Georgia Health Cost Baseline — specification v1.0.1

A public, reproducible per-county metric: what one reference buyer — buying on Georgia Access without a premium tax credit — pays for a year of coverage in Georgia in a healthy year (the Floor) and in a bad year (the Ceiling, the lowest in-network worst case), computed only from the public plan-year files.

License. Specification (this document): CC0 1.0 (public domain dedication). Dataset (the committed artifacts data/evidence/ga-health-cost-baseline-<planYear>.json and the /dataset routes that serve them): CC BY 4.0. Build script (scripts/build-ga-health-cost-baseline.mjs): license designated when the build repository is published (Apache-2.0 proposed). Anyone may reproduce, republish or extend the specification; the reproduction recipe is in §9.

1. Reference buyer

Age 40, non-tobacco, self-only, buying on Georgia Access (Georgia's state-based Marketplace) without a premium tax credit — unsubsidized, gross premium: the gross filed premium; no advance premium tax credit, no cost-sharing reduction, no ICHRA or employer contribution — an on-Marketplace individual medical plan effective January 1 of the plan year, in the plan's standard "-01" variant.

The age-40 premium is the IndividualRate on Rate PUF rows whose tobacco marker is Tobacco User/Non-Tobacco User or No Preference (both carry the non-tobacco rate) and whose rate window covers January 1 of the plan year.

2. Universe per county

A plan is in a county's universe when all of the following hold:

1. Listed. It appears in the same-year Georgia Access Plan Attributes PUF; or, where Georgia Access lists nothing for a plan that is priced, in the CMS State-Based-Exchange QHP PUF bundle for Georgia (GAPlans…csv). Georgia Access wins on overlap — a CMS row is used only for a standard component Georgia Access does not list, and a CMS service-area identity is used only when Georgia Access does not file that (issuer, ServiceAreaId). This mirrors the engine's GeorgiaCmsFallbackSelector (lib/surely/market-data/reconciliation.ts).
2. Priced. The same-year Georgia Access Rate PUF carries exactly one positive age-40 non-tobacco rate for it in the county's rating area. Conflicting age-40 rates (two different positive rates, or a zero and a positive rate) for a plan in a rating area exclude the plan in that rating area.
3. Not excluded. Excluded: dental-only plans, Catastrophic plans, off-exchange-only plans (QHPNonQHPTypeId = Off the Exchange), child-only offerings (ChildOnlyOffering = Allows Child-Only), plans that lack an integrated in-network individual maximum out-of-pocket (MedicalDrugMaximumOutofPocketIntegrated ≠ Yes, or TEHBInnTier1IndividualMOOP empty / not applicable), plans not effective on January 1, and plans with no "-01" variant row.
4. Offered throughout the entire county (a full-county service area row or CoverEntireState; §3).

Plans priced but listed nowhere (in neither Georgia Access nor the CMS bundle) are excluded and reported per county as pricedUnlisted (a count) and pricedUnlistedIssuerIds (issuer ids only — never plan or carrier names).

3. Placement — full-county only

Using the Service Area PUF, a plan is offered throughout the entire county only via a CoverEntireState row or a full-county service area row (PartialCounty ≠ Yes). Partial-county rows do not count for the headline figures. They feed one flagged companion column, floorAnywhereMonthly: the lowest premium among plans offered in any part of the county (full-county and partial-county rows together), or null when the county has no partial-county-only plan. There is no representative-ZIP rule — the Baseline never picks one ZIP to stand for a county.

4. The two figures

• FLOOR (healthy-year cost) = 12 × the lowest filed monthly age-40 premium among the county's universe plans, for the reference buyer buying on Georgia Access without a premium tax credit (unsubsidized, gross premium). Published as floorMonthly (the filed monthly premium, to the cent) and floorAnnual (12 × that, to the cent), with floorPlanId.
• CEILING (bad-year cost) = the lowest in-network worst case: the minimum, over the county's universe plans, of (12 × premium + the plan's in-network individual out-of-pocket maximum); out-of-network care, non-covered services and balance billing are outside it. Published as ceilingAnnual (to the cent; MOOPs are integers as filed), with ceilingPlanId. The Ceiling plan is often not the Floor plan.

Both are per plan — the buyer picks one plan and keeps it all year. Neither figure nets out subsidies, and neither models any particular pattern of care: the Floor is "you never use it", the Ceiling is "you hit the plan's in-network cap".

5. Grain and statewide summary

All 159 Georgia counties, every plan year; a run that cannot price 159/159 counties aborts. County rows are ordered by name (case-insensitive), FIPS tie-break — the order of the county index. Slugs follow countySlug in app/(marketing)/georgia-renewal-index/counties.ts.

