Skip to content

Latest commit

 

History

History
57 lines (47 loc) · 3.29 KB

File metadata and controls

57 lines (47 loc) · 3.29 KB

Estimator Contract

Paper DOI: https://doi.org/10.5281/zenodo.20476200

v0.4.0 adds deterministic estimator helpers. They are not a substitute for the paper's scientific assumptions. Each helper accepts declared JSON input and emits a CertificateReport with metrics and a JSON-serializable certificate artifact. If a required assumption, overlap condition, support condition, calibration bridge, evidence budget, or target-validity field is missing, the report fails closed.

CLI

uv run altk estimate finite-sample examples/estimators/finite_sample.json
uv run altk estimate proxy-bridge examples/estimators/proxy_bridge.json
uv run altk estimate causal-effect examples/estimators/causal_effect.json
uv run altk estimate transport-diagnostics examples/estimators/transport_diagnostics.json
uv run altk estimate guard-risk examples/estimators/guard_risk.json
uv run altk estimate federated-pooling examples/estimators/federated_pooling.json
uv run altk estimate portfolio-selection examples/estimators/portfolio_selection.json
uv run altk estimate foundry-phase examples/estimators/foundry_phase.json
uv run altk estimate reproduction-phase examples/estimators/reproduction_phase.json
uv run altk estimate cara-time-to-target examples/estimators/cara_time_to_target.json
uv run altk estimate alpha-budget examples/estimators/alpha_budget.json

Estimator Kinds

Kind Output artifact Fail-closed condition
finite-sample confidence-accounting empty sample, invalid range, invalid confidence
proxy-bridge proxy-bridge-record invalid calibration or nonpositive bridged lower bound
causal-effect causal-certificate-shaped artifact missing identification, unsupported mode, no overlap
transport-diagnostics transport-diagnostics uncovered target support, radius above threshold, excessive dimension
guard-risk guard-risk-record invalid alpha, too little calibration data, risk above budget
federated-pooling federated-pooling invalid record, invalid correlation, nonpositive pooled lower bound
portfolio-selection submodular-selection-result malformed portfolio or budget
foundry-phase phase-control-record missing source/sink or min-cut below demand
reproduction-phase recombination-estimate invalid matrix, capacity mismatch, unidentified recombination
cara-time-to-target CARA-process-shaped artifact invalid target, raw-net insolvency, no time-to-target margin
alpha-budget evidence-budget exhausted finite evidence budget

Non-Python Rule

A non-Python implementation may use different numerical libraries, but L5 conformance requires the same input/output shape and the same fail-closed decision on the bundled examples and conformance fixtures. Numerical values may be represented in the host language's native floating type, but report keys, predicate names, and artifact fields must remain stable.

Scientific Boundary

Estimator success means the declared data passed the reference arithmetic. It does not certify unobserved assumptions. Causal identification, transportability, root authority, recombination structure, and ALT-CARA target validity must still be declared and independently justified before a packet can enter settlement.