Resolution: Killed
Caveats: Killed on the registered point-estimate clause; honest weights follow. (1) Guard margin: the resolvability floor cleared by exactly one fake (21 vs 20). One fewer flagged prestige fake and this test would have voided, so the verdict rests on the thinnest evidence the registration itself deemed admissible. The floor is a registered bright line and is honoured as such, but at 21-vs-35 fakes the >=5 point clause is statistically soft: the log-scale 95% CI on R is roughly (2.20, 6.50), which straddles the pin, and if the true ratio were exactly 5 an observed R <= 3.78 would arise about one time in six (one-sided p ~ 0.16). The kill is per the registered point-estimate rule - the same way the house's 10x pin died at observed 9.4 - not a statistically decisive exclusion of 5x. The direction, by contrast, is decisive: the CI floor is ~2.2, well above 1; prestige/seal classes really do carry a multiplied fake-rate. (2) Guard-design selection: under a no-enrichment null the prestige classes expect ~16.6 fakes, below the 20 floor, so the guard clears only ~23% of the time in null worlds - the test as designed mostly voids unless the direction already leans positive. That blurs the void/refuted boundary at the low end and means the guard is not hypothesis-neutral; it does not touch the observed R = 3.78. (3) Detected vs attempted: is_artifact_fake records forgeries CAUGHT, a per-class lower bound on forgeries attempted. Detection plausibly concentrates on prestige/seal objects (market value, publication and collection scrutiny) and under-reaches the 89,307-row administrative mass; both plausible biases inflate the measured R, so the true attempted-forgery ratio is, if anything, further below 5. The kill is robust in the plausible bias direction; only an unlikely prestige-under-detection regime could push the true ratio above the pin. (4) Concentration within the union class: 14 of the 21 prestige fakes sit in the Cylinder-seal cell (13.06x the admin rate); the genre arm proper (Official or display 6, Literary 1, all other genre cells 0) carries 7 fakes across 7,927 rows at only 2.08x admin. Cylinder seals alone would have blown past 5x but, at 14 fakes, would have voided the guard; the registered union is precisely what cleared the guard and simultaneously diluted the ratio below the pin - no sub-reading of the pinned classes both resolves and supports. Substantively, 'forgery has a genre' survives as direction, but the mechanism the data shows is object-market forgery (carved seals as collectibles) more than textual-genre prestige. (5) Disjointness pin: 5,074 admin-genre rows sit inside the prestige class (via seal artifact-type or a co-occurring prestige genre token; the artifact's univariate cells cannot split these) and were removed from the admin denominator by construction - a pro-prediction allocation. The 6,872 admin-genre rows excluded from class T carry 14 fakes at a 0.00204 rate (~4.8x class T itself), consistent with substantial overlap with the 14 Cylinder-seal fakes - much of the 'prestige' signal may be admin-genre seals. Even with this favourable construction, R fell short of 5. (6) Copy structure: the 126,000-row copy is a two-pass harvest (sequential pages 1-60 in artifact-id order plus a seeded uniform-random 80-page draw; 46,000 rows reachable only via the sequential pass) of a 421,501-artifact catalogue (~30%). The prediction explicitly self-scoped to this copy, so the verdict's internal validity is untouched; but the sequential arm over-represents low-id, long-catalogued material whose fake flags are likely more mature, so neither the class rates nor R should be quoted as full-CDLI estimates without a uniform-design draw. (7) Coverage: the pinned classes contain 56 of the copy's 160 flagged fakes; 104 fakes lie outside both classes. A follow-up owes the flags a full genre-by-artifact-type cross-tab (the compute ran 10 univariate cells, per the registered literal reading) before the 'genre of forgery' question is considered settled. (8) Disclosed priors: triage graded this ADJACENT and the registration is marked is_calibration=True; the aggregate fake count (160 of 126,000) was self-measured in-house before registration (the leaked-by-construction tension recorded in known_priors), so the guard floor was set knowing the total. The class-level rates and the ratio itself were computed blind, after registration (registered 11:23 UTC, computed 11:45 UTC 2026-07-19, ordering verified read-only). Calibration: fifth multiplicative-magnitude pin to die (>=3:1, >2x, <1/3, 10x, now >=5x - observed 1.69, 0.87, 1.28, 9.4, 3.78); direction decisively right, magnitude overpinned by a 1.32x shortfall; a >=3x pin would have survived on these numbers.
