Methodology · v6.3← rounds

2-minute read · the audit trail

How a brief is made.

PubMed indexes thousands of papers a day. Maybe ten of them matter to your subspecialty. Maybe one changes how you see a patient on Monday.

This page is how we find those ten, how we check ourselves before we ship them, and how you can check our work.

126

Journal slots (36 unique)

Bias flags per paper

Validator gates

Ranker signals

Briefed per Pro reader, daily (3 on Free)

Why this page exists

Most toolsMost tools rewrite the abstract in different words.
rounds.We extract structured data, then compose a fixed five-field summary checked against eight deterministic gates.
Most toolsMost tools optimize for engagement (more time on page, more clicks).
rounds.We optimize for whether you would cite the paper in clinic. Topics that already appeared in your week are penalized.
Most toolsMost tools hide the source behind their own UI.
rounds.Every brief points back to the source paper. One click, no paywall, no rewrite layered on top.
Most toolsMost tools ingest invented numbers when the upstream model hallucinates.
rounds.Every number, dose, and drug name in a card must trace back to a value extracted from the source paper. If it cannot, the card is regenerated.

Our promises to you

01
Every number, drug name, and dose traces to the source paper.
Eight deterministic gates check each summary before it ships. If a gate fails, the card is rewritten or falls back to a deterministic template, marked as such.
02
No advertiser pays to put a paper in your brief.
Source authority (society guidelines, top-tier journals) carries a small positive weight; preprints carry a small negative one. That is editorial, not commercial.
03
You decide. We summarize.
rounds. is an editorial reading product, not clinical decision support. Every brief points back to the full source paper, one click away.
04
Your reading history is not a marketing segment.
Profile data lives only in your account. We do not sell it. Privacy notice.

What we read.

Every 24 hours we ingest newly indexed papers from PubMed and MEDLINE, along with registered trials from ClinicalTrials.gov. Before a record enters the pipeline it has to clear the ingest filter: animal-only studies and known predatory journals are dropped, non-research item types are excluded, and duplicates are merged on DOI or PMID. We curate a reference list of 36 core ophthalmology journals across the 10 subspecialties (126 subspecialty slots, since the flagship titles recur across several); subspecialty relevance itself is decided by the classifier, not by a journal allow-list.

The 10 subspecialties

·Cataract & Refractive
·Cornea & External Disease
·Glaucoma
·Medical Retina
·Surgical Retina & Vitreous
·Uveitis & Ocular Immunology
·Neuro-Ophthalmology
·Oculoplastics & Orbit
·Pediatric Ophthalmology & Strabismus
·Ocular Oncology

Basic-science papers enter only when they read out on something clinically actionable within 24 months. Retracted papers are filtered before they reach the ranker. Preprint sources (medRxiv, bioRxiv) are post-launch roadmap.

How a paper gets through.

Seven stages on a fixed daily schedule. Each writes its own audit trail. Reproducible, inspectable, versioned.

01Ingest

02Filter

03Classify

04Grade

05Rank

06Summarize

07Deliver

01: IngestDaily fetch from PubMed and ClinicalTrials.gov.
02: FilterIdentifier deduplication, predatory-journal and non-human exclusion, paper-type filter.
03: ClassifySubspecialty, study design, evidence level, and the clinical question the paper addresses.
04: GradeBias-flag inventory and composite quality score from extracted study metadata.
05: RankPer-reader ranker over fourteen signals, with diversity and a cold-start formula.
06: SummarizeStructured generation gated by eight deterministic validators.
07: DeliverWeb brief, article pages, and email brief.

How we judge a paper.

Two scores travel with every paper. An evidence level on the CEBM 1 to 5 hierarchy (RCT outranks cohort outranks editorial), and an 18-flag bias inventory across six clinical axes. Both are deterministic. No language model decides whether a study is biased.

Evidence levels

L1
RCT / Meta. Randomized trial or systematic review of randomized trials.
L2
Cohort. Prospective cohort, or well-designed retrospective with registered protocol.
L3
Cross-sectional. Cross-sectional, case-control, or large observational registry.
L4
Case series. Case series, small retrospective chart reviews, early-phase devices.
L5
Editorial. Expert opinion, editorials, guideline narratives, framing commentary.

Bias inventory · 18 flags across six axes

Sample size and power

·Small sample size (under 50 patients)
·Underpowered: fewer than 100 patients with a non-significant primary outcome
·High loss to follow-up (over 20%)

Study design

·Retrospective cohort or case-control
·RCT without placebo, sham, or vehicle comparator
·Unblinded patient or clinician assessment
·RCT without intention-to-treat analysis
·Observational design without confounder adjustment

Pre-registration and reporting

·RCT without a trial registration (NCT or EudraCT)
·Many p-values reported, or primary outcome changed after registration
·No quantitative effect size reported
·Dropout reported without a missing-data discussion

Outcome integrity

·Surrogate endpoint used as the primary outcome
·Post-hoc or exploratory subgroup analysis

Independence

·Industry-funded with disclosed support
·No conflict-of-interest section detected

Center and follow-up

·Single-center study
·Short follow-up (under 6 months chronic, under 1 month acute)

Every flag is deterministic. It fires when the structured study data we extract or a targeted text scan of the abstract says so. No language model judges the paper.

