What a Healthy Evidence Chain Looks Like: Vaccine Topology vs Antidepressant Topology

The Question

In M_MED1, we found that the antidepressant evidence chain has a circular structure: the monoamine hypothesis is the #1 hub in the justification layer (d1), the endpoint-selection layer (d2), AND the citation-as-support layer (d3). The founding assumption justified the research program, selected the endpoints, and was also the most-cited evidence for the research program. Same node. Every layer. That is the structural signature of a self-referential evidence loop.

But a finding like that only becomes meaningful when you can show what the alternative looks like. What does a healthy evidence chain look like in the same structural framework? Does the topology actually differ -- or does every evidence chain look circular by IRDME's definition?

M_MED2 answers that. We mapped the childhood vaccine evidence chain using the same three layers and ran four pre-registered hypotheses.

Pre-Registration

Pre-registered before any analysis was run.
• Hash: 615d18e3 (full: 615d18e38568367f0c0e01ef77b3925cbbf3c934da5091276540df30d7c78853)
• Timestamp: 2026-06-02T18:10:01 UTC
• Public record: https://github.com/vladi160/preregistrations/commit/caa26e2

The four hypotheses were committed to the public repo, hashed, and pushed before the engine saw the data.

The Dataset

Eight nodes representing the vaccine evidence chain:
• adaptive_immunity_mechanism -- the founding theoretical claim
• antigen_antibody_response -- the specific immune response induced
• immunological_endpoint -- seroconversion / antibody titer measurement
• phase3_efficacy_trial -- large-scale RCT measuring disease prevention
• safety_surveillance -- post-market adverse event monitoring (VAERS, Yellow Card)
• fda_regulatory_approval -- FDA/EMA regulatory decision
• clinical_vaccination_guidelines -- CDC/ACIP/WHO recommendations
• population_disease_reduction -- real-world surveillance showing disease incidence drop

Three relational layers:
• justifies (d1): which claims provide theoretical justification for others
• selects_endpoints (d2): which claims constrain which outcome measures are used
• cites_as_support (d3): which claims are cited as empirical evidence for others

What IRDME Found

All 4 hypotheses confirmed.

h1 -- FPL gradient: CONFIRMED r(justifies, selects_endpoints) = 0.7454, Spearman = 0.8053, p = 0.014. Statistically significant at n=8 -- a stronger signal than most n=8 graphs produce. The mechanistic justification and endpoint selection layers are tightly coupled, as predicted by the FPL.

h2 -- adaptive_immunity_mechanism is the justification hub: CONFIRMED Rank #1 in justifies. It drives the evidence design: it justifies the antibody response claim, the immunological endpoint, the trial design, and the safety monitoring design. High degree, high betweenness.

h3 -- clinical_vaccination_guidelines is the citation hub: CONFIRMED Rank #1 in cites_as_support. Clinical trials, safety surveillance, population data, and regulatory decisions all feed citations into the guidelines. The empirical institutional outcome -- not the founding theory -- is the top citation node.

h4 -- FDA approval rank <=3 in cites_as_support: CONFIRMED Rank #3. This confirmed the structural arrangement: citations flow toward empirical institutional outcomes (guidelines, regulatory approval, surveillance), not back to the theoretical mechanism.

Key Finding

The justification hub and the citation hub are different nodes.
• justifies hub: adaptive_immunity_mechanism (rank #1)
• cites_as_support hub: clinical_vaccination_guidelines (rank #1)

Adaptive_immunity_mechanism is completely absent from the cites_as_support top-5. The founding mechanistic theory disappears from the citation layer because citations go to empirical data -- the trials confirmed the vaccine works, so those trials and their institutional outputs are what gets cited. The theory is not cited in support of the theory.

This is the opposite of M_MED1:
• Antidepressants: monoamine_hypothesis rank #1 in ALL three layers.
• Vaccines: adaptive_immunity_mechanism rank #1 in justifies, absent from cites_as_support top-5.

Interpretation

The FPL provides a structural test that behaves differently across these two evidence chains:
• Healthy chain (vaccines): r = 0.75, p = 0.014. Strong gradient between mechanistic coupling (justification + endpoint selection) and citation coupling.
• Circular chain (antidepressants, M_MED1): r = 0.41, p = 0.44 (not significant). Weaker gradient.
• Circular chain (opioids, M_MED3, companion experiment): r = 0.14, p = 0.86 (not significant). Gradient collapse.

The structural interpretation: in a circular evidence chain, the founding assumption homogenizes all epistemic layers -- it justifies things, selects endpoints, and is also the top citation. This homogenization reduces the FPL gradient. In a healthy chain, the layers have distinct roles: theory for justification, empirical data for citation. This separation creates the strong gradient the FPL detects.

A note on what this does and does not claim. IRDME identifies structural patterns in how evidence claims relate to each other. It does not claim vaccines work because of their topology, or that antidepressants fail because of theirs. The causal claims belong to clinical trials and pharmacology. IRDME shows that the STRUCTURE of the evidence network differs between these two cases -- and that structural difference is detectable from graph topology alone, without reading any paper.

Limitations and Next Steps

n=8 is a representative model, not an exhaustive literature graph. The nodes were selected to capture the key structural claims; the topology would be similar but larger with n=20+.

The vaccine evidence chain used is idealized -- real vaccine safety debates (adjuvant concerns, schedule decisions) would add more complex structure.

The three-experiment series (M_MED1 antidepressants, M_MED2 vaccines, M_MED3 opioids) establishes the pattern as provisional. A formal benchmark against known replication failure rates across 10+ drug classes would be needed to move from provisional to confirmed.

Next question: does the FPL gradient score (r value in d1-d2 vs d1-d3 comparison) correlate with eventual replication failure rates across drug classes? If yes, this becomes a structural quality audit tool for pharmaceutical evidence chains -- detectable before the replication crisis happens.

Pre-Registration Record

Experiment: M_MED2_v1

Dataset: vaccine_evidence_chain.json, n=8, 3 layers, 21 relations

Hash: 615d18e38568367f0c0e01ef77b3925cbbf3c934da5091276540df30d7c78853

Timestamp: 2026-06-02T18:10:01.618670+00:00

Commit: https://github.com/vladi160/preregistrations/commit/caa26e2

Results: h1 CONFIRMED r=0.7454 p=0.014, h2 CONFIRMED (adaptive_immunity_mechanism rank #1 justifies), h3 CONFIRMED (clinical_vaccination_guidelines rank #1 cites_as_support), h4 CONFIRMED (fda_regulatory_approval rank #3 cites_as_support)

Verified: hypotheses unchanged, hash intact