The statewide summary is the nearest-rank median across counties (unweighted by population — every county counts once; for 159 counties the 80th smallest value) plus the min and max with county names, for floorMonthly, floorAnnual and ceilingAnnual. No care profiles, no other ages, no SBC scenarios.

6. Renewal Shock

With --prior <previous plan-year artifact>, each county row carries renewalShock: priorFloorAnnual, deltaDollars, deltaPercent (one decimal), and — when both years have a Ceiling — priorCeilingAnnual, ceilingDeltaDollars, ceilingDeltaPercent. A county absent from the prior artifact gets null. A statewide average is never applied to a county. The top-level renewalShock block summarizes the county deltas (nearest-rank median, min/max with county names) and pins the prior artifact's filename, sha256, plan year and completeness.

7. Labels on every figure

Completeness (completeness, validated against the inputs actually supplied — the flag must agree with the files or the run aborts):

The January 16 freeze is the OEP-close provisional snapshot: the artifact built from the Georgia Access files as posted when open enrollment closes, before the CMS bundle posts (typically May–June).

Rate status (rateStatus, rateStatusSource): proposed vs approved is read only from the county index's proposed2027 block (data/evidence/ga-county-index-2026.json#proposed2027) through the node twin of lib/evidence/rate-status-core.ts (scripts/lib/rate-2027-block.mjs) — validated, never typed. The block describes Georgia's 2027 rate filings: for a plan-year 2027 artifact it labels the figures themselves (proposed until Georgia's Insurance Department approves); for a plan-year 2026 artifact it is context and the artifact's notes say so.

8. Inputs, failure modes, determinism

• Inputs are pinned in the artifact as { label, filename, url, sha256, bytes } — the Rate, Plan Attributes and Service Area PUFs, the CMS bundle and its two members, the geography snapshot and the county index. The PUF CSVs and the bundle are not committed; download them into gitignored .context/ from the URLs in the artifact (lib/surely/market-data/manifest.ts carries the same pins).
• Literal header names are pinned per file dialect and drift aborts: Georgia Access (BusinessYear, StateCode, IssuerId, PlanId, RatingAreaId, Tobaco — the Rate PUF's spelling — Age, IndividualRate, RateEffectiveDate, RateExpirationDate; StandardComponentId, MarketCoverage, DentalOnlyPlan, ServiceAreaId, MetalLevel, QHPNonQHPTypeId, ChildOnlyOffering, PlanEffectiveDate, PlanExpirationDate, CSRVariationType, MedicalDrugMaximumOutofPocketIntegrated, TEHBInnTier1IndividualMOOP; CoverEntireState, County, PartialCounty) and CMS SBE (BUSINESS YEAR, STATE CODE, ISSUER ID, STANDARD COMPONENT ID, SERVICE AREA ID, METAL LEVEL, QHP NONQHP TYPE ID, CHILD ONLY OFFERING, MEDICAL DRUG MAXIMUM OUT OF POCKET INTEGRATED, TEHB INN TIER 1 INDIVIDUAL MOOP, COVER ENTIRE STATE, COUNTY, PARTIAL COUNTY, DENTAL PLAN ONLY, …). Extra columns are tolerated; a missing pinned column is not.
• Fail closed: any Georgia row with the wrong BusinessYear; any unusable age-40 rate row (bad date, plan id, rate or tobacco marker); a county with no universe plan (named); a rating-area label the geography does not know, or a county whose exactly-placed plans do not file a rate in its rating area; fewer or more than 159 counties (--expect-counties exists for the 3-county test fixture only and is written into notes[]).
• Deterministic: no timestamps beyond --as-of, fixed key order, numbers as numbers, two runs on the same inputs are byte-identical (the test asserts it; the audit CSV too). Zero network calls. Issuer ids and HIOS plan ids only — no carrier or plan names anywhere in the artifact or the audit.
• The audit CSV (--audit-out) lists every plan considered in every county — countyFips, countyName, ratingArea, planId, issuerId, classifier, placement, premium40, moopIndividual, outcome, reason — so any county figure can be traced to the row that produced it. Its row count and sha256 are in the artifact (audit).

