The Tyranny of the Few

Resolution criteria: VERBATIM PREDICTION: 'in each of the three corpora - witnesses per Sumerian literary composite in CDLI, copies per work among the DCLP literary papyri on papyri.info, and occurrences per epigram type in DBBE - the top decile of works ranked by witness count carries at least 50% of all witness records, and the three top-decile shares fall inside one band 20 percentage points wide; coverage guard: a corpus counts only if it supplies at least 300 distinct works and 1,500 witness records with secure work-to-witness assignment.' KILL VERBATIM: 'any qualifying corpus below 50%, or a cross-corpus spread wider than 20 points, kills.' OPERATIONALIZATION: CDLI = witnesses (exemplar tablets) per Old Babylonian Sumerian literary composite, via CDLI's composite(Q-number)-to-exemplar structure, query route documented; DCLP = copies per literary work using DCLP/Trismegistos work-identity as exposed on papyri.info, unassigned items excluded and the assignment-coverage rate reported; DBBE = occurrences per epigram type via type-occurrence links (the same instrument route as tonight's A-009 census). TOP-DECILE RULE (mechanical, shared): rank works by witness count descending with ties broken by stable identifier order; the decile = the first ceil(N/10) works; share = their witness records over all witness records. FULL PARTITION: corpora meeting the guards qualify; fewer than 3 qualifying -> INCONCLUSIVE-BY-COVERAGE; all 3 qualifying: SUPPORTED iff every share >= 50% AND (max - min) <= 20pp, KILLED otherwise (the kill clause is the exact complement - no dead zone). FIREWALL NOTE: the compute measures the identical statistic on FOUR corpora (these three plus the sibling B-002's Cantus data) in one blind run that is not told which corpus any hypothesis distinguishes.

Known priors disclosure: Triaged adjacent 2026-07-20. Disclosed neighbors, no numbers carried: the 2026-07-04 pilot resolved a DBBE/Pinakes heavy-tail item (supported-as-calibration), so the registrant knows DBBE occurrences-per-type is qualitatively heavy-tailed, and knows the pure-Yule-fails-in-Pinakes nuance; no top-decile share has ever been computed on any of the three pinned corpora. Preferential-attachment copying is established theory (the claim's premise); the wager is the cross-medium BAND, which is uncomputed. Sibling registration B-002 (same batch) wagers the opposite direction for liturgical transmission on the same statistic - the paired law-and-exception design; one blind compute serves both.

Register-before-compute held (registration 18:04:34Z, packet 9290244). Mechanical top-decile rule as registered (rank by witness count, stable-id ties, first ceil(N/10)). Results: top-decile shares CDLI 66.32%, DCLP 62.74%, DBBE 55.66%; all three corpora meet the guards (>= 300 works, >= 1,500 assigned records); spread = 10.66pp. REGISTERED PARTITION: all three qualifying, every share >= 50%, spread <= 20pp -> SUPPORTED.

Dataset: Three corpora, one statistic, one blind run: CDLI Old Babylonian Sumerian literary composites (391 works, 4,059 witness records from the advanced-search CSV export matching the engine's own count; 443 composite root-records correctly excluded; 18.61% unassigned excluded); DCLP via a sparse clone of papyri/idp.data at commit 64be870 (2,091 works with Trismegistos authorwork identity, 10,853 assigned witness records, 46.2% honestly unassigned - genuine scholarly non-identification, 0 parse errors); DBBE types/occurrences via its search_api (5,224 types, 13,084 type-linked occurrence records; the exact unassigned rate bounded-not-fabricated after a server-side pagination bug was politely abandoned).

computed 2026-07-20