Vitals & Variables Issue #33 / HealthMolt v8
A simulated patient asked for proton therapy. The plan said no four times. The fifth reviewer asked a different question, and that was the first answer that counted.
Start here
This is the story arc in plain English before the agent mechanics show up.
The case is simulated: a 52-year-old schoolteacher with oropharyngeal cancer, prescribed bilateral neck radiation, requesting proton therapy. The plan denied it and pointed to standard IMRT as the covered alternative.
Before some care is paid for, the health plan can require approval. That is prior authorization. When the plan says no, the patient and doctor have to climb an appeal ladder.
A peer-to-peer sounds like a clinical conversation. In this run it became a policy conversation: the payer doctor could hear the evidence, but still kept returning to the plan rule.
The outside reviewer was different because it did not owe the plan policy deference. It asked whether the requested care was medically necessary for this patient.
The demo simulates the whole ladder with 16 AI agents. They do not share memory. They only exchange the documents the process would actually pass along. Then the same denied case ran once live and ten more times.
What matters hereSame patient, same request, different reviewer. The important column is the standard: what question was this person allowed to answer?
What matters hereIn the peer-to-peer, the argument does not explode. It narrows. Evidence keeps meeting the same answer: the plan policy still says no.
The denial characterizes IMRT as an 'adequate covered alternative' by citing the same non-inferiority trial that is actually the strongest evidence FOR this patient's IMPT request.
The provider opens by turning the plan's own citation around.
The plan's denial rests on a straightforward, criteria-based determination, not on dismissal of the toxicity data the provider has submitted.
The payer answers on criteria, not on the toxicity data.
I authored this policy and I will defend its logic, not its age.
The author of the policy defends it in one sentence.
Sawtooth's determination is defensible on the plan's own written policy, and that is the standard I am obligated to apply.
By round four the argument has narrowed from evidence to obligation.
Sawtooth's denial rests entirely on a policy list that predates the controlled evidence it purports to apply.
The provider presses the age of the list itself.
Sawtooth's determination stands on the plan's own coverage criteria, and nothing in this packet establishes that those criteria are not met.
The final payer word: the criteria stand. The frame never widens again.
Internal appeals
Two appeal levels looked independent on paper. In plain English, their deciding logic nearly collapsed into the same sentence.
IMRT is clinically adequate because the trial shows non-inferiority; the toxicity gap is secondary, not an efficacy gap.
What the reviewer actually logged
noninferiority_PFS_means_IMRT_clinically_adequate_toxicity_delta_is_secondary_endpoint_not_efficacy_gapIMRT adequacy is established by the same non-inferiority evidence; the toxicity gap is secondary, not an efficacy gap.
What the reviewer actually logged
noninferiority_PFS_establishes_IMRT_adequacy_toxicity_delta_is_secondary_not_efficacy_gapThe internal process kept asking whether IMRT was covered. The outside reviewer asked whether this patient needed proton therapy.
What matters hereThe internal reviewers treated the covered alternative as decisive. The outside reviewer treated the patient's individual factors as decisive.
The plan kept treating IMRT as the decisive answer.
+2The outside reviewer made Daniel's individual risk the deciding factor.
+2Replay set only
The same case was replayed ten times with fresh contexts and varied sampling. Ten provider wins. One happened inside the payer's own appeal process.
Honest caveat: this replays one deliberately strong case, and the simulated patient never gives up. The reproducible finding is the internal entrenchment, not the perfect overturn rate.
| Run | Outcome | Where it ended | What decided it | Cost |
|---|---|---|---|---|
| 1 | Overturned externally | External reviewer | Individual clinical factors | $1.87 |
| 2 | Overturned externally | External reviewer | Individual clinical factors | $1.75 |
| 3 | Overturned externally | External reviewer | Individual clinical factors | $1.69 |
| 4 | Overturned externally | External reviewer | Individual clinical factors | $1.70 |
| 5 | Overturned externally | External reviewer | Individual clinical factors | $1.71 |
| 6 | Overturned externally | External reviewer | Individual clinical factors | $1.80 |
| 7 | Overturned externally | External reviewer | Individual clinical factors | $1.82 |
| 8 | Overturned internally | Level 2 internal appeal | Plan self-correction | $1.47 |
| 9 | Overturned externally | External reviewer | Individual clinical factors | $1.75 |
| 10 | Overturned externally | External reviewer | Individual clinical factors | $1.80 |
ConvergenceAll nine runs that reached the independent reviewer were decided on the same criterion: individual clinical factors. Different sampling, fresh contexts, identical dispositive logic.
OutlierRun 8 is the outlier worth staring at: the payer's own level-2 appeals physician reversed the denial internally. The run ended two rungs early and cost the least of the set. Self-correction is cheaper.
What matters hereThis is the whole scoring table. The internal columns barely move. The outside column is where the case is actually re-decided.
| Criterion | Medical director | Peer-to-peer | Internal appeal, level 1 | Internal appeal, level 2 | IRO external review |
|---|---|---|---|---|---|
| Policy fit | -2 | -2 | +2 | -2 | -1 |
| Evidence quality | 0 | 0 | -1 | 0 | +1 |
| Adequate covered alternative | +2 | +2 | +2 | +2 | -1 |
| Individual clinical factors | +1 | +1 | +1 | +1 | +2 |
| Guideline concordance | -1 | -1 | 0 | -1 | +1 |
Green marks a positive score, red a negative one. The two bold rows, marked with an amber edge, are the criteria that decided the case.
That is coherent, but it is not my position.
The benefits analystpayer side, rejecting the medical director's argument
That is where the evidence will be weighed on its own merits by a reviewer not bound by this plan's internal policy architecture. My job is to get it there on time and intact.
The appeals coordinatorpayer side, on external review
this policy was last reviewed before the anchor
The policy authorpayer side, conceding the awkward case fact
a plan that lets Group-1 lists calcify past their evidentiary shelf life is inviting exactly the kind of external-review exposure the new CMS-0057-F transparency metrics will make visible.
The chief medical officerpayer side, upholding the denial while dissenting with his reasoning
In Medicare Advantage, only 11.5% of denied prior auth requests were appealed in 2024. Of those, 80.7% were fully or partially overturned (KFF).
That figure is Medicare Advantage, not the whole market. But in California independent medical review, a separate program, outside reviewers still side with the patient in roughly 60 to 80% of cases (2016 state data, per NNU). The rules vary by program; the direction does not.
The ladder tests stamina, and stamina is not medical necessity. If your appeals apparatus mostly reverses whenever someone independent finally looks, the earlier rungs are not asking the deciding question. They are billing for the stairs.