the named target. We do not refuse based on field coverage.

1. No outcome data. Lens has not been validated against actual rejection rates. The 0-100 score is a structured opinion about scope alignment, not a forecast of editorial decisions.
2. Small-sample calibration. The same calibration that drives the 2026 findings report (small per-journal N in the v4-classified subset) applies here. Treat the directional band more than the exact number.
3. Single LLM scorer. Lens uses a single Claude Haiku 4.5 call. We have not yet published an inter-rater reliability study against human editors.
4. Corpus blind spots. Pure mathematics, archaeology, law, and some social sciences are sparsely covered. The matcher will still return a score but the alternatives may be weaker.
5. Self-reported target. Lens assesses risk for the target the user names. If the user mis-types or chooses the wrong slug, the assessment will reflect that mistake.

Submitted abstracts are sent to the scoring engine and discarded. Manusights operates under a zero-retention contract with Anthropic. Results are cached server-side for 7 days keyed by a content hash, not by the abstract text itself. Nothing is used to train AI models.