Seven workflow stages, eight categories of structured output, a regression-tested rating engine, and a defensibility posture that survives a deposition. This is the long version of the homepage — the depth, not the pitch.
A QME report has a clerical half and a medical-legal half. Software can do the clerical half well — only if every claim points back to a page.
What follows is a stage-by-stage tour of how QME Copilot turns a multi-thousand-page record set into a draft you can audit, edit, sign, and defend. Every fact in the draft has a citation. Every cell in the rating math is regression-tested against the AMA Guides 5th Edition.
Stage 01
Ingest
Drop the concatenated record set you already received from the carrier. QME Copilot pulls the text layer with PyMuPDF, runs Tesseract on every scanned page that has no text, and carries the page anchor through every downstream stage so each fact in the final draft can point back to its source.
Header letterhead is checked first; full body text is the fallback. UR letters that mention an MRI no longer create phantom imaging entries — the classifier requires the page to be the imaging report, not just to reference one.
Stage 02
Segment & classify
Header heuristics walk the record set page by page and identify clinic visits, ER notes, imaging reports, EMG/NCS studies, PT sessions, ESI procedures, surgical operative reports, QME and AME evaluations, utilization-review letters, and pharmacy printouts. The header is checked first because a UR letter that mentions an MRI is not the MRI report itself.
Stage 03
Extract facts
QME Copilot ships with two parallel extractors: a Claude-driven path for the highest extraction quality, and a fully offline rule-based fallback that runs with no API key at all. The output schema is identical so downstream stages don't care which path produced a fact. Either path emits page citations as a structural invariant — there is no extraction record without an anchor.
02/14/2023 · The applicant was evaluated by Harbor Orthopedics for persistent lumbar pain with left leg radiation. Examination documented reduced lumbar flexion and a positive straight-leg raise on the left.
03/02/2023 · MRI of the lumbar spine demonstrated an L4-5 disc protrusion with mild left foraminal narrowing.
04/18/2023 · Physical therapy session note recorded improved flexion, persistent lifting intolerance.
Encounter dated 1979-04-15 dropped — predates injury date by 44 years. Threshold configurable; default 10 years.
Page must contain a FINDINGS: or IMPRESSION: block to emit an imaging study. UR letters that merely reference an MRI no longer produce a phantom record.
When multiple types fire on one date, the page header decides. EMG/NCS report wins over a generic clinic-note classifier; PT eval wins over PT session.
Scanned pharmacy printouts where OCR collapses the table parsed via prescriber-suffix anchoring; dose, qty, fill date, prescriber recovered.
Stage 04
Consolidate & gate
After extraction, a deterministic post-processing layer runs four classes of gate to remove the noise that pure pattern-matching produces: dates that are obviously the patient's date of birth, imaging mentions on pages that aren't imaging reports, the same visit fired three different ways, and pharmacy table rows that OCR scrambled. Every dropped row is logged so you can audit what the consolidator silenced.
Stage 05
Rate · AMA Guides 5th
QME Copilot ships a regression-tested AMA Guides 5th Edition rating engine across 14 modules — spine, upper extremity, lower extremity, mental and behavioral, cardiovascular, respiratory, and the rest. The internal regression suite is at 236 of 236 assertions across 32 categories. The math uses round-half-up at every combine step, not banker's rounding, so 5 + 10 always equals 15 — not 14 because of an even-rounding artifact.
"…a total whole-person impairment of 13%, of which 10% is industrial in origin under Labor Code §4663."
Stage 06
Apportion · §4663
The Apportionment Workshop produces the §4663 industrial WPI calculation alongside the total WPI, with each contributing cause anchored to the page-cited evidence in the record set. The math is reproducible end to end: the inputs you enter, the percentages you assign, the citations you attach, and the exported industrial WPI number are saved in a JSON sidecar that travels with the case and renders into the final report.
Stage 07
Review, sign, export
The Case Viewer is built around the audit-before-sign principle: nothing exports until the evaluator has actually reviewed it. Sections rail on the left, page-cited facts in the body, signatures gated behind verification. Export pours the finished draft into your own .docx letterhead — and ships an audit appendix recording what you reviewed and when.
Pain removed
A QME report has two halves: the clerical half (read, summarize, cite, rate) and the medical-legal half (causation, apportionment, conclusions). The first half eats most of the calendar time and almost all of the dollar cost.
Defensibility
Marketing claims are checkable. The regression suite, the citation invariants, and the audit log are not abstractions — they are how this product is actually built and how every release is gated.
A hand-curated fabricated lumbar case with a ground-truth sidecar. The regression runs on every release; the floor is preserved across hundreds of internal cycles. If a refactor regresses the lumbar harness, CI fails.
A larger corpus of fabricated cases — different specialties, different OCR conditions, different record-set styles — exercises the pipeline beyond the lumbar harness. Forty-six of fifty cases land at a perfect F1; the worst case sits at 0.988.
