How Manusights Compass works

The corpus: 1,321 venues

OpenAlex journals + curated CS conferences

Compass ranks venues from a fixed corpus of 1,321 entries: 1,291 academic journals and 30 hand-curated top-tier computer-science conferences. The journal set is drawn from OpenAlex and filtered to venues that have an ISSN, more than 500 published works, and more than 50,000 lifetime citations — a deliberately conservative cut that keeps the corpus to established, indexed venues rather than every entry in OpenAlex.

The 30 conferences (NeurIPS, CVPR, ACL, and peers) are added by hand because in computer science the conference, not the journal, is the primary publication venue, and OpenAlex does not model them the way it models journals. Compass treats journals and conferences as equal ranked candidates.

Each corpus entry carries a one-sentence scope description, up to five subfield tags, an impact-factor or citation-rate value, an acceptance-rate estimate where one is known, and DOAJ / MEDLINE indexing flags. The scope sentence is what the ranking model reads, and the verbatim scope text is what every fit explanation quotes back (see Ranking, below).

Impact-factor provenance

Verified JCR 2024 where available, OpenAlex proxy otherwise

Every venue in a Compass result shows an impact-factor value with an explicit source label, because the two sources are not interchangeable:

• jcr_2025 — a verified Clarivate Journal Citation Reports 2024 Journal Impact Factor. Used for the curated set of journals where we hold a confirmed value.
• openalex_proxy — a 2-year citation rate computed from OpenAlex, used for every other venue. It is directionally useful for venue selection but is not the official JCR number and should not be quoted as one.

We label the source rather than silently mixing the two because a proxy presented as a JCR Impact Factor is a factual error, and venue selection that turns on a tenth of a point off the wrong number is selection done badly. Treat either value as a rough tier signal, not a precise rank.

Ranking: one model call, explanation-first

Claude Sonnet 4.6

Compass does not embed-and-search. It sends the entire 1,321-venue corpus, your title, and your abstract to Claude Sonnet 4.6 in a single call and asks it to return the top five venues by scope fit. Each result carries a fit score (0–100), a tier (stretch / realistic / safe), a one-sentence reason, and a one-sentence “what to strengthen” note.

This is a deliberate choice. A vector search can pick a plausibly-relevant journal, but it cannot tell you why it fits or what the venue is looking for. The corpus is small enough to fit in one prompt, so we trade embedding-style precision for an explanation you can actually act on. The cost of that choice: Compass is only as good as its scope text and the model’s reading of it, and it is not tuned for sub-corpus semantic recall the way a dedicated retrieval system would be.

Auditable explanations

Every “why it fits” sentence is required to quote 3–6 verbatim words from the venue’s scope text, in single quotes (for example, matches the journal’s coverage of ‘single-cell RNA sequencing of immune cells’). Generic phrasing like “good fit for biomedical research” is rejected. The point is that you can check the claim against the corpus entry rather than taking the model’s word for it.

What the tier means

Tier is a realism signal, not a ranking. It answers “how hard is this venue for this paper?”, not “which is best?”

• Stretch. A top-tier outcome if everything goes right, but the acceptance and significance bar is high. Top-decile venues (Cell, Nature, NEJM, JACS, NeurIPS-class) default to stretch.
• Realistic. Strong fit, and the work plausibly meets the bar based on what the abstract claims.
• Safe. Clearly within scope and the paper exceeds the typical bar.

Confidence flags and refusal cases

• Low confidence (top score under 60). Compass surfaces an amber banner. A top score under 60 usually means the corpus does not cover the paper’s field well, most often machine learning and computer science beyond the 30 curated conferences, pure mathematics, social sciences, or earth sciences. The low score is itself the signal: it tells you to look outside the corpus rather than trust a forced match.
• No corpus match (422). If no venue clears a non-zero fit, Compass returns a scope disclosure instead of a list of zero-scored cards.
• Not an abstract (400). Marketing copy, placeholder text, gibberish, or a prompt-injection attempt is rejected with a clear error. A real abstract from an uncovered field is not treated this way — it gets honest low scores instead.
• Under 200 characters. The minimum-length gate is per-character (some pastes drop newlines), so paste the full abstract for signal.

What Compass does not do

• It does not assess scientific merit or novelty — only scope alignment and tier realism.
• It does not read your full manuscript, only the title and abstract you paste.
• It does not predict acceptance probability. Tier is editorial judgment from the model, not a calibrated probability, and we have not published an inter-rater reliability study against human editors.
• It does not know special-issue calls, editor relationships, or a venue’s current desk-rejection patterns. Treat results as one input, not a verdict.
• It cannot recommend a venue that is not in the corpus. Coverage is strongest in biomedicine, chemistry, physics, materials, and core computer science.