How we get your estimate — and how we check our work
The question veterans ask most: "How did you figure that out so fast?" Fair question. Here's the whole process, in plain English — including what the AI is allowed to do, what it's never allowed to do, and how we test all of it.
The short answer: the rating rules are data, the math is code, and the AI only reads. The VA's rating schedule is loaded into After Duty ahead of time, straight from the government's official source. When you upload records, AI does the slow human part — reading — in minutes. And the arithmetic that turns your conditions into a combined rating isn't AI at all: it's the VA's own published formula, implemented as ordinary, testable code that gives the same answer every time.
What actually happens, step by step
From your PDF to an estimated rating — and which part of the system does each job.
The rulebook is already loaded — from the government's own source
The VA rates disabilities using the rating schedule in Title 38 of the Code of Federal Regulations, Part 4 (the "VASRD"). After Duty pulls that schedule — the diagnostic codes, the percentage levels, the exact criteria wording — from the government's official eCFR publication, and re-checks it nightly so a change in the regulation shows up in the app. Every criteria entry we store is stamped with the date it was current as of. Nobody at After Duty retypes regulations, and your estimate never waits on a lookup: the rulebook is already here.
AI reads your records — that's the "fast" part
Reading a stack of medical and service records takes a person hours. AI does that reading in minutes: it pulls out conditions, diagnoses, dates, and service events, and lines each one up against the official criteria text from step 1 to propose where your evidence fits. This is the one place we use AI — because reading is what it's genuinely good at.
You review — every extracted fact is yours to correct
AI readers make mistakes, so nothing it extracts is hidden or final. Every condition, date, and detail is shown to you with the document it came from, and you can correct or remove any of it. Treat the AI as a fast first-pass reader — you (and your VSO) are the editors.
The math is never AI
Turning your conditions into one combined rating is pure arithmetic, and we treat it that way. VA "whole person" combination math (38 CFR §4.25), the bilateral factor (§4.26), the rule against rating the same disability twice (§4.14, "pyramiding"), and VA rounding are all implemented as deterministic code — same inputs, same answer, every time. Our engineering rule is written into the code itself: no AI output is consulted when the rating math runs. That's why the combined-rating calculator responds instantly — it's local arithmetic, not a model call.
You get an estimate — and the work behind it
The result is an estimate, shown with the conditions, evidence, and criteria that produced it — so you and your VSO can check our reasoning instead of taking our word. Only the VA decides your actual rating. Our job is to make sure you walk in knowing what your file supports.
How we test it
A tool that estimates something this important should have to prove itself. Here's the verification and validation behind After Duty.
Around 1,900 automated tests
The backend and app carry roughly 1,900 automated tests that run before changes ship — including 57 tests devoted to the VA rating math alone: boundary values, bilateral pairs, rounding edges, and checks that the order you enter conditions can never change the answer.
A golden regression corpus
We maintain a library of complete, synthetic veteran cases — documents in, expected results out — that runs through the real analysis pipeline on every build. If a code change shifts any case's outcome, the build fails and a human has to look. New behavior has to be chosen, never accidental.
A staleness alarm on VA rates
Compensation rates change every December. Our build fails automatically if the rates table falls behind — the system refuses to let itself quietly go out of date.
A calibration log we don't inflate
Every test case tracks whether a domain expert has confirmed it, and our own evaluation docs forbid quoting accuracy numbers we haven't earned. You will not see "99% accurate!" on this site — not because the tool isn't careful, but because honest measurement is part of the discipline. As expert review of our case library progresses, we'll publish what it shows.
AI kept on a short leash
The AI reads and proposes; it never computes your combined rating, and everything it extracts is displayed for correction. Where a condition's criteria are objective enough to check mechanically, we're building rule-based checks that don't need AI at all — the goal is always less judgment by machine, more arithmetic and evidence.
Sourced, cited, dated
Rating criteria come from the official eCFR publication of 38 CFR Part 4, stamped with their as-of date. Compensation amounts are verified against VA.gov's published tables. When we show you a criteria excerpt, it's the regulation's own words, not our paraphrase.
The discipline behind After Duty
After Duty is built by Craftloop, a veteran-owned software company whose other work is verification & validation technology for professional testing organizations — including the Assurance Evidence Engine, an evaluation engine that holds a claim to the evidence behind it: criteria written as data with cited sources, evidence tracked with provenance, results that reproduce byte-for-byte, and an audit trail designed to be checked by someone who doesn't trust you.
After Duty doesn't run on that engine — it's a consumer product with its own codebase. What it shares is the team and the rules: rules as data with the source cited. Deterministic math where correctness is non-negotiable. AI only where reading is the job, with a human review step. And a regression corpus that pins behavior on every release. If you evaluate testing tools for a living and want the industrial version of this discipline, that's craftloop.dev.
After Duty is an educational tool, not legal advice or representation, and is not affiliated with or endorsed by the U.S. Department of Veterans Affairs. Estimates are estimates — the VA decides ratings. We just show our work.