9. Reproduce (PY2026, RECORD)

mkdir -p .context/ga-puf-2026
cd .context/ga-puf-2026
curl -sL -A "Mozilla/5.0" -O https://georgiaaccess.gov/wp-content/uploads/sites/3/sites/4/2025/09/RATE-PUF_20250926212630_2026.csv
curl -sL -A "Mozilla/5.0" -O https://georgiaaccess.gov/wp-content/uploads/sites/3/sites/4/2025/12/PLAN-PUF_20251216063410_2026.csv
curl -sL -A "Mozilla/5.0" -O https://georgiaaccess.gov/wp-content/uploads/sites/3/sites/4/2025/12/SA-PUF_20251205153240_2026.csv
curl -sL -A "Mozilla/5.0" -O https://www.cms.gov/files/zip/georgiasbepuf2026.zip
shasum -a 256 *
#   b40eff4c2871f0e97ae95ed019425e128f0474967a38af4b1d5031eafc2a6b20  RATE-PUF_20250926212630_2026.csv   (17,665,307 bytes)
#   a3c540b9e4440055fb68a9922018965ee3c78586bdb5709753f604c5e9ecb3d6  PLAN-PUF_20251216063410_2026.csv   (958,712 bytes — the Dec 16 export, not the Aug 28 re-export)
#   ab268965108a70027d244f3f4ccd1ab16f2cd95ca3956268eae0997d3d836b3a  SA-PUF_20251205153240_2026.csv     (144,978 bytes)
#   4e33017f3c9a05d38a6f765df26982f258eaa999bd39dc6b8221e191f9011933  georgiasbepuf2026.zip              (1,352,514 bytes)
cd ../..

pnpm evidence:ga-health-cost-baseline -- --plan-year 2026 --as-of 2026-09-02 --completeness record \
  --rate .context/ga-puf-2026/RATE-PUF_20250926212630_2026.csv \
  --plan .context/ga-puf-2026/PLAN-PUF_20251216063410_2026.csv \
  --service-area .context/ga-puf-2026/SA-PUF_20251205153240_2026.csv \
  --sbe-zip .context/ga-puf-2026/georgiasbepuf2026.zip \
  --replay data/evidence/ga-county-index-2026.json \
  --out data/evidence/ga-health-cost-baseline-2026.json \
  --audit-out output/ga-health-cost-baseline/2026/audit.csv

Expected (about 0.3 s on a 2024 laptop; Node 22 or 26):

sha256 2c38167878ae90b81b9b040ab9b8c89b39846194ff3d0c3d6a3c2e7181158ee0   data/evidence/ga-health-cost-baseline-2026.json
audit  22079 rows; sha256 bab30245fd4aeae5098e2cb849e9f13d1b511fbbc90398b1e403277e654f0698
statewide: median Floor $450.58/mo (min Gordon $370.65, max Carroll $573.63); median Ceiling $13,029.44/yr; counties 159; plan universe 203
plans: rate PUF 240; by classifier {"cms-sbe":16,"georgia-access":189}; excluded {"dental-only":29,"catastrophic":6}; rate conflicts 0
replay vs data/evidence/ga-county-index-2026.json: 157/159 match (Cobb, Fulton — partial-county-only 45334GA0020011)

shasum -a 256 data/evidence/ga-health-cost-baseline-2026.json must print the sha above; a second run is byte-identical (cmp). The audit CSV is regenerated by the same command and is gitignored.

10. What it is not

• Not a population average. Nobody is "the median county"; the statewide figures weight each county once, not by residents, and no enrollment data enters.
• Not a prediction. The Floor and Ceiling are the filed plan terms for the plan year, read from public files; they do not forecast what any person will spend, what carriers will file next year, or what regulators will approve.
• Not subsidized cost. Gross premium only. Most Marketplace enrollees pay less after the advance premium tax credit; an ICHRA or employer contribution changes the picture entirely. The Baseline is the unsubsidized reference, deliberately.
• Not a care-profile simulation. There is no deductible arithmetic, no drug tier, no visit count, no SBC coverage example. The Floor assumes zero use; the Ceiling assumes the in-network individual cap is reached. Anything in between is the Surely Cost Engine's job, not this metric's.
• Not a plan recommendation. floorPlanId and ceilingPlanId name the plans that set the figures; they say nothing about networks, formularies or fit.

11. Changelog

• 1.0.1 (2026-09-02) — wording precision (offered throughout the county; in-network worst case; not a plan recommendation; Georgia Access named as the marketplace). No computational change; figures identical to 1.0.0.
• 1.0.0 (2026-09-02) — first public version. PY2026 RECORD artifact from the Georgia Access Rate (Sep 26, 2025), Plan Attributes (Dec 16, 2025) and Service Area (Dec 5, 2025) PUFs plus the CMS PY2026 Georgia SBE bundle (posted 2026); replay against the engine-built county index 157/159 (REPLAY.md §9).