Update 6 — Analytics: What Your Corpus Can Tell You About Itself

Metadata

Update number: 6
Update type: feature
Date: 2026-08-19
Version: 1.3.0 → 1.4.0
Scope: analytics, reference exports, Electron security

Summary

resmon has always recorded a great deal about every paper it finds — the source, the DOI, the authors, the publication date, the subject categories, whether the paper was new to you at the time, and the moment resmon itself first saw it. Until now it did almost nothing with any of it: a list, a count, and a Markdown report.

This release adds an Analytics page that answers four questions from data already on your disk, and BibTeX / RIS / CSV export so the papers a sweep finds can go straight into a reference manager.

Nothing on the Analytics page queries a repository. Opening it costs no API quota and works offline.

Motivation

Two of the questions here are ones a researcher cannot easily answer any other way.

Which of your repositories are actually earning their place? resmon can query fifteen sources. Some of them will be re-reporting papers that another source already gave you. Every one of those costs time and rate-limit budget on every single sweep, and contributes nothing. Nobody can tell you which, because it depends entirely on what you search for.

How long does each source take to surface a paper? This one resmon is unusually placed to answer. Because it stamps first_seen_at itself, it knows how long each repository took to deliver a paper to you after that paper was published. No repository publishes this figure about itself. It has a direct practical consequence: a source with a two-week median will not deliver anything sooner because a routine runs hourly.

Changes

The Analytics page

Which sources earn their place. Per repository, papers that nothing else found, against papers that also arrived from somewhere else. Papers are identified across sources by DOI, falling back to a normalized title, because the same paper arriving from arXiv and OpenAlex is two rows by design. A source with almost no unique contribution is a candidate for deselecting.
How quickly each source surfaces a paper. Median, fastest, and slowest days between publication and resmon first seeing it. Negative lags are kept rather than discarded — some sources date a paper later than they post it, and dropping those would flatter the source.
Routine health. New results per run, per routine, with a sparkline and an explicit count of consecutive runs that found nothing. A routine marked Quiet has returned nothing new for several runs, which usually means its keywords are too narrow or its field has gone still.
Publication volume over time. Papers per month, stacked by source or by subject category, with a legend and a table view of the same figures. A paper in two categories is counted in both. An undated paper stays in your corpus totals but is left off the timeline rather than assigned a guessed date.

On figures that say “not enough data yet”

This is the part you will notice, so it is worth being explicit about the rule.

Counts are always shown. “Fourteen papers from arXiv” is true whether your corpus holds fourteen papers or fourteen thousand, so there is no reason to withhold it.

Averages and percentages are held back until they mean something. A median of three numbers is not a finding, it is three numbers. Below the threshold resmon shows the count and what it still needs — “3 (needs 5)” — rather than printing a confident median computed from almost nothing.

A percentage of nothing is undefined, not zero. DOI coverage on an empty corpus shows an em dash. Rendering it as “0%” would be a claim about your data rather than an admission that there isn’t any yet.

The same applies to routines: one quiet run is not evidence that a routine is finished, so resmon will not call a routine healthy or quiet until it has three completed runs. Until then it says “Too early to tell” and states how many runs it needs.

The thresholds are deliberately low — five dated papers for a lag median, three runs for a routine verdict. They exist to stop a single data point being presented as a trend, not to withhold information from a small corpus. Every figure also carries the sample size it was computed from, so nothing is hidden.

On an empty corpus the page does not draw empty charts. It explains what it will show once you have run a search, and links to Deep Dive and Deep Sweep. An axis with no data on it reads as breakage, not as emptiness.

Reference exports

Results & Logs gains BibTeX, RIS and CSV buttons alongside the existing export. These give you the papers, in the formats reference managers read, rather than the report about them:

BibTeX — LaTeX, Overleaf, JabRef, and Zotero’s importer.
RIS — EndNote, Papers, Mendeley, and most publisher sites.
CSV — a spreadsheet, or your own scripts.

Selecting several runs produces one file. Papers with a DOI are exported as journal articles and those without — usually preprints — as generic entries, because claiming a venue resmon does not know would be worse than leaving it out. Cite keys are made unique within a file, BibTeX special characters are escaped including inside URLs, and abstracts are collapsed to a single line, since multi-line values break several importers and RIS outright.

Security

The Electron window no longer disables the browser’s same-origin policy. That flag dated from when the interface was loaded from a file:// URL and could not otherwise reach its own backend; the interface has since been served by a local HTTP server, which made the flag unnecessary. Removing it was verified two ways: a cross-origin request succeeded in a browser with the policy enforced, and the running app was observed issuing proper CORS preflight requests — which a browser only sends when the policy is in force.

Verification

	Before	After
Backend tests	404	423
Renderer tests	13	21
CI	3.10, 3.11, 3.12 green	unchanged, still green
Analytics endpoints	—	6
Export formats	Markdown, PDF, LaTeX	+ BibTeX, RIS, CSV

The analytics tests treat the empty and thin corpus as first-class cases rather than edge cases, because that is the state every new install is in.

Follow-ups

The Analytics page has no tutorial video yet; its entry shows the standard placeholder.
Analytics are computed on each page load with no caching. On a few thousand papers this is instant. On a corpus a hundred times larger the author count walks every row, and that will want revisiting.
The corpus-wide faceted explorer and cross-execution comparison described in the Phase 2 plan are not in this release.