The pilot spans 24–36 policy lineages selected from a preregistered candidate frame, stratified so no institution, domain, or policy area is systematically excluded. Cases are chosen by materiality and methodological-learning value; the frame retains screened-out events so the pilot can test whether available data favor certain institutions or groups.
Scaffolding implemented in @gii/pilot · the coverage below is computed live from the illustrative frame (30 candidates, 6 screened out).
Frame lineages by institution
The pilot must span the branches, not concentrate where records are easiest to obtain — a flat profile here is the point of the design.
30 lineages in total · The frame
Frame lineages by instrument
Fiscal, regulatory, judicial, administrative, trade, and monetary actions each behave differently under the scoring rules, so each must appear.
30 lineages in total · The frame
Frame lineages by direction mix
Every lineage carries exactly one direction mix, so this ring does partition the frame. A pilot of only benefit cases would calibrate GII-D on half the world.
- mixed1137%
- benefit827%
- burden620%
- direction neutral517%
30 lineages in total · The frame
The materiality screen#
What may enter the frame at all is decided by materiality screen v0.1.0 — no longer prose. Three quantified gates, any one of which admits a candidate; the 0–100 materiality score then ranks candidates for selection but never gates. Two rules do the safety work: an unknown fact never passes a gate, and a missing fact earns zero score — being unmeasured is not a way to look material.
| Gate | Criterion (any one admits) |
|---|---|
| reach | Estimated exposure reaches at least 0.1% of the relevant base (reachShare ≥ 0.001). |
| scale | Estimated annualized fiscal or economic scale is at least $100M per year (annualScaleUsd ≥ 1e8). |
| durable-binding | A binding legal change expected to persist at least 3 years and reach at least 0.01% of the relevant base (binding ∧ durationYears ≥ 3 ∧ reachShare ≥ 0.0001). |
A screened-out candidate’s disposition is machine-generated from the measured facts — every failing gate with its value, every unmeasured fact named. For a fictional candidate reaching 0.02% of its base at $8M/yr, non-binding:
below materiality screen v0.1.0: reach 0.020% < 0.1% of the relevant base; scale $8.00M < $100M/yr; not a binding legal change
Every threshold below is a starting hypothesis for pilot calibration, challengeable through the methodology-review process (review packet, O8). The score reuses the codebook’s quantified anchor machinery — see magnitude anchors.
Stratification coverage#
A selection of eight illustrative candidates, and how it spans each design dimension:
| Dimension | Covered | Missing |
|---|---|---|
| Institution 7/7 | congress, president, executive-agency, independent-regulator, lower-court, supreme-court, federal-reserve | — |
| Domain 6/6 | people_households, us_economy, public_entities, nonprofits, international_trade, private_enterprise | — |
| Instrument 6/6 | fiscal, regulatory, judicial, administrative, trade, monetary | — |
| Exposure 2/2 | national, geographically-staggered | — |
| Direction mix 4/4 | benefit, burden, mixed, direction-neutral | — |
| Data environment 2/2 | high-data, low-data | — |
Preregistration#
Preregistration is mechanical, not rhetorical: an analysis plan — outcome map, orientation rules, design — is content-hashed and timestamped through a public third-party registry (e.g. OSF) and GII’s own publication registry before any outcome is inspected. Once locked, amending the plan changes the hash and is detectable, and a decision cannot be preregistered twice. The same hash-timestamp seals each release package.
Walking skeletons#
Before scaling ingestion, three end-to-end walking-skeleton lineages exercise the whole stack — register the decision, link its lineage, preregister the plan, score it, and build a reproducible release — on one lineage at a time. The @gii/pilot assembly wires registries, lineage, scoring, and release together.
Success criteria#
The methodology advances beyond beta only if it demonstrates all of the following:
- Acceptable reviewer agreement on magnitude components and orientation.
- Reproducible estimates and score builds.
- Stable lineage and non-duplication rules.
- Sensitivity ranges that do not routinely overturn central interpretations.
- Documented handling of forecasts, revisions, stays, and supersession.
- Coverage across institutions and domains without systematic silent exclusion.
- Understandable public presentation of magnitude versus direction.
- A viable correction and governance process.
Reviewer agreement is assessed with named statistics, not an undefined standard:
| Statistic | Applies to |
|---|---|
| Krippendorff's α (interval) | E, R, P, S magnitude components and I, Q, F, B, T confidence components |
| Weighted κ / ordinal α | evidence-level and maturity assignment |
| ICC(2,k) | composite magnitude |
Provisional acceptance is graded against two floors — a lower one for pilot learning and a higher one for production. The floors themselves are internal: a publication gate’s position states exactly how much reviewer disagreement this index tolerates before it stops publishing. No target score distribution is imposed; if a case cannot be scored credibly, withholding is a successful control, not a pilot failure.