Registered against the IN-HOUSE CDLI catalogue copy (no external permission needed; lesson-4 trivially satisfied): ManuscriptRecord rows with tradition='cuneiform', source slug 'cdli', in the project's Docker Postgres - verified at registration 2026-07-19: source status active, exactly 126,000 rows, per-row raw fields genres_for_ws_e, artifact_type, is_artifact_fake (boolean), retired; copy harvested 2026-07-04 from the cdli.earth REST API (pages 1-60 sequential plus an 80-page random sample of the 421,501-artifact catalogue). The prediction self-scopes to this copy. PREDICTION VERBATIM: "Prediction (in-house CDLI copy): the per-row fake-rate among 'Official or display' (royal and monumental), Literary, and Prayer/Incantation genres, and among seal artifact-types, is >= 5 times the fake-rate among Administrative tablets (primary clause: prestige-genre:administrative fake-rate ratio >= 5). Disambiguation: fake-rate = flagged fakes divided by total rows within each genre or artifact-type class; classes are read from the genre and artifact-type fields. Coverage guard: with only 160 flagged fakes the per-cell counts are small, so require >= 20 fakes across the prestige classes combined for the test to resolve, else void; and note that the flag records DETECTED forgeries, so the pattern is of forgeries caught - a lower bound on forgeries attempted." CLASS DEFINITIONS PINNED AT REGISTRATION from the copy's actual value vocabulary (inspected today: values only - no fake-by-class cross-tabulation was run or seen): ROW UNIVERSE = the 126,000 rows minus rows with raw.retired = true (pinned; retired entries are catalogue-withdrawn). GENRE MEMBERSHIP = split genres_for_ws_e on ';', strip whitespace, exact member match. PRESTIGE CLASS (union; each row counted once) = rows with a genre member in {'Official or display', 'Royal Inscription'} (the prediction's own gloss 'royal and monumental' maps to both CDLI values) OR in {'Literary', 'Literary letter and letter-prayer'} OR in {'Prayer', 'Incantation'} OR with artifact_type in {'Cylinder seal', 'Stamp seal', 'seal (not impression)'} (seal OBJECTS; impression carriers 'sealing', 'Jar Sealing', 'bulla' are excluded - the forged collectible is the seal itself). ADMINISTRATIVE CLASS = rows with genre member 'Administrative', NOT in the prestige class (disjointness pinned; the overlap count is reported), and artifact_type in {'', 'tablet', 'tablet & envelope', 'envelope'} ('Administrative tablets' verbatim; blank artifact_type retained - CDLI leaves the default tablet type blank). QUANTITIES: fake-rate_class = rows with is_artifact_fake true / rows in class; R = fake-rate_prestige / fake-rate_administrative; if fake-rate_administrative = 0 with prestige fakes >= 20, R = +infinity (pinned).
Resolution criteria — the registered fine print
Resolution criteria: COMPUTE RECIPE (blind agent, read-only DB access): (1) select the row universe (tradition='cuneiform', source 'cdli', retired excluded) and report its size; (2) build the two pinned classes exactly as registered; (3) count is_artifact_fake in each; (4) compute R. CLAUSE PRECEDENCE: (1) INCONCLUSIVE_BY_DESIGN if fakes across the prestige classes combined number fewer than 20 (the prediction's own guard); (2) SUPPORTED if R >= 5 (including the pinned R = +infinity case); (3) REFUTED if R < 5. REPORTED, NON-VERDICT: the full per-cell table (each genre family and each seal type separately, with fake counts and rates), the prestige-and-Administrative overlap count absorbed by the disjointness pin, the retired-row count, and the copy's sampling structure (sequential pages 1-60 + 80 random pages) as a representativeness caveat on any generalization beyond the copy - the registered claim is about the in-house copy by the prediction's own scoping, so this caveat bears on interpretation, not on the verdict. Compute firewalled from threshold and direction.
