Transparency

How we decide what matters

A regulatory monitor is only as good as what it refuses to send you. This page documents the scoring and the gates in full — including the ones we added because our own engine got it wrong. The thresholds below are read directly from the running code, so this page cannot quietly fall out of date.

The path an item takes

1Poll. 10 government and legislative feeds, daily. Each is named publicly, with the endpoint we read.
2Diff. Changes are keyed by a content hash, so an amended bill reads as a change and a re-published notice does not read as news.
3Classify. Each item gets a materiality tier (A/B/C) and a signal score, plus a plain-language summary and a suggested action.
4Review. Tier A and Tier B are held for a person. Nothing in those tiers reaches the public feed without human approval.
5Publish. Approved items appear on the feed with a link to the primary source.
6Deliver. A separate decision from publishing. Most items are never emailed to anyone.

The three materiality tiers

Volume is not thoroughness. Most of what a regulatory sweep catches is real, sourced, and irrelevant to a practice — so we tier it and say which is which, rather than presenting a border-refusal record as though it were a rule change.

Tier A

Critical

default route: Today’s batch

A rule changed, and there is a concrete action a practice must take.

For example: A state enacts a supervision requirement with a January 1 effective date. A GLP-1 comes off the shortage list, ending the compounding basis for it.

Review: Human review, always.
Where it goes: Feed, the weekly Brief, and — only if it clears two more gates below — the daily email.

Tier B

Monitor

default route: Include in Brief

A real regulatory development that does not yet demand an action.

For example: A scope-of-practice bill is introduced and referred to committee. An FDA guidance document opens a comment period.

Review: Human review, always.
Where it goes: Feed and the weekly Brief. Never a standalone email.

Tier C

Background

default route: Feed only

Real, sourced, and on the record — but not something a practice acts on. Mostly single-shipment border actions and routine facility citations.

For example: One shipment of lidocaine refused entry at the border. Nine observations at a manufacturing facility.

Review: Two capped exhaust sources publish automatically (below). Everything else keeps its human.
Where it goes: Feed only. Structurally cannot email.

What has to be true before we email you

All three, always. Any one failing means the item still appears on the feed — it just does not interrupt anyone.

1. It is Tier A
A rule changed. Tier B and Tier C are structurally ineligible, whatever their score.
2. The recommended action is specific
The action has to point at something concrete — a lot number, a named firm or product, a bill number, a state and a deadline, a dated step, or a shortage-status change. Text that only says monitor for updates, stay informed, review your current practices, or verify your suppliers are licensedfails. That last one is a disclaimer wearing an action’s clothes, and it is the literal text on most of the capped border records in our corpus.
The rule of thumb the code encodes: if the item would still be true and unchanged tomorrow with nobody doing anything, it is not Tier A.
Why we don’t trust the classifier here. The engine tags its own actions “specific” or “generic”. Measured across the corpus, that tag disagreed with the definition above on 58% of tagged items — essentially all in the permissive direction. It rated “monitor future legislative sessions for renewed activity”as specific. So the gate is an independent check, not the classifier’s self-assessment.
3. The event is current
Published within the last 90 days — far tighter than the 18-month window the feed displays, because “worth putting on a page” and “worth interrupting someone today” are different bars. A future effective date overrides the age: a bill signed months ago whose duty starts next January is precisely the alert worth sending.
Why this gate exists.Two 2021 state bills were sitting in the emailable set, one of which said “comply before May 11, 2021”. The specificity gate could not catch it — a past date is still a concrete date.

Approving is not sending

These are two separate decisions, and fusing them is what turns a monitor into a nuisance. Approval says “this is true and public”. Delivery says “and here is how it reaches people”. Each approved item is routed explicitly:

Feed onlyAppears on the public feed. No email, ever.
Include in BriefAppears on the feed and in the weekly Brief.
Today’s batchRides the once-daily batch — one email covering the day, not one per item.
Send now (break-glass)A standalone send, reserved for break-glass. Most weeks: zero.

Tier A defaults to the daily batch rather than a standalone send, on purpose. Ten critical items on a busy day should be one email, not ten.

The one thing that publishes without a human

Tier-C background exhaust from exactly 2 capped sources — FDA Import Refusal and FDA Inspection Citation — publishes automatically. These are high-volume, low-consequence records that would otherwise bury a human queue. Every other item in every other tier and source waits for a person.

An auto-published item still has to clear every one of these, and anything that fails simply stays in the queue for a human:

Tier C, from one of the two capped sources named above
A signal score at or below 45 — a Tier-C item scoring higher means the tier and the score disagree, so we trust neither and escalate it
A primary source URL on an https government host — checked against an allowlist, not inferred
A content hash, so we know the deduplication is intact
A real title and body
A parseable date that is not in the future — a future-dated regulatory event is a parsing bug, not news

Tier C cannot email by construction — the automated publisher does not call the notifier at all. So the worst case for this path is an extra background record on the feed.

Severity labels

The badge on each item is a display of the signal score: CRITICAL at 90 or above, CAUTION at 70–89, MONITOR below that. These cuts were raised deliberately: at the original threshold, 63% of the corpus read CRITICAL, which makes the word mean nothing.

What we will not do

Fabricate a date. If a source gives no date, the item shows no date. We never infer one, and we never substitute the date we found it.
Show an unreviewed item as reviewed. The public feed can only read approved items — the database enforces it, not just the application code.
Count coverage as results. We monitor all 50 state legislatures. The number of states with a change actually on record is smaller, and it is the number we print.
Give legal advice. This is regulatory monitoring. Every item links its primary source, and the right next step for anything consequential is your own counsel.

See the pipeline’s current state on data status, or the feeds themselves on data sources.