Methodology

The model may be wrong. The evidence boundary must still be right.

CascadeLens is a transparent stress compiler. It makes a deliberately narrow promise: preserve what was known, what was assumed, how a result was produced, and what would reverse a decision.

01 · Evidence

Grades control eligibility, not decoration.

Official observations and entity reports enter the lower bound. Independently verified third-party records may enter the central bound. Text extraction and model inference remain bounded-only and cannot support the primary estimate.

official observedentity reportedthird-party verifiedtext extractedmodel inferred

02 · Time

Valid time and knowledge time are separate.

A fact can describe the past yet become available only later. WorldGraph records when a relationship was valid, when the exact source version became available, and when it was retrieved. Replays gate on availability at the frozen cutoff while allowing a later archival retrieval of that exact version.

03 · Cascade

Bounded daily propagation over a typed directed graph.

The built-in engine refreshes valid and knowable graph visibility on every simulated day, activates and retires shocks from their declared intervals, and solves each day to a declared tolerance under a strict iteration cap. Results distinguish the time-weighted mean, within-horizon peak, and final-day impact. Flow-share sums above one are reported rather than silently normalized.

impactₜ(v) = 1 − Π(1 − contributionᵢ,ₜ)
contributionᵢ,ₜ = impactₜ(u) × edge_weight × transmission

04 · Interventions

Activation first, Pareto frontier second.

Up to sixteen interventions are exhaustively enumerated. Budget, count, units, and mutual exclusion are checked before evaluation; work begins at the frozen decision cutoff and each feasible effect activates only at that cutoff plus its declared lead time. The product retains the do-nothing baseline and reports a separate cost–risk frontier and recommendation for every horizon.

05 · Observability

Ask which missing fact could change the action.

A candidate relation is evaluated across present and absent branches to estimate expected value of perfect information, probability of decision change, worst-case impact reduction, and decision-uncertainty reduction net of acquisition cost. The branch is counterfactual: CascadeLens does not relabel the candidate as verified evidence.

06 · Benchmark

Outcomes live in a separate partition.

Post-event observations cannot support graph inputs or shocks. Scoring requires a declared metric and horizon, a complete outcome window beginning with the first shock, availability only after that window closes, and an outcome-only source. With valid separated outcomes, the benchmark reports error, rank, direction, interval coverage and width, empirical coverage calibration error, and regret versus an explicit zero-impact baseline. With no comparable outcome—or for a synthetic stress—the result is explicitly scenario-only.

07 · RiskPack

Recomputation is necessary and still not validation.

A RiskPack holds the scenario, sealed graph, source manifest, assumptions, model card, all recomputation inputs, cascade bounds, intervention and observability outputs, benchmark status, limitations, rebuild command, and relative SHA-256 checksums. Verification deterministically recomputes every derived output. An external expected digest can detect a self-consistently repacked input, but neither mode proves predictive validity, publisher identity without that external receipt, or adoption.