How the brain was built
No single model wrote these rules and marked its own work. The brain is built by separate roles that check each other, with people as the final say — and everything each role did is on the record.
Decides which sources may be used at all. Licensed or confidential material cannot be added — the system has no way to express it.
Download each source from its publisher and convert it to text. Every file is fingerprinted, so a changed source is noticed.
Reads one source and writes one rule per requirement, each with a word-for-word quote. It cannot publish anything.
Confirms every quote really appears in the source. A rule with an invented citation cannot get past this.
Checks each rule against its source without seeing how it was drafted. Passes it, sends it back with required changes, or rejects it.
Decide anything the two AIs could not agree on, every blocker-level rule, and a random sample of the rest.
The only door into this site and into reviews. Refuses any rule whose quote, status or links do not check out.
Distinctions the sources are asking for
When a source draws a line our vocabulary can't express, the drafter says so instead of bending a rule to fit. These are waiting for a person to weigh.
Fair balance is scoped to print, so it never fires on a digital piece
Filed 2026-09-22 by the brain agent's first real question ("why would a fair balance rule apply to an HCP email?"). Observation, not a decision — for the 2026-10-04 axes checkpoint.
All four fair-balance rules under 21 CFR 202.1(e)(7) — US-FB-002, US-FB-004, US-FB-005, US-FB-007 —
carry medium: [print] (US-FB-005 print/broadcast). An HCP email is medium: digital, so none of
them are selected for one, and the same holds for an HCP web page or a banner.
The drafter was reading the source correctly: 202.1(l)(1) defines "advertisement" by naming published
media, and 202.1(l)(2) treats brochures and mailers as labeling. The question is whether our medium
facet should carry that distinction at all:
- For keeping it: the regulation's own words draw the line; a rule that claims to cover email where the citation does not is exactly the overreach the verifier exists to catch.
- Against: FDA has applied fair balance to digital promotion for years (untitled letters on
websites and banner ads are in
precedent/), so a brain that stays silent on a promotional email is wrong in the way that matters to a writer.
Probably the right shape: keep the citation honest and add the digital application as its own
rule resting on the OPDP internet/social-media guidance, rather than widening the CFR rules' scope.
That needs the guidance in sources first.
Check before deciding: how many precedent letters in the brain are about digital pieces and cite fair balance. That is evidence, not argument.
Observation: some scope is implicit in the source, so "absent = any" over-includes
- raised by: the Builder's framework counts, 2026-09-20 — for the 2026-10-04 axes checkpoint
The us-pre-approval framework reports ~83 rules in scope, almost all from 21 CFR 202.1. They
are in scope only because those rules carry no launch-stage at all — and they carry none
because the regulation never says "post-approval": advertising a prescription drug presupposes an
approved one. The Drafter was told to scope a rule only where the source text limits it, and it
obeyed.
So there are two kinds of scope: what a source states, and what a source presupposes by its subject matter. Today only the first is expressible. Options for the checkpoint:
- Let a source declare default scope in the registry (e.g. 202.1 →
launch-stage: post-approval,promotional: promotional), inherited by its rules and shown as inherited. - Make frameworks exclude as well as select.
- Treat
launch-stageas derived from the governing source rather than as a per-rule parameter — which would mean it is not a fundamental axis at all.
Option 3 is the interesting one for the question Surya asked: are Darren's axes the fundamental ones? This is the first piece of evidence that one of them may be a property of which law applies rather than of the piece.
Vocabulary gaps reported by drafters during revision, 2026-09-20
Raised while revising rules from the Product Name Placement guidance. Both were handled by stating the limit in the rule's prose instead — which keeps the rule honest but makes the limit invisible to scoping. For the 2026-10-04 axes checkpoint.
- No prescription-vs-OTC distinction. The guidance (footnote 3) excludes OTC products; four
rules needed the limit and had nowhere to put it. Candidate:
rx-status: [prescription, otc]. Nearly every source in this brain is prescription-only, so this may be better expressed as a source-level default (see_observation-implicit-scope.md) than per rule. mediumhas noaudiovisual. The guidance treats broadcast advertisements and non-broadcast audiovisual promotional labeling (in-office video, web video) alike. Two rules fell back tomedium: [broadcast]with a sentence saying broadcast means all audiovisual — that is bending a parameter, which the Drafter prompt forbids; the packet's fallback told it to. Candidate: addaudiovisual, or splitmediumfromregulatory-class(advertisement vs. promotional labeling) — the second is the distinction FDA's own rules turn on.product-typehas nodevice(raised earlier by the Presenting Risk Information drafter).
