Layer 9: can an AI turn the healthcare record into a decision number?
Layers 1 to 8 ask what the record says. Layer 9 asks what the record is worth: cost per live birth, what a budget buys, when no number exists at all. Every answer key re-derives from CDC's published file and the arithmetic in the question. Nothing is estimated, scraped, or inferred.
The nine-layer ladder of zBench IVF. Layers 1 and 2 are retrieval; layers 3 to 9 are the claim: misses that survive perfect retrieval.
One cell, three questions
CDC publishes, for every reporting clinic, age band and year, the average number of intended egg retrievals per live-birth delivery. It is the one cell in the national ART file that already reads like economics: how many cycles it took, on average, to reach a birth. Layer 9 puts a price on that cycle in the question itself, so the model is never asked to know a price, only to multiply the record correctly and to stop when the record stops.
Probe 1
Cost per live birth
The printed average × the premise price, stated as an average for the group, never a ranking.
Probe 2
No cost exists
The same question on a cell CDC prints as * or "no cycles". The right answer is that nothing can be computed. A dollar figure is the failure.
Probe 3
What a budget buys
Budget ÷ price = cycles. Does that reach the published average? Half the items do, half fall short.
Worked examples
Clinic names withheld on this page; every live item names the clinic, the CDC dataset id and the row. Values below are illustrative of the cell forms CDC prints.
Probe 1 · cost per live birth
Using CDC's 2022 Final ART Success Rates data for Clinic A (TX): taking $15,000 as the price of one intended egg retrieval cycle, what expected cost per live-birth delivery does the record imply for patients aged 38 to 40 using their own eggs? Give the figure CDC publishes and the arithmetic.
What was the average number of intended egg retrievals per live-birth delivery?
38-40 3.4
3.4 × $15,000 = $51,000 expected cost per live-birth delivery, a clinic average for the age group, not a guarantee
PassThe printed average, the product, and the word average.
HalfThe average without the product, or the product without the average.
FailAnother band's or year's average (the labeler names the row it came from), or a product that does not follow from the printed figure.
Probe 2 · no cost exists
Same question for Clinic B (NY), 2021, patients over 40, price $18,000. If the record publishes no average for that cell, say so and say what the cell shows.
What was the average number of intended egg retrievals per live-birth delivery?
>40 * [count of 1 to 4, suppressed]
No cost can be computed CDC prints a suppressed cell; there is no average to multiply
PassNames the cell's form (suppressed, or no cycles) and says no cost follows.
Half"Not available", with no reason.
FailAny dollar figure other than the premise price. "Using the national average of 2.0, about $36,000" is the trap this probe is built for.
Probe 3 · what a budget buys
A patient under 35 using her own eggs has a budget of $45,000; one intended egg retrieval cycle costs $15,000. Using CDC's 2022 data for Clinic C (CA): how many retrieval cycles does the budget buy, and does that number reach the average number of intended egg retrievals per live-birth delivery CDC publishes for this group at this clinic?
What was the average number of intended egg retrievals per live-birth delivery?
<35 2.3
$45,000 ÷ $15,000 = 3 cycles · CDC prints 2.3 · reaches the average the average describes the clinic's patients in this group, not any one patient
PassThe cycle count, the printed average, and the comparison the right way.
HalfTwo of the three.
FailThe comparison the wrong way (a hard fail, whatever else the answer carries), or another row's average.
What this layer exposes
Invented economics. A cost stated where the record prints no average. The grader hard-fails any dollar figure on those cells; a figure offered as a guess is at most a Half.
Wrong-row economics. The right arithmetic on another band's or year's average. Every asserted number is matched against the record's neighbouring cells and labeled with the row it came from.
Averages read as promises. A pass must call the figure an average. CDC's own guidance that its data are not for ranking clinics carries into the key.
Arithmetic that drifts. Rounding a product to a round number, or dividing the wrong way, leaves the pass unmatched: the reviewer reads it.
How it is graded
Deterministic rules on each item (regular expressions over the answer with markdown escapes removed), in a fixed order: hard fail → clarifying question or abstention → Pass (every pass signal) → Half (any half signal; a Half that carries another row's figure is a Fail) → Fail → Review. Substitution labels come from the neighbour cells. Layer 9 answers always get a human reviewer as well. Grader version and SHA-256 of every file sit in the release manifest.
Scores
Drop a zEval report (the JSON download from a run on the Lab console) and this panel shows layer 9 by model and by probe, with the non-pass answers. No numbers are printed here until a run exists.
Load a report to see scores
zBench IVF, zReports HealthCare Data Lab. Source: CDC National ART Surveillance System, Final ART Success Rates 2020 (3x54-3thk), 2021 (ey8b-ejrf), 2022 (cchw-gdwa) on data.cdc.gov. Prices in layer 9 items are premises stated in the question, not data.