Reference Class Model

Assumptions

Everything the model believes, stated before the math. If you do not buy one of these, do not buy the output.

1. Selection dimensions are pre-outcome attributes only - project type, state, capacity band; an outcome field can never define the analog set.
2. The cohort freezes before outcomes are read, at every hierarchy level.
3. Semantic mismatch is rejected: an analog must share the subject's asset class unless a documented shared-target contract exists - none does in this registry.
4. Sparse analogs stay UNKNOWN with the fallback path shown; no shrinkage is passed off as measurement.
5. Alternative-cohort sensitivity prints at every level, including levels looser than the one used.
6. In-flight analogs are censored, so future outcomes cannot leak into the read.
7. Embedding similarity is discovery, not comparability - no embeddings are used here.

Derivation

Analogy is decided on attributes knowable before outcome: project type, state, capacity band (small under 100 MW, medium under 500 MW, large above). The hierarchy relaxes in a fixed order - same type + state + band; same type + state; same type + band; same type - and the first level with at least 10 resolved analogs is the cohort. Type equality is non-negotiable at every level.

For the chosen cohort the outcome distribution is the withdrawn share among resolved members with a Wilson interval (shared with M22):

over the resolved members only, with censored members printed, never dropped. Every level of the hierarchy prints its own counts and, where resolvable, its own interval - the sensitivity of the answer to cohort choice is on the page, not behind it.

Identification strategy

Why the estimate supports the claimed meaning - and where the model refuses to claim more than the data carries.

• Pre-outcome selection is the anti-leakage rule: because no outcome field can define membership, the cohort cannot be chosen to fit the answer.
• The hierarchical fallback is the honesty mechanism for sparsity: when the strict cohort is thin, the model shows exactly where it relaxed and prints that transfer is limited accordingly - instead of quietly widening the net.
• The minimum of 10 resolved analogs gates any distribution: below it the answer is UNKNOWN with the hierarchy shown, never a shrunken number presented as measured.
• Rejection of semantic mismatch keeps unlike assets out: a gas plant's history never becomes a solar project's prior without a documented shared-target contract.

Inputs and outputs

Calibration and gate evidence

Validation design

• Gate-verified discipline: the hierarchy prints at every level with sensitivity, at most one level is used, outcome fields are excluded from selection, censored analogs print, and a sparse subject stays UNKNOWN.

Evidence on record

• Smoke gates at seal (7 m34-specific): hierarchy prints at every level with sensitivity; exactly one level used or none; outcome fields excluded from selection; censored analogs printed, not dropped; a sparse or unknown subject stays UNKNOWN, never shrunk; determinism (identical input hash on re-run); audit row written for every run.

Gates at seal

• smoke: 7 m34 gates incl. pre-outcome selection, sparse-stays-UNKNOWN, determinism + audit row (suite counts at seal per the build agent's seal report)
• sealed at 36389cf on atlas-nl-shock-redesign

Failure modes

The ways this model declines, degrades or goes UNKNOWN - and what it prints when it does.

• No hierarchy level reaches 10 resolved analogs: no analog set is claimed; the hierarchy shows where it thinned.
• Subject attributes missing (unknown type or capacity band): cohorts built on partial keys degrade gracefully and the fallback path prints.
• In-flight analogs: censored and printed, so the resolved-only rate is not inflated by unfinished projects.

Sensitivity

• The choice of hierarchy level moves the interval: looser cohorts are wider-population but weaker-resemblance, and both the choice and its cost print.
• Capacity-band edges matter near the 100 / 500 MW boundaries; a subject near a boundary may belong with the neighboring band, and the relaxed levels show what that change would do.