Proposed scoping parameter: ad-type
- values: reminder, bulk-sale, compounding, standard
- raised by: drafter, from ecfr-21-202.1, 2026-09-21
21 CFR 202.1(e)(2) creates distinct exemption categories (reminder, bulk-sale, prescription-compounding advertisements) with different content rules than standard ads; the existing scoping vocabulary has no way to flag a rule as applying only to, or excluding, these specific ad types.
Proposed scoping parameter: rx-status
- values: prescription, otc
- raised by: drafter, from ecfr-21-314.81, 2026-09-21
21 CFR 314.81(b)(3)(i) expressly limits the timing requirement for submitting specimens to 'the advertisement for a prescription drug product,' a scope distinction not covered by the existing approval-pathway/product-type vocabulary.
Proposed scoping parameter: product-type
- values: device
- raised by: drafter, from fda-guidance-presenting-risk-information-prescription-drug-medical, 2026-09-21
The guidance applies equally to medical device advertising and promotional labeling, but the provided product-type vocabulary (drug | biologic) has no value for devices, so device-specific scoping could not be recorded on any rule.
Proposed scoping parameter: rx-status
- values: prescription, otc
- raised by: drafter, from fda-guidance-presenting-risk-information-prescription-drug-medical, 2026-09-21
The guidance states it does not address OTC drug promotional labeling and instead covers prescription drug and restricted-device promotion; the provided vocabulary has no parameter capturing this Rx/OTC scope limit.
Proposed scoping parameter: request-channel
- values: public, non-public
- raised by: drafter, from fda-guidance-responding-unsolicited-requests-off-label-information, 2026-09-21
The guidance sets substantially different, section-specific recommendations for non-public one-on-one requests (Section V) versus public-forum requests, including those via emerging electronic media (Section VI). The provided applies_when vocabulary has no parameter capturing this public/non-public distinction, which the source itself uses to scope entire sections of rules.
Recent activity
| 2026-10-20 05:04 | builder | brain | ok | script |
| 2026-09-23 04:40 | builder | brain | ok | script |
| 2026-09-23 00:46 | builder | brain | ok | script |
| 2026-09-22 20:22 | builder | brain | ok | script |
| 2026-09-22 04:28 | builder | brain | ok | script |
| 2026-09-22 03:30 | builder | brain | ok | script |
| 2026-09-21 19:35 | builder | brain | ok | script |
| 2026-09-21 19:11 | builder | brain | ok | script |
| 2026-09-21 16:51 | builder | brain | ok | script |
| 2026-09-21 07:49 | builder | brain | ok | script |
| 2026-09-21 07:49 | verifier | US-SCI-003 | revise | claude-code:opus |
| 2026-09-21 07:49 | verifier | US-MIS-026 | pass | claude-code:opus |
| 2026-09-21 07:49 | verifier | US-DISC-015 | revise | claude-code:opus |
| 2026-09-21 07:49 | drafter | US-MIS-054 | revised | claude-code:sonnet |
| 2026-09-21 07:49 | drafter | US-MIS-049 | revised | claude-code:sonnet |
| 2026-09-21 07:49 | drafter | US-FB-012 | revised | claude-code:sonnet |
| 2026-09-21 07:47 | builder | brain | ok | script |
| 2026-09-21 07:47 | verifier | US-RISK-034 | revise | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-SUB-011 | revise | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-RISK-022 | revise | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-RISK-021 | revise | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-RISK-020 | revise | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-IND-013 | revise | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-RISK-033 | pass | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-RISK-030 | pass | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-RISK-029 | revise | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-MIS-054 | revise | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-SUB-016 | revise | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-SCI-030 | pass | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-SCI-020 | pass | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-IND-014 | revise | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-MIS-053 | pass | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-MIS-052 | pass | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-MIS-049 | revise | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-MIS-044 | revise | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-FB-014 | revise | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-FB-012 | revise | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-DISC-013 | revise | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-DISC-012 | revise | claude-code:opus |
| 2026-09-21 07:47 | verifier | US-DISC-009 | revise | claude-code:opus |