cvprofiles · construct-validity profile

Trust — country + LLM

Admissible measurement set under a stated nomological network, and the image of β on that set. Empty sets and wide ranges are scientific features.

Construct-identified range

[L, U] · method = min/max B*

[0.106829, 0.481254]

Width = 0.374424 · 5 survivors · bootstrap percentile band = [-0.039153, 0.730078]

Target β = ols_coef on outcome log_gdp_pc. Range uses only measures in M*; rejected β values are shown below but never enter [L,U].

Run freeze

run_idde21a795a0dee6202efdf3928900da59ba74c174ed5ba48124c7e28adb5806a2
package3.0.2
scores_hash0b115ce85abef8fdd43764332a20f660534099656635cc553d279bc5f6254860
network_hash27490da9292c755027cb295162238dbfda45e5f032d037775bf1d10e21b951e4
beta_hash40040c2105c1d9b34e212dddd8ca9487fdb315d4db2ebee0ebaf29129df5cafe
δ0.0
seed20260814
networktrust_country_llm_extension

Bootstrap inference · additive to headline range

Units-only Efron resampling; percentile band over non-empty replicates. The headline [L, U] above remains min/max β(M*) on the full sample.

replicates1000
seed20260814
quantiles[0.025, 0.975]
non-empty999
empty1 (0.001)
degenerate0 (0.000)
band[-0.039153, 0.730078]

Uncertainty band · additive to headline range

Per-side α/2 quantiles over non-empty bootstrap replicates (α = 0.1). Heuristic band — not a confidence interval, not a holdout-robustness band. The headline [L, U] above remains min/max β(M*) on the full sample.

α0.1
κ2.0
quantiles[0.05, 0.95]
band[-0.016141, 0.698761]
empty rate0.001
degenerate rate0.000

Boundary attribution · |margin| ≤ κ·SE

measuremarginSEboundary
m_gps_trust -0.197064 0.156197 yes
m_trust_general 0.306865 0.121647 no
m_trust_in_group 0.097667 0.116052 yes
m_trust_out_group 0.219973 0.114168 yes
m_trust_institution -0.562602 0.125151 no
m_trust_composite -0.051991 0.121806 yes
m_llm_trust_llama 0.196624 0.132026 yes
m_llm_trust_phi -0.158092 0.188466 yes
m_noise -0.373876 0.157017 no

Admission frequency p̂m

measurep̂m
m_gps_trust 0.127
m_trust_general 0.997
m_trust_in_group 0.909
m_trust_out_group 0.945
m_trust_institution 0.001
m_trust_composite 0.788
m_llm_trust_llama 0.901
m_llm_trust_phi 0.167
m_noise 0.007

Heuristic uncertainty band over non-empty bootstrap replicates (selection uncertainty on the pooled sample; holdout verdict is a full-sample point finding outside the band). Not a confidence interval; not a holdout-robustness band. Boundary attribution: |margin_m| <= kappa * SE_m.

θ-grid sensitivity · diagnostic viewport

λ scales threshold magnitudes only. λ is never auto-selected; the headline is always λ=1.0. This grid is not part of the freeze preimage.

λ|M*|emptyLU
0.5 5 no 0.106829 0.481254
1.0 5 no 0.106829 0.481254
1.5 4 no 0.191812 0.481254
2.0 3 no 0.266456 0.481254

θ-anchors · pre-data audit

One pre-data anchor per restriction (hash 68e155f6775d…). The engine enforces completeness + hash; "pre-data" is a process commitment, not a temporal claim.

restrictionkindpre-dataanchor
conv_rule_of_law literature yes aghion2010_h5 — Trust–institutions correlations typically ≥0.4 (Aghion, Algan, Cahuc, Shleifer 2010). H5 historical floor 0.30.
disc_gini literature yes bjornskov2008_h5 — Bjørnskov 2008: negative trust–inequality relationship. H5 historical θ=0.10 on corr_sign.
conv_edu_hold literature yes falk2018_table_v — Falk et al. 2018 Table V: trust increases with cognitive ability in almost all countries. Leftover country-level education bar θ=0.20.

Menu admission

M* size = 5 / 9. Slack rule: admit if sr(m) ≥ −δ for all r.

measure status β(m) failed restrictions
m_gps_trust rejected 0.109023 conv_rule_of_law
m_trust_general admissible 0.390598 —
m_trust_in_group admissible 0.191812 —
m_trust_out_group admissible 0.266456 —
m_trust_institution rejected -0.092285 conv_rule_of_law
m_trust_composite admissible 0.106829 —
m_llm_trust_llama admissible 0.481254 —
m_llm_trust_phi rejected 0.386442 disc_gini
m_noise rejected 0.028351 conv_rule_of_law, disc_gini

Slack matrix sr(m)

Positive / zero ≈ satisfied at δ=0.0. Negative = violation magnitude.

measure conv_rule_of_law disc_gini conv_edu_hold
m_gps_trust -0.176415 0.003543 -0.197064
m_trust_general 0.404767 0.421874 0.306865
m_trust_in_group 0.182494 0.509629 0.097667
m_trust_out_group 0.358501 0.219973 0.328627
m_trust_institution -0.424613 0.175068 -0.562602
m_trust_composite 0.104768 0.407739 -0.051991
m_llm_trust_llama 0.475050 0.196624 0.221417
m_llm_trust_phi 0.361173 -0.158092 0.050590
m_noise -0.373876 -0.004827 -0.101694

Network (researcher-owned)

Engine does not invent restrictions. Agent may author oracle R for synthetic DGPs only.