Metric Definitions
Pinned formulas, scopes, missing-data rules, and version identifiers.
entropy_delta_mean.v1
- Display name
- Signed mean entropy delta
- Version
- v1
- Formula
mean_i(H_i(B) - H_i(A))- Inputs
- latest valid entropy_delta reading per active model
- Unit
- bits
- Scope
- latest valid entropy_delta reading per active model
- Missingness
- exclude null, non-numeric, and non-finite readings
- Aggregation rule
- unweighted arithmetic mean over valid models
- Source bundle
- runtime-supplied
- Valid / excluded
- runtime-supplied / runtime-supplied
mean_abs_entropy_delta.v1
- Display name
- Mean absolute entropy delta
- Version
- v1
- Formula
mean_i(abs(H_i(B) - H_i(A)))- Inputs
- same included entropy_delta readings as entropy_delta_mean.v1
- Unit
- bits
- Scope
- same included readings as entropy_delta_mean.v1
- Missingness
- exclude null, non-numeric, and non-finite readings
- Aggregation rule
- unweighted arithmetic mean of absolute values
- Source bundle
- runtime-supplied
- Valid / excluded
- runtime-supplied / runtime-supplied
cii_mean.v1
- Display name
- Mean Composite Interest Index
- Version
- v1
- Formula
mean_i(CII_i)- Inputs
- valid per-model CII readings
- Unit
- unitless
- Scope
- valid model CII readings in the selected telemetry window
- Missingness
- exclude null, non-numeric, and non-finite readings
- Aggregation rule
- unweighted arithmetic mean over valid model CII values
- Source bundle
- runtime-supplied
- Valid / excluded
- runtime-supplied / runtime-supplied
window_entropy_gap.v1
- Display name
- Character-window entropy gap
- Version
- v1
- Formula
mean_j(abs(H(window_j(A)) - H(window_j(B))))- Inputs
- paired response A, paired response B, window_chars
- Unit
- bits
- Scope
- paired non-overlapping character windows of equal window_chars
- Missingness
- pair through min(number_of_A_windows, number_of_B_windows)
- Aggregation rule
- unweighted mean over paired complete windows
- Source bundle
- runtime-supplied
- Valid / excluded
- runtime-supplied / runtime-supplied
window_entropy_gap_pointwise_p.v1
- Display name
- Pointwise permutation p-value
- Version
- v1
- Formula
(1 + count(T_null >= T_observed)) / (1 + B)- Inputs
- observed window entropy gap, B paired-label permutations
- Unit
- probability
- Scope
- one separately tested window_chars value
- Missingness
- not evaluated when either response has no complete window
- Aggregation rule
- one plus exceedances divided by one plus B
- Source bundle
- runtime-supplied
- Valid / excluded
- runtime-supplied / runtime-supplied
window_entropy_gap_maxT_z_p.v1
- Display name
- Standardized max-T family-wise permutation p-value
- Version
- v1
- Formula
(1 + count(max_w(z_null,w) >= z_observed,w)) / (1 + B)- Inputs
- observed gap at each estimable window_chars, joint paired-label permutations, per-window null mean and standard deviation
- Unit
- probability
- Scope
- joint window_chars family
- Missingness
- not evaluated when no window size is estimable
- Aggregation rule
- maximum standardized null statistic per permutation across the jointly estimable window family
- Source bundle
- runtime-supplied
- Valid / excluded
- runtime-supplied / runtime-supplied
- Status
- candidate_global_test_pending_human_selection
relevant_char_fraction.v1
- Display name
- Matched-condition differing-character fraction
- Version
- v1
- Formula
differing_character_positions / compared_character_positions- Inputs
- paired response window A, paired response window B
- Unit
- fraction
- Scope
- one paired response window
- Missingness
- not evaluated when both paired windows are empty
- Aggregation rule
- position-wise comparison; missing trailing characters in the shorter paired window count as differences
- Source bundle
- runtime-supplied
- Valid / excluded
- runtime-supplied / runtime-supplied
window_dilution_regression.v1
- Display name
- Window-gap stimulus-dilution regression
- Version
- v1
- Formula
gap = intercept + slope * relevant_char_fraction + residual- Inputs
- window_entropy_gap observations, relevant_char_fraction.v1 observations
- Unit
- bits per differing-character fraction
- Scope
- paired-window observations for one provider/model sweep
- Missingness
- slope is undefined when fewer than two x values vary
- Aggregation rule
- unweighted OLS with nonparametric paired-window bootstrap uncertainty and model-level residual reporting
- Source bundle
- runtime-supplied
- Valid / excluded
- runtime-supplied / runtime-supplied
- Status
- diagnostic_only_no_threshold_tuning