Public evidence record

person_scores.k1 — Frozen capability contract v3.6, real_world_cross_sectional_point_estimation. PHQ-9, CES-D-8 and EURO-D across NHANES 2015-2016, NHANES 2017-2018 (locked), ESS Round 7, SHARE Release 9 waves 9 and 8 (locked); 78 country-cycles. Agreement with conventional estimators 0.92-0.94 on the two NHANES cycles, 0.97-0.998 on ESS7, and 0.9851 pooled over 66,812 persons in 28 countries on SHARE wave 9.

• Agreement with a conventional estimate is not accuracy against truth: METER and its comparators share graded-response assumptions and can be wrong together.
• Cross-sectional only; one construct family (depressive symptomatology); two instrument families.

person_scores.k2_supplied — Frozen capability contract v3.6. Expressible and returned with a named warning; the confirmatory synthetic benchmark established 3-5 factors, not 2. No confirmatory evaluation of its own under the v0.6.0-beta.1 generation.

• Returned with a named warning; no K=2-only confirmatory person-score evaluation exists in any generation.

person_scores.k3_to_k5_supplied — Frozen capability contract v3.6; confirmatory synthetic benchmark plus one external transfer (IPIP, artifact D9E). Person scoring only. The structure is supplied by the user and is never discovered by METER.

• Person scoring only: no item-parameter claim of any kind attaches to a multi-factor structure.
• Evidence is confirmatory synthetic plus a single external transfer, not the multi-dataset real-data transfer that supports single-factor person scoring.

item_discrimination.k1 — B4X-3 confirmatory evaluation, criterion C2, sealed corpus B4X_EVAL_V1; reports/b4x/b4x3_evaluation.json at commit 2d12386. PASS on the across-seed median and on all three individual seeds. Per-seed medians 0.922385 / 0.908772 / 0.929348 against a line of >= 0.864551 (higher is better); panel n = 202. The capability rests on all three seeds passing, not on the seed that happens to serve.

• Prospective synthetic known-truth evidence only. External real-data item-parameter validation remains ongoing.
• Estimates are reported on the model's own logit metric; the latent scale is the run's own standardization, not an external anchor.
• The C2 criterion judges recovery of item ordering, which is a correlational claim, not agreement in absolute units.
• The line is a non-inferiority bar against a conventional mirt specialist (specialist minus 0.10), not an absolute quality bar.

item_thresholds.k1 — B4X-3 confirmatory evaluation, criterion C3, sealed corpus B4X_EVAL_V1; reports/b4x/b4x3_evaluation.json at commit 2d12386. PASS on the across-seed median and on all three individual seeds. Per-seed threshold RMSE medians 0.459003 / 0.376687 / 0.342942 against a line of <= 0.500245 (lower is better); panel n = 202. The capability rests on all three seeds passing, not on the seed that happens to serve.

• Prospective synthetic known-truth evidence only. External real-data item-parameter validation remains ongoing.
• Estimates are reported on the model's own logit metric.
• C3 measures RMSE after a per-world linear alignment to the generating truth. That alignment requires known truth and therefore cannot be performed on an uploaded dataset.
• The line is a non-inferiority bar against a conventional mirt specialist (specialist plus 0.15), not an absolute quality bar.

item_thresholds.k2 — B4X-3 confirmatory evaluation, criterion C4. The across-seed median passes (0.420973 against a line of <= 0.448627) and two of three seeds pass individually, but seed 20260901 returns 0.450421, worse than the line by 0.001794. The frozen rule requires the across-seed median AND every individual seed, so C4 is ineligible. Per-seed values are in evidence_record.b4x3.criteria; they are not a ranking, and which seed serves is unrelated to this criterion.

• METER withholds these estimates rather than serving them under a designation the evidence did not earn.

item_discrimination.k_ge_2 — B4X-3, reported descriptively and explicitly excluded from adjudication. Per-seed medians 0.787175 / 0.761324 / 0.810682 on a panel of n = 181, all below the 0.842115 line. Relative to the historical 0.7470-0.7875 development band, one median (0.810682) sits above it and two (0.787175, 0.761324) sit inside it. Under the frozen rule this is evidence about corpus difficulty, not about METER capability, in either direction; it could neither rescue nor veto the designation.

• METER withholds these estimates.

factor_discovery — Milestone 7 closure (2026-08-06); frozen capability contract v3.6, battery_structure_discovery_1_to_5. Never offered. Structure is an input you declare, not an output METER produces. Four smoke runs each reported every world as one factor, with balanced count accuracy 0.200.

• Refused for every dataset, in every release to date.

latent_structure_discovery — Owner directive, release 0.2: latent-structure discovery is not supported unless separately governed by another released capability. No released capability governs it. Listed separately from factor-NUMBER discovery so that neither can be read as permitted by the other's absence. METER is told the structure; it does not infer which items belong to which factor, nor how many factors exist.

• Refused for every dataset. If a future released capability governs structure discovery, it will carry its own ledger row and its own evidence.

item_parameters.k_ge_3 — B4X-3 closeout section 9: 'Anything at K>=3 remains NOT TESTED on any estimand.' The B4X-3 corpus acceptance rule admitted only worlds with K <= 2, so no K>=3 estimand was evaluated in this generation.

• Absence of evidence, not evidence of absence. Withheld because untested, not because it failed.