Scriptome track record
Every pre-registered prediction, scored
The falsification loop, extended to the Scriptome. Each row below was written down and frozen before the evidence that scored it existed. Wins and losses are shown with identical weight.
Live-read bets (excavated-corpus estimator, Waves 3b–5)
Read directly from the model YAML and ScriptomeMethodRun outputs every time this page renders — these numbers cannot drift out of sync with the underlying record.
The anti-artifact bet is vindicated: the ID-ordered Wave-3a batch measured Ur III at 83.9%; this wave's seeded random sample measures Ur III at 26.232% (95% CI 18.6-33.9%), inside the registered 20-28% band and far below the 40% ceiling. 5 of 9 period-mix bands hit; both directional claims passed. Scored exactly per the pre-registered, pre-evidence protocol -- no band or bucket-mapping rule was adjusted after seeing this outcome.
Scored mechanically as written against a seeded random sample -- a real, pre-registered result -- but not epistemically blind: the site-selection queries this project ran to design later bets (Karanis) show this project does look at live data while designing its own tests. Sound-not-blind is the honest description, not "blind," and not "invalid" either.
LOSE: the actual Tebtunis census count sits far above the band predicted by transporting Oxyrhynchus's own census-yield-per-recovered-piece rate. This is model-class information, not a reporting problem (goal's own framing): it directly tests whether the excavation x publication PRODUCT is safely treated as tradition-constant (the v1 estimator's working assumption) or varies materially by site. The measured answer here is that it varies -- Tebtunis's own yield rate is measured several-fold higher than Oxyrhynchus's.
A second LOSE, even after widening the band with the Wave-4b Oxyrhynchus->Tebtunis df=1 heterogeneity estimate: actual Karanis census 2,281 sits 3.7x above even the widened p95 ceiling (611), and ~21x the p50. Read plainly as model-class information (per the goal's own framing, exactly as Wave 4b's original LOSE was read): the SINGLE two-site heterogeneity estimate this project owns (Oxyrhynchus vs. Tebtunis, +/-0.71 dex) is NOT wide enough to bracket a third site's true deviation -- either genuine site-to-site dispersion in the excavation x publication yield rate is substantially LARGER than one pair can measure, or Karanis's own anchor (already flagged as the weakest in the table, banded from a collection-wide ceiling rather than a site-specific figure) is itself too low, or both. Both are real, named, undecided candidate explanations -- not adjudicated here, per the same discipline Wave 4b applied to its own LOSE. This does NOT widen site_heterogeneity's own prior after the fact (no re-editing, per the scoring rule above); it is named as a recommended Wave-6+ reconsideration instead (see the wave report's Phase 2 section).
Earlier bets (Waves 1–2)
Hand-recorded from each wave's own report; cited so any reader can verify the original scoring.
The century histogram implied by the Pinakes witnesses' own Siecle dating, predicted before the ManuscriptRecord mint, matched the mint's actual per-century counts exactly (20/20 periods, c01-c20). Honesty note recorded at the time: prediction and mint read the SAME Siecle field through the same parser, so this demonstrates pipeline fidelity (no drift between readout and mint), not forecasting skill -- the goal explicitly commissioned it as the cheap version of G1.
Source: docs/generated/scriptome_v1_wave1_report.md#3.5
Registered 2026-07-03T12:46:07Z before any vHMML fetch. Mostly wrong, reported plainly: the single biggest miss was scope, not magnitude -- the prediction modelled vHMML as an Eastern-Christian catalogue and did not anticipate that HMML's Reading Room dataset is numerically dominated by its West African (Mali/Timbuktu) Arabic-script digitization program (119,322 records), entirely unpredicted, alongside an unpredicted islamic majority (138,291). Only the armenian range hit; dated-fraction and century-mass both missed high (cataloguing better-dated and more recency-concentrated than predicted).
Source: docs/generated/scriptome_v1_wave2_report.md#2.6