Pilot dress rehearsal

The historical pilot waits on three things this repository cannot manufacture: real 2015–2025 decision records, provisioned source credentials, and blinded human reviewers. It does not have to wait to answer a fourth question — whether the pipeline survives a pilot-sized corpus. So it runs one. Thirty fictional policy lineages, spanning the entire stratification, go through preregistration, the codebook gate, scoring, the withhold rule, sensitivity, series assembly, reviewer calibration, review, dossiers, and release packages together, and the page below reports what happened.

Run at build time from @gii/rehearsal · corpus digest c8436c23b71f83c21995a497… — a rebuild reproduces it byte for byte.

Withholding, at scale#

Half the corpus does not publish an observed headline — and that is the design working, not failing. A single worked example can demonstrate one condition of the rule. A corpus is what demonstrates every one of them, and this corpus is authored so each fires somewhere — the table below is the count, not a claim about it.

Conditions overlap: a forecast-stage record typically trips several at once, and every withheld case states all of its reasons rather than the first one found.

What reaches the series#

Running thirty lineages forces a question one lineage never raises: does a case whose headline is withheld still contribute to the national series? It does not. forbids averaging forecast and observed values, and withholds an observed headline for cause — letting a withheld case into the aggregate anyway would launder it back into publication. The series therefore carries the 15 publishable observed headlines, and the other 15 are excluded and counted.

National impact pressure, 2015–2025

The rebasing is provisional and reference-window dependent; the raw measure is the one that means something. Exposure profiles taper, because attributable exposure to a specific decision decays as later actions accumulate over it.

Corpus-scale invariants#

These are the claims that only a corpus can test — properties of thirty lineages together rather than of any one of them. Each reports what was observed, whether or not it held.

Reviewer calibration at corpus size#

Five synthetic reviewers rate every lineage, disagreeing more on the cases the corpus marks as hard to judge. The point is not the numbers — real reviewers will produce different ones — but that the statistics compute, grade, and separate on a pilot-sized rating matrix at all, and that they say something specific about which judgments are the difficult ones.

Graded against a pilot-learning floor and a higher production floor. See calibration for the statistics themselves, validated against documented reference datasets.

The corpus#

Authored, not sampled. Each lineage names the methodological lesson it carries, and the frame it projects into is the same one the pilot design page renders — a test pins them together so they cannot drift.

30 lineages plus 6 screened-out events kept with their dispositions, so the pilot can detect whether available data favor certain institutions or policy areas.

Run it#

npm run gii -- rehearse                  # the summary above, in a terminal
  npm run gii -- rehearse --cases 10       # a bounded run
  npm run gii -- rehearse --json           # the full report

  GET /api/v1/rehearsal
  npm run gii -- export pilot-corpus.json  # the corpus itself
  npm run gii -- export rehearsal-cases.csv