You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[Parity #478] Prove estimator hyperparameter, stopping-rule and work-budget equivalence in every performance row #501
A performance comparison is invalid if Flow and sklearn silently perform different configured work. Extend the parity contract to record and compare every performance-relevant setting: solver/algorithm, regularization, tolerance, max_iter, n_init, n_estimators, max_depth, feature subsampling, bootstrap, kernel/gamma, random seed, convergence/stopping rule and effective iteration/tree/support-vector counts.
Defaults must be materialized rather than assumed, because sklearn defaults can change by version.
Acceptance: each canonical/operation row contains resolved sklearn and Flow config plus effective-work diagnostics; mismatched work budgets are not performance-eligible; version changes that alter sklearn defaults fail/flag contract regeneration; benchmark docs expose deliberate semantic differences rather than hiding them.
Concrete result from the work-equivalence audit: timing_comparable currently conflates two distinct claims. Product-level timing only needs semantic parity, while an implementation-efficiency claim additionally needs configured-work and realized-work parity.
Current benchmarks/parity_contract.json has at least four rows that cannot yet obtain a strict configured-work certificate from the committed metadata despite timing_comparable: true: Iris LogisticRegression (lbfgs_no_line_search, max_iter=200 vs lbfgs, max_iter=1000), Digits LogisticRegression (same mismatch), Iris PCA (power_iteration vs unresolved auto), and Diabetes Ridge (Flow iterative work fields vs no resolved sklearn solver/work metadata).
Also, equal configured budgets are insufficient: e.g. identical KMeans n_init/max_iter/tol/seed can still realize 7 vs 19 Lloyd iterations, so C_flow = C_ref does not imply W_flow = W_ref.
Defaults such as solver=auto should be materialized into a resolved config before certification; effective work should remain an estimator-specific vector rather than a universal scalar.
Full derivation + executable Flow certificate model are now in OpenProjects: ml/work-equivalence-certificates-for-benchmarks.md ml/work-equivalence-certificates-for-benchmarks.flow
Activity
godofecht commentedon Aug 25, 2026
Concrete result from the work-equivalence audit:
timing_comparablecurrently conflates two distinct claims. Product-level timing only needs semantic parity, while an implementation-efficiency claim additionally needs configured-work and realized-work parity.Current
benchmarks/parity_contract.jsonhas at least four rows that cannot yet obtain a strict configured-work certificate from the committed metadata despitetiming_comparable: true: Iris LogisticRegression (lbfgs_no_line_search, max_iter=200 vslbfgs, max_iter=1000), Digits LogisticRegression (same mismatch), Iris PCA (power_iterationvs unresolvedauto), and Diabetes Ridge (Flow iterative work fields vs no resolved sklearn solver/work metadata).Also, equal configured budgets are insufficient: e.g. identical KMeans
n_init/max_iter/tol/seedcan still realize 7 vs 19 Lloyd iterations, soC_flow = C_refdoes not implyW_flow = W_ref.Proposed derived eligibility:
product_eligible = semantic_parityimplementation_eligible = semantic_parity && configured_work_parity && realized_work_parityDefaults such as
solver=autoshould be materialized into a resolved config before certification; effective work should remain an estimator-specific vector rather than a universal scalar.Full derivation + executable Flow certificate model are now in OpenProjects:
ml/work-equivalence-certificates-for-benchmarks.mdml/work-equivalence-certificates-for-benchmarks.flowNo hosted CI/workflows were touched.