Every number on this site, and how it was computed.

These are not aspirations. They are enforced in the build scripts that produce the data files this site reads, and every one of them costs us a number we would otherwise get to display.

1. 1Published sources only. Every input is either a published dataset, a peer-reviewed paper, or data that applicants and programs reported themselves. Nothing is simulated, imputed across years, or filled in to make a page look complete.
2. 2Effect sizes with intervals, not p-values. Each estimate ships with a confidence interval. p-values are not used anywhere on this site. An interval tells you how much the number could move; a p-value only tells you whether someone should be surprised.
3. 3Small samples get pulled toward the mean. A program with nine reports does not get to claim a 90 percent signal conversion rate. Estimates are shrunk toward the specialty average with a weight set by sample size, so thin cells regress instead of shouting.
4. 4A null result is reported as a null result. Where a credential's interval crosses no effect, the site says "no measurable independent effect" and the number does not move. Across every specialty and stratum modeled, 145 of 322 credential estimates (45 percent) come back that way, and all of them are labeled.
5. 5Observational data is never called causal. Applicants who signal a program differ from applicants who do not, in ways no dataset records. Every lift on this site is an association measured under those conditions. The word "causes" does not appear in any model output.

Six source families feed everything on the site. Each one is used for the questions it can actually answer, and none of them is asked to carry a claim it cannot support.

Ophthalmology does not run through the main match algorithm, so running it through the main model would be wrong in a way that is invisible to the user. It gets a separate engine.

The central question of the modern application: does signaling this program actually change your odds of an interview there? For every program in 23 specialties, the model estimates the invite rate for applicants who signaled and for applicants who did not, and reports the difference.

A signal does not act alone. Applicants who signal a program are also more likely to have trained near it or rotated there, and if you attribute all of that to the signal you will overestimate what a signal buys a stranger.

How much does one more first-author publication, or an honor society election, independently change your odds of an interview, holding the program and everything else fixed? This is the most heavily defended model on the site, because it is the one people most want to over-read.

How many people applied, how many were interviewed, how many seats exist. Two paths, tried in order, and the page tells you which one produced the number.

Whether rotating at a program changes your odds of an interview there. The honest answer has three different qualities depending on what a given program has published, so the model carries three states and labels which one you are looking at.

Away-rotation decisions are not published by anyone. The only way this data exists is that applicants maintain it themselves, which makes provenance and hygiene the entire methodology.

Your portfolio is not a list of independent coin flips, so the plan is simulated rather than summed.

Model coverage is uneven, and pretending otherwise would be the easiest way to mislead you. This table is generated from the model files themselves at build time, so it cannot drift from what the site is actually serving.

Signaled and unsignaled are the stratification-eligible row counts behind each specialty’s lift model: rows with a classifiable applicant type and a known Step 2 band. Credential rows are the applicant-by-program rows in that specialty’s credential fit. Tiers records whether the invite model distinguishes gold from silver signals for that specialty.

• ■They are not causal. Applicants who signal, rotate, or apply to a given program differ from those who do not in ways no dataset records: mentorship, geography, family, advice they were given. Every lift here is an association measured in the presence of those differences.
• ■The sample is self-selected. Applicant-reported data over-represents people who finished a cycle and chose to report it. Away-rotation data over-represents accepted rotations. Both skews push reported rates upward, and neither is correctable from within the data.
• ■Programs change faster than data is published. A new program director, a changed signal policy, a class-size shift: none of it appears in a dataset until at least a cycle later. The most recent cycle is always the least well measured.
• ■Nothing here models the parts that decide interviews. Letters, personal statements, interview performance, away-rotation impressions, and program-specific priorities are not in any of these datasets. They are also, by most program directors’ own accounts, where the decisions actually get made.
• ■A probability is not a prediction about you. A 40 percent estimate does not mean you will match 40 percent of the way. It means that among applicants who look like you on the measured axes, roughly four in ten matched. You are one draw, and the measured axes are not all of you.
• ■This is not medical or advisory guidance. Use it as one input alongside your program director, your specialty advisor, and mentors who know your file. Where this site and your advisor disagree, your advisor has information this site does not.

• Your own rows are yours. Every per-user table is row-level-secured on your account id. Nobody else can read your tracker, worksheet, profile, or match plan, and no aggregate query returns a user id.
• Community numbers are aggregates only. Away and signal community views are served by database functions that return counts and rates, never rows. Where an aggregate would be thin enough to identify someone, the interface flags it instead of rendering it.
• Anonymous submissions stay write-only. Rotator reports, post-cycle data submissions, and feedback can be written by anyone and read back by no one, including through the public interface.
• The most sensitive fields never leave your browser. Parts of the application worksheet are kept in local storage on your own device and are never written to the server.
• Nothing is sold, ever. There is no paywall, no premium tier, and no data broker on the other end of this.

Every model file carries a method version and the exact method string it was built with, and this page renders those strings directly rather than a summary of them. When a model changes, its version increments and this page changes with it. The signal-lift files on this build are method version 3; the credential-lift files are method version 2.

If a number here contradicts something you know to be true, that is worth an email. Program corrections, data errors, and methodology challenges all go to the same place and all get read.