Known-priors disclosure — what the registrant already knew
Known priors disclosure: Triage 2026-07-17 graded ADJACENT: that forgers target prestige classes is documented at length (Muscarella, The Lie Became Great, 2000; Michel-Friedrich, Fakes and Forgeries of Written Artefacts, 2020) - direction printed; but the prestige:administrative fake-rate ratio on CDLI's is-fake field is un-run and unpublished. DISCLOSED TENSION with the wave flags digest (breadth_w2_confidence_flags): the digest lists items 3/12/13 among 'self-measured in-house items ... leaked-by-construction'; the triage adjudication - the eligibility authority for this batch - held 012 ADJACENT because only the AGGREGATE fake count (160 of 126,000, stated in the prediction text itself) was self-measured by the generating agent, while the primary ratio was never computed; this registration follows the triage verdict and records the tension. The registrant's vocabulary inspection today retrieved DISTINCT VALUES of the genre and artifact-type fields only - no cross-tabulation with is_artifact_fake was run, so no verdict-bearing quantity is known. CALIBRATION NOTE: the verdict clause is a MULTIPLICATIVE >= 5x pin - the historically overpinned class (3:1 died at 1.69, 2x at 1.28, 10x at 9.4x); registered faithfully per the anti-threshold-shopping rule; selection accepted the risk because forgery-targeting base rates typically separate by an order of magnitude and the prediction's own >= 20-fake guard absorbs small-cell noise; the registrant's expectation is supported-but-not-certain, with the honest failure mode being an administrative fake-rate inflated by mass-produced fake Ur III administrative tablets, a documented market phenomenon.
Method and dataset — how it was measured
Mechanical application of the registered clause precedence to the blind compute artifact (docs/generated/compute_cdli_fake_rates_20260719.json), with arithmetic independently re-verified from the artifact's raw counts. (1) Coverage guard: prestige-class fakes 21 >= 20 floor, so the test resolves (margin 1). (2) Primary clause: R = rate_P/rate_T = (21/13,073)/(35/82,435) = 3.7834 < 5, so registered clause (3) REFUTED fires => verdict 'killed'. rate_T > 0, so the pinned R=+infinity branch is moot. Pre-registration ordering verified read-only in the DB: ConjecturePrediction (slug cuneiform-cdli-w2-012-forgery-has-a-genre, variant_key v1) registered_at 2026-07-19T11:23:15.496843+00:00, status open, no prior resolution row, is_calibration=True; artifact computed_at 2026-07-19T11:45:52.475906+00:00 strictly postdates it. prompt_hash convention: SHA-256 hex over the UTF-8 bytes of the resolution brief from the '## The registration' header line through the final line of the '## Your task' section ('...nothing else.'), headers inclusive, reproduced verbatim as 8 lines (3 section headers, 3 one-line section bodies, 2 blank separator lines), every line including the last terminated by a single LF.
Dataset: In-house CDLI catalogue copy: apps.scriptome ManuscriptRecord rows with tradition='cuneiform', source slug 'cdli' (source active), exactly 126,000 rows in the project's Docker Postgres; harvested 2026-07-04 from the cdli.earth REST API as the union of two passes (sequential pages 1-60 by artifact-id order, plus a seeded [20260704] uniform-random draw of 80 pages from the 422-page / 421,501-artifact space; 46,000 rows touched only by the sequential pass). Fields used per row: raw.genres_for_ws_e (';'-split, whitespace-stripped, exact label match), raw.artifact_type (NULL coerced to ''), raw.is_artifact_fake (boolean; 160 true), raw.retired (boolean; 0 true, so the registered retired-exclusion is vacuous and the universe is all 126,000 rows).
computed 2026-07-19