Event-Localized Information Value

Use event_localized_information_value to identify both how much a declared event is worth learning and which regions of a finite uncertainty support create the value of resolving a coordinate. The v1 evaluator is exact and Python-only.

import json
from pathlib import Path

from voiage import event_localized_information_value

request = json.loads(
    Path("specs/frontier/event-localized-information/v1/fixtures/normative/input.json")
    .read_text()
)
result = event_localized_information_value(request)
payload = result.to_contract_dict()
print(payload["event"]["perfect_gross_voi"])
print(payload["density"]["information_value"])

The canonical policy-relative density is f(x) [max_a g_a(x) - g_a*(x)], where a* is the declared baseline-optimal reference action. It is nonnegative at every finite atom. The separately labelled centered diagnostic f(x) [max_a g_a(x) - V0] can be negative locally even though its sum recovers the same coordinate information value. The result retains complete ties, integral errors, modes and vectors from the declared base coordinate.

The v1 objective is higher-is-better expected value or expected utility and therefore has direction: "maximize". Convert a cost/loss objective to a higher-is-better score before calling the evaluator; raw minimization is rejected. The reference action must attain the true baseline maximum. tie_tolerance lies in [0, 1e-6], while integral_tolerance must lie in (0, 1e-6]. Assurance integrals use the unrounded atoms; tolerance cleanup is applied only to the portable presentation.

The same contract evaluates perfect revelation of a declared event and every requested accuracy of a symmetric imperfect binary channel. Accuracy 0.5 is uninformative; p and 1-p have equal gross value because their signal labels can be inverted. The symmetry diagnostic uses tolerant accuracy matching and is null, not zero, when the requested grid contains no complementary pair. The portable result includes state/coordinate membership evidence, allowing its strict validator to re-evaluate the declared threshold or state set and every event, channel and density marginal rather than trusting reported summaries. Equal coordinates must appear as one grouped density atom.

voiage --format json calculate-event-localized-information request.json

plot_information_density(result) and plot_event_accuracy_curve(result) use the result only. They do not re-evaluate the model or infer observations from a plot. This finite probability-mass implementation does not claim continuous density estimation. Monetary buying-price information is owned by the #595 expected-utility pricing family. Rust, R and Julia are unsupported; Mojo is an external upstream boundary. Scientific review and stable promotion remain separate gates.