The next epistemic frontier

Declare the measurement before you calculate it.

A joint number is not one thing. This workbench makes you name the system, population, event, and observation mode before it will emit a result.

It computes a static all-miss result only when every stated guard saw every stated item. For a deployed route or an adaptive attack, it records the missing protocol fields and refuses to launder those questions into a static score.

Choose what was observed

The same product labels can appear in three different measurements. Select one; the workbench will make its scope visible in the resulting packet.

Mode 01 · calculable here Static full exposure Every guard scores every applicable item independently. A reported union identifies the static all-miss count for this declared set. Mode 02 · trace required Deployed route An actual order, gate, router, or intervention determines which component sees which artifact and which terminal outcome occurs. Mode 03 · adversary required Adaptive holdout An attacker model, access budget, feedback channel, and sealed holdout are part of the measurement — not optional commentary.

Protocol draft

Stack Study v0.1

Static mode is selected. Enter a shared positive population and per-guard catches; a union is optional but is required to identify the observed static all-miss count.

Local only

JavaScript unavailableThe protocol fields remain readable, but this local calculator is intentionally unavailable without JavaScript. Use the tested reference implementation ↗ for the finite bounds instead; this page does not submit a form anywhere.

Study label

Declared system under test

Positive event definition

Keep it shared across the listed guards. A system label alone is not an event definition.

Positive items in the declared population

Version / configuration / route digest

Evidence access posture

Route or protocol identifier

Every listed guard was evaluated on the same named population and the same positive event definition. Every listed guard saw every applicable item. No earlier block prevented a later guard from scoring an item.

Static counts

Use counts, not percentages. The calculator will expose what follows from marginals alone, then show an exact static result only if you also retain the union.

Per-guard catch counts on the declared positive population.
Guard / component	Pinned version or threshold	Catches among positives	Remove guard

Union catches among positives (optional, but identifies static all-miss)

Do not supply an independence estimate here. The union is an observed aggregate from this exact shared-item evaluation.

No calculation yet. This tool does not retain your entries.

The study contract this enforces

A first real study should preserve enough structure to distinguish what was actually measured from what remains hypothetical. The preflight makes the first layer executable; the next layers are intentionally more demanding.

A · SUT

Freeze the workflow

Name components, versions, policy/configuration, shared dependencies, and route semantics. A changed configuration is a different system under test.

B · Mode

Keep observation modes apart

Mark static full exposure, an actual deployed route, replay/ablation, or adaptive holdout. Do not render one as evidence for another.

C · Trace

Retain route and terminal facts

For a route, retain stage sequence, intervention, missingness, and terminal outcome. A blocked input is not an observation for a later stage.

D · Revision

Make correction additive

Bind protocol, data manifest, runner, and report. Release a linked revision with reason rather than silently rewriting an evidentiary record.

MLCommons AILuminate technical users guide ↗ — a system under test can be a single model, an ensemble, or a mixture of models and algorithms.
NIST AI RMF: Measure ↗ — calls for documented measurements, uncertainty, and conditions similar to the deployment setting.
AWS Bedrock: ApplyGuardrail ↗ — illustrates that a guardrail can be placed at different points in an application flow and applied to input or output.
UK AISI Inspect logs ↗ — an example of append-only post-evaluation edit history with author and reason.