How we summarize it.

Every summary is the same five-field shape. A headline. A why-now line. A take-away. A one-sentence limitation. A one-sentence prior-evidence framing. The summaries are automated and physician-designed. That is what you see, every weekday.

Sample · this is what one looks like

Faricimab cuts injection burden 30% versus aflibercept in nAMD.

Why now: Multi-center RCT (n=1,329) showed 16-week dosing achievable in 79% of eyes at year 2.
Take-away: For treatment-naive nAMD, prefer faricimab over aflibercept; expect about 30% fewer injections per year.
Limitations: Two-year horizon; long-term durability unproven.
Prior: Extends TENAYA / LUCERNE 1-year results to year 2.

Eight deterministic gates check each card before delivery. If a gate fails, the card is rewritten. If rewriting also fails, the card falls back to a deterministic template populated from the extracted source rows; that card carries an Auto-template badge so you see which tier produced what you read.

The eight gates · what each one catches

01
SchemaOff-shape output. Wrong word counts, missing fields, or fields filled in the wrong order.
02
Forbidden phraseGeneric openers, hedging filler, and promotional language. A high-severity match sends the summary back to be rewritten with the failing phrase named.
03
Editorial qualityHeadlines that paraphrase the article title. Why-now lines without a number or comparator. Take-aways without a decision verb.
04
Fact-checkNumbers, percentages, effect sizes, drug names, or device names that do not trace back to a value extracted from the source paper.
05
Drug safetyDrug names or dosages that do not appear in the source abstract. Treated as fabricated; the summary is regenerated.
06
Device safetyDevice or implant names that do not appear in the source abstract. Treated as fabricated, like drug names; the summary is regenerated.
07
Observational guardCausal verbs (caused, reduced, prevented) on an observational study, where the design supports association rather than cause.
08
Number policyNumbers typed into the prose, or a statistic referenced with no value behind it. Every figure has to arrive through the deterministic source-traced block, never the model’s free text.

We don't publish simulated numbers. From launch, validator agreement and the correction rate per thousand summaries are tracked, and published once enough live operation has accrued to be meaningful.

Why you see what you see.

The morning ten are chosen by a reader-specific ranker. Fourteen weighted signals along four axes. No advertiser boost. No paid placement. Ever.

Personal fit

Your reading profile (a continuously updated picture of the papers you have rated as relevant), your declared subspecialties, the topics and drugs you read about often, the patient populations you treat, and a saved clinical question if you have one.

Evidence quality

A composite newsworthiness score (society guidelines, safety-relevant retractions, large high-powered trials, and replication signals) carries the highest single weight. Then the L4 quality grade, recent society guidelines, the CEBM evidence level, and source authority.

Diversity

Topics that already appeared in your brief in the last fourteen days are penalized so the week reads as a window onto the field, not a loop. Papers that look different from what you have read are favored.

Recency

Exponential decay with a one-week half-life since the paper was indexed. Older papers are not excluded; they just have to be more relevant than fresh ones to make the cut.

Reading profile. Your profile moves toward the papers you mark as relevant and gently away from the ones you mark as not relevant. The more you rate, the faster it adapts. Opens, saves, and shares are recorded but do not yet move your profile. Never trained across users.

Active clinical question. Save a question on your profile and matching papers get a positive weight for twenty-one days. After that the boost decays to zero until you refresh it.

First week. In your first seven days, or until you have rated at least five papers, the ranker leans on your declared subspecialty, paper quality, and your active question rather than a profile it has not had time to learn.

Source authority weighting (society guidelines, top-tier journals) carries a small positive weight; preprints carry a small negative one. That is editorial weighting of source quality, not a commercial relationship. Exact weights are versioned and audit-logged; every change bumps the version.

When something is wrong.

We will be wrong sometimes. Our job is to make it rare, catchable, and fixable. Every article carries a Report this summary link. Reports are reviewed as soon as we can, typically within a few working days.

How a correction lands

01
Editorial sampling.
Every card clears eight automated validator gates before it ships. From launch, a rotating sample of shipped cards is additionally reviewed by a clinician against the source, with each outcome (pass, rewrite, correct) logged.
02
Reader reports.
Reports go to the same queue. Response window is one working day.
03
Confirmed corrections.
The card is regenerated. The article page renders a Corrections block under the brief showing the date, the field that changed, and the before/after diff. The original text is preserved in the audit log, not deleted. Nothing is rewritten in silence.
04
Quarterly review.
Subspecialty taxonomy, the curated journal list, ranker weights, validator thresholds, and editorial style guide are reviewed every quarter. Material changes bump the methodology version.

What rounds. is

rounds. is an editorial reading product for medical professionals. Not a medical device. Not a clinical decision-support system. Not a diagnostic tool. It does not replace clinical judgment, examination, or guideline-based care.