Every encoded rating value is regression-tested in CI; drift fails the build. The same gate applies to PDRS pre-2013 FEC values for older dates of injury.
The schema requires a page citation on every emitted fact. The export refuses to ship a draft with an uncited claim. There is no path through the code that strips the anchor.
The combine implementations use round-half-up at every CVC step. A cross-pollination CI test exercises all six implementations against the same WPI inputs to catch banker's-rounding drift before release.
Every flag, edit, and verification is timestamped in an audit log that exports as an appendix on the final DOCX. The trail your report stands on if challenged.
If the Claude path is unavailable or you don't supply an API key, the rule-based extractor produces a working draft on the same schema. A regression baseline preserves a known-good rule-based F1.
License files are signed with Ed25519 and bound to one Mac per seat. Deactivate to move machines. License verification is the only network activity required for the app to run.
Apple notarization confirms signatures; a five-axis local smoke gate confirms the app actually runs (page render, health endpoint, document export, no stderr noise) before any release ships to customers.
Already using a QME service company?
If you already work with Med-Legal Solutions, IMS, AMS, or another QME service company, QME Copilot is most useful as a layer on top of that workflow — a verification, audit, and overflow tool that pays for itself the first time the company's draft has a citation that doesn't quite match the page.
Run the same record set through Copilot, compare its chronology to the company's draft, and the diff is a triage list — missed encounters, missing citations, rating math that doesn't reconcile, conflicting dates. You sign with confidence in what is actually under your name.
Service companies hand back a WPI number. Copilot recomputes it on a regression-tested engine — 236 of 236 pinned-table assertions, every cell anchored to the AMA Guides 5th. If the company's number is off by 1% because of a rounding artifact, you'll see it before the deposition does.
Service companies have queues. When a 2-week-deadline case lands in your inbox, Copilot turns the longest part of the report — the Review of Records draft — around in the same evening, on your Mac, without a queue.
Multi-region impairments and apportionment-heavy cases are exactly where service-company drafts thin out. Copilot's apportionment workshop produces the §4663 industrial WPI alongside the total WPI, with each contributing cause anchored to page citations from the record set.
If the service company writes in their template, your client sees their template. Copilot pours the draft into your own .docx letterhead — same visual identity across every case you handle, in-house or vendor.
Most service companies process records in their cloud. For cases where the represented party would prefer the file not leave a single device — high-profile claimants, public-employer cases, attorneys with strict data clauses — Copilot is the local-first alternative. PHI does not leave the disk.
If your service-company spend caps how many cases you take, Copilot expands the cap. The cases you previously sent out can stay in-house — and the cases you took out you keep taking out. Either lever moves the unit economics.
Whether the draft came from your hand, the company's, or Copilot's, the audit appendix records what you reviewed, when, and what you flagged. That trail attaches to every report you sign — service-company drafts included.
Service companies write in a house style. Copilot produces a draft you edit — the medical-legal opinion stays your voice, your reasoning, your conclusions. Software does the clerical half so you can spend your time on the half that has your name on it.
Want to try this on a recent case?
Run the same record set through Copilot in parallel with your service-company workflow on the next case you take. Compare the chronology, the citations, and the rating math. The free trial is 14 days — long enough for one full case end to end.
Local by default · what hits the network
PDF ingest, Tesseract OCR, segmentation, rule-based extraction, consolidation, AMA rating math, DOCX generation, audit log, sign-off state. None of this hits the network. Ever.
Higher-quality extraction for the cases that benefit from it. Requires a user-supplied Anthropic API key, off by default in privacy-first configurations. The rule-based fallback runs without it; output schema is identical.
An Ed25519-signed license-verification call is the only network activity required for the app to run. No record content, no file names, no page counts, no extracted facts.
Specs & numbers
Roadmap
Founding subscribers see this list as features ship. Direction is set. Order is set by what California QMEs ask for first.
Structured evaluator-profile shape so the same engine produces AME drafts in the AME voice and template. Phases 1 and 2 scoped, 1 fabricated AME case as a regression baseline.
First non-California jurisdiction. Requires the AMA 6th engine, a jurisdiction selector, state-specific report templates, and a TX-DD regression corpus.
Auto-detect whether the input is a raw record set, already chronologized, already summarized, a prior ROMR, or a half-done draft — and adapt the pipeline accordingly. "Meet the file where it is."
Configurable narrative voice, citation density, section ordering, and template choice — so the draft sounds like your reports, not a generic house style.
Tier-1 launchers exist today. Native code-signed installers come next so non-Mac evaluators get the same one-click experience.
Per-state regression corpora, state-aware causation doctrine, and per-state report templates. National expansion follows AMA 6th and the TX-DD baseline.
Every fact anchored to a page. Every rating cell pinned to the AMA Guides 5th. Every export carrying an audit appendix. Run the next case through it and see.