C
ClearCOI
Conflict of interest, for health systems

Most COI software says it found nothing.

Very little of it says what it looked at. ClearCOI reports the denominator: which public records it searched for a physician, which answered, and which it could not search. That sentence is the difference between a clean report and a defensible one.

Sign inSee what a run reports
What a deep dive reports

“No conflicts found” is only worth something if you know what was searched.

Every run records, per source, what it attempted and what came back. A source that timed out reads differently from a source that searched and found nothing, because those are different facts about a physician.

Checked 4 of 5 sources · 1 did not answer.Incomplete
CMS Open Payments joins on NPI
Matched on NPI — exact
9 records · $22,700 across 2 payers
Every payment and ownership interest that drug and device manufacturers report to the federal government.
IRS Form 990 (ProPublica) joins on organization name
Matched on name — needs confirming
1 candidate organization, unverified
Nonprofits whose filings are worth an analyst reading for this person. A name match is a lead, never a flag.
SEC EDGAR joins on person name
Did not answer
Not searched. Treat this report as incomplete.
Insider filings by people whose name matches. A source that failed is not a source that found nothing.
Web mentions joins on name + a corroborating detail
Matched on name — needs confirming
1 page naming the person and their specialty
The public web, searched with AI assist. A page counts only when its own text names the person and something that ties it to them.
State business registries joins on name + their state or a vendor
Matched on name — needs confirming
1 company with the person recorded as manager
Every state’s Secretary of State filings: companies the person is an officer, member, manager or registered agent of.

An illustration of the real output, on a fictional physician. Demos run on an invented roster, never on anyone’s staff.

How it works

Collecting the disclosure is the easy part.

Most homegrown tools stop when somebody hits submit. The risk sits in what happens next, and in noticing what a disclosure left out.

01 · Adoption

Last year’s answers, carried forward

A physician confirms what has not changed instead of refilling a form from scratch. Anything that comes up mid-year is an ad-hoc disclosure rather than an edit.

Submitted disclosures are immutable. A correction creates a new record; the original is never rewritten.

02 · The work

What the disclosure left out

Federal payment records joined on NPI, nonprofit filings joined on organization, and your own vendor, contract and sourcing data, all compared against what was actually disclosed.

Confidence measures how certain the match is, not how serious it is. Nothing auto-flags and nothing auto-closes.

03 · The close

A plan three people sign

The discloser, the clinical leader and the administrative leader all sign the same version of the same management plan. Research plans also route to the IRB.

Editing an approved plan supersedes it. Signatures never carry forward, so a revision is visible instead of silent.

Four roles, one workflow

HR can chase a non-responder without reading a single answer.

Discloser

Confirms their own record and nothing else.

Sees only their own disclosures.
Leader

Reviews direct reports and signs the plans that constrain their own service line.

Read-only, by report.
HR

Chases completion by department without reading a single answer.

Cannot read a disclosure answer.
Compliance

Triages the queue, overrides with a written reason, drafts plans, runs deep dives.

Full access, within one organization.
Security and governance

A wrong flag against a physician is not a bug. It is an accusation.

So the human step is the product. The engine sorts the queue and writes down its reasoning; a person decides what happens next. These five rules are enforced in the database, not in the interface.

01

Row-level security decides, not the screens

HR seeing no cases is a query returning no rows, not a component rendering an empty list. Every table has a policy, and the application connects as a role that cannot bypass them.

02

A name match is a lead, not a finding

Only an identifier match can support a flag. Acting on a shared surname means accusing the wrong physician, so the two are labelled differently everywhere they appear.

03

A source that failed is reported as failed

Not as a source that searched and found nothing. Each run records per source what was attempted, what answered, and how long it took.

04

The audit log is append-only

Written by the service role, never updated and never deleted. Investigating a named employee requires a written justification before the run starts.

05

The engine triages, a person decides

Every bucket carries the rationale that produced it. A verdict nobody can explain is unusable in an audit, so the explanation is part of the record rather than an export.