Criticality Spectrometer

Measure node criticality as a curve across adaptation horizons in AND/OR dependency systems.

Most network rankings assign one importance score per node. Criticality Spectrometer instead removes each node, sweeps the time at which substitutes become available, and records mission loss at every horizon. The result distinguishes a node that must be protected now and later from one whose risk can be reduced by enabling alternatives.

Install

pip install criticality-spectrometer

That puts the criticality-spectrometer command and the Python API on your system. From v0.1.2, the demo fixtures are also bundled with the package, so the 60-second demo needs no clone:

criticality-spectrometer example canonical --output model.json
criticality-spectrometer run model.json

The full worked examples, tests, and tutorial ship with the source; clone the repository to explore those:

git clone https://github.com/AMBRA7592/criticality-spectrometer.git
cd criticality-spectrometer

Run it in 60 seconds

criticality-spectrometer run examples/canonical/model.json

The canonical model is a seven-node, hand-verifiable fixture. Its bottleneck has impact 1 at tau=0 and 0 after its backup activates at tau=12:

node                     impact               shape                OR gap
bottleneck               [1, 0, 0]            fully_adaptable      [1, 0, 0]

Use JSON output for a reproducible artifact:

criticality-spectrometer run examples/canonical/model.json --format json > result.json

JSON reports identify the instrument and schema versions, model SHA-256, horizons, and run configuration. The contract is schema/result.schema.json.

Prefer the Python API?

from criticality_spectrometer import load_model, run_sweep

model = load_model("examples/tutorial/model.json")
result = run_sweep(model)
print(result.curves["runner_primary"].impact)
# [1, 0]

The narrated model-your-first-system tutorial builds that ten-node CI pipeline from scratch.

CLI exit codes are stable: 0 means success, 2 means invalid CLI input or an invalid model, and 3 means the model failed the positive, constant-baseline requirement. criticality-spectrometer --version prints the installed version.

Explain a curve

run tells you what a node's curve is; explain shows why — per horizon: lost and restored sinks, casualties grouped by cascade round, every unsatisfied requirement group, and which substitutes are active versus actually rescuing a group. Rounds are propagation stages, not unique-causality claims.

criticality-spectrometer explain examples/canonical/model.json bottleneck

tau=0   impact 1   lost sinks: sink
  round 1: assembler — unsatisfied: stage_bottleneck (members: bottleneck; pending: backup@12)
  round 2: sink — unsatisfied: final (members: assembler)
  ...
tau=12   impact 0   lost sinks: none   restored: sink
  rescuing substitutes: backup -> assembler.stage_bottleneck @12

--format json emits a self-identifying document conforming to schema/explain.schema.json.

What the instrument returns

For every node, the sweep reports an impact curve and a conservative shape class:

These labels describe model output. They are not policy recommendations or empirical claims by themselves.

Model contract

A model contains:

• nodes;
• identified requirement groups with AND or OR logic;
• substitutes targeted to a specific requirement group and activation time;
• one mission outcome: served_sinks or ordered_served_sinks.

The machine-readable contract is in schema/model.schema.json. The formal cascade, outcome, baseline, and comparison semantics are in docs/method.md.

Worked example: AI compute supply chain

The repository includes a 52-node worked example expressed entirely as model data; the engine contains no semiconductor-specific entities. Three missions separate topology, an advanced-fab path, and the primary ordered frontier stack.

The primary stack reproduces the prior case study's seven named acceptance tests at the shape level. For example, the modeled EUV corridor is persistent, TSMC is fully adaptable over the specified horizons, and germanium has no impact on that mission. The example is an application, not cross-domain validation.

• Parity analysis and divergences
• Evidence ledger
• Frozen source manifest
• Non-claims and empirical boundaries

Rebuild and verify the example:

python examples/ai_compute/build_ai_case.py
pytest -q tests/test_ai_case.py

Worked example: Kubernetes Bookinfo

The second domain compares endpoint redundancy declared by an Istio Bookinfo Kubernetes Service with the route observed for one bounded request. In the selector view, reviews_v1 is non-critical because three versions are eligible; in the observed all-v1 route view, its curve is [1, 0] because a modeled route change to the already-deployed v2 restores the mission after five minutes.

The example preserves pinned manifest/source inventory and the normalized request record separately, labels the five-minute adaptation time as an assumption, and uses explain to expose four failure rounds and the restoring substitute. Those rounds preserve the declared productpage Service layer.

• Bookinfo result and non-claims
• Evidence ledger
• Declared selector model
• Observed route model

Rebuild and verify it:

python examples/kubernetes/build_bookinfo_case.py
pytest -q tests/test_kubernetes_case.py

How it differs from common network measures

Scope

This is an alpha research instrument with a canonical fixture and two bounded worked domains. It does not infer dependencies, estimate activation times, prove causal claims, or turn shape classes into policy prescriptions. The current model contract also lacks connectivity-only edges; the AI example logs where that boundary matters. The Kubernetes example uses one externally reported request and is an external-validity probe, not broad validation.

See docs/nonclaims.md for the full boundary and CHANGELOG.md for release history.

Development

python -m pip install -e ".[test]"
pytest -q

Contributions are welcome, especially independent examples that exercise the frozen model contract without adding domain logic to the engine. See CONTRIBUTING.md.

Citation and license

Citation metadata is provided in CITATION.cff. Released under the MIT License.