The audit trail
How a brief is made.
Every summary you read passes through the same five-stage pipeline and is gated by a validator stack that checks evidence integrity and drug safety. This page is the canonical description. Nothing here is abstract. Every claim is measurable, and every number updates monthly.
01What we read.
Every 24 hours we ingest newly indexed papers from PubMed, plus the tables-of-contents of 47 ophthalmology-adjacent journals. Preprints from medRxiv and bioRxiv enter a separate lane.
If it is relevant to a practicing ophthalmologist, it is candidate material. Basic-science papers enter only when they read out on something clinically actionable within 24 months.
02The five-stage pipeline.
Every accepted paper passes through the same sequence. Reproducible, inspectable, versioned.
Ingest
Full text, figures, supplement. Registered in content-addressable store.
Classify
Subspecialty, study design, evidence level, clinical question vector.
Summarize
Structured extraction: design, PICO, results, limitations. Fixed rubric.
Validate
A validator stack checks evidence integrity and drug safety. Summary blocked if checks fail.
Publish
Delivered at 04:00 local. Archive is immutable; corrections append.
03Evidence level taxonomy.
Every brief carries one of five evidence levels. We use a hybrid of the Oxford CEBM 2011 framework and the GRADE hierarchy, simplified for reading speed.
Level 1
Randomized trial or systematic review of randomized trials.
Level 2
Prospective cohort or well-designed retrospective with registered protocol.
Level 3
Cross-sectional, case-control, or large observational registry studies.
Level 4
Case series, small retrospective chart reviews, early-phase devices.
Level 5
Expert opinion, editorials, guidelines narratives, framing commentary.
04The validators.
After a summary is generated, it is blocked until two independent checks pass. If either fails, the summary is discarded and the paper is flagged for editorial review. No fallback, no silent pass.
Evidence integrity
Every number, effect size, and result claim in the summary must be traceable to the source paper. Fabrication, transposition, and wrong-table errors are the explicit failure modes this check rules out.
Drug safety
Every drug name, dose, and route mentioned in the summary is cross-referenced against the source abstract. Hallucinated drug names, wrong dosages, and unmarked off-label use are the failure modes this check rules out.
Validator agreement rates and false-negative rates will be published after the first 90 days of live operation, calibrated against a held-out editorial-review sample. We do not publish simulated numbers.
05Ranking: why you see what you see.
The morning ten are chosen by a reader-specific ranker, not a universal top ten. The ranker is a learned function of four signals.
Subspecialty weight
The subspecialties you declared in onboarding, nudged by what you actually open.
Evidence prior
Level 1 papers outrank Level 5 all else equal.
Reading trajectory
An EWMA-weighted embedding of the papers you opened, saved, or marked. 30-day half-life. Never trained across users.
Active clinical question
Papers that address a question you have searched or saved recently get a small boost for up to 21 days.
No advertiser boost. No journal boost. No favored authors. We do not accept sponsored placement. Not in the brief, not in the archive, ever.
06Editorial oversight.
rounds. is an editorial product. Every summary passes the same pipeline, and a clinician-led editorial review sits on top of it.
Daily. A rotating sample of summaries is reviewed against the source paper before delivery. The review log is kept for 12 months and is available on request.
On report. Any summary a reader flags is re-reviewed the same working day and, if warranted, corrected with a dated appended note. Nothing is rewritten in silence.
Quarterly. The taxonomy, ranker weights, validator thresholds, and editorial style guide are reviewed. Material changes bump the methodology version.
rounds. is an editorial reading product for medical professionals. It is not a medical device, not a clinical decision-support system, and not a diagnostic tool. It does not replace clinical judgment, examination, or guideline-based care.
07Errors, corrections, and your role.
We will make mistakes. Our job is to make them rare, catchable, and fixable. Yours is to report them.
Every article has a Report this summary link. Corrections are published as an appended note with a dated diff. Nothing is rewritten in silence.
08What we do not do.
rounds. is not a clinical decision-support system. We do not recommend treatments for individual patients.
rounds. does not generate novel claims. Every sentence in every summary is tethered to a source.
rounds. does not sell your reading history, your subspecialty, your clinic, or your name. Our privacy notice is the same contract in legal form.