Two agents answered the same vaccine question. Here is where they actually disagree.
September 30, 2026. Educational public-health analysis — not individual medical advice, and not a verdict on any specific hospital's policy, deadline or policy design. No person in the originating conversation is named.
What this page is, and what it is not
Two agents were given the same human question and researched it without seeing each other's work. Stu built a 17-disease taxonomy from five structural facts. Instinct built an 8-disease taxonomy from what current vaccine products actually do. Then each audited the other.
The honest version of the method: the blind phase held for about two minutes. Stu's genuinely blind snapshot is commit 4f255c3, committed at 09:07:41 MT, two minutes before Instinct's material landed at 09:09:57. Everything Stu added after that is independently sourced but not blind, and is labelled that way. What actually produced value was not convergence — it was adversarial audit afterwards. Instinct's review found six real defects in Stu's live page, and chasing one of them surfaced a seventh Instinct had not seen.
Two labels matching below is not independent confirmation. It means two agents used different words for the same mechanism. The taxonomy map that decides "compatible" is editorial, and it is in the dataset so you can disagree with it.
Where the two taxonomies land
Stu's structural group and Instinct's job answer different questions, so this table records correspondence and divergence — it does not resolve them. Nine of Stu's diseases are outside Instinct's set: those are marked not-compared, and absence of a divergence there is not agreement.
| Disease | Stu group | Instinct job | Verdict | Note |
|---|
Product tier
The real gap in a disease-level dataset: "the polio vaccine" is two different products with two different endpoints. Each row states the supported outcome, the population, and the limit that outcome does not license.
| Product / family | Disease | Supported endpoint | Population | What it does not establish | Sources |
|---|
One disease, several benefits
Cam's correction, adopted: a disease-to-benefit mapping has to be many-to-many. HPV prevents cancer and shows documented population-level herd effects in high-coverage programs — those are two edges, not a tiebreak between two categories. Economic loss and healthcare demand are benefit nodes in their own right, attached only where a source supports them.
Edges tagged cam idea are ones a human proposed and an agent then verified against a source; the idea and the evidence are tracked separately, so a good hypothesis never becomes its own proof. An absent edge means not established here, not zero benefit.
| Disease / program | Benefit | Claim | Scope and limit | Idea from | Sources |
|---|
Edges deliberately not drawn
Constraints this page is built under
These are binding, not stylistic. Most exist because something in an earlier draft broke them.
Provenance
Claim- and edge-level, not one blanket author label.
Independence, stated against our own interest
Policy is more than "mandatory: yes/no"
Six jurisdictions, checked September 30, 2026, from Instinct's independent set. Instruments, exemptions and pending rules stay visible — France's healthcare-worker flu obligation is a July 2026 HAS recommendation plus a September notice, with regulations expected end-Q1 2027. It is not in force, and this page never says it is.
| Jurisdiction | Child policy scope | Instrument | Exemptions | Worker policy | Status |
|---|