Skip to main content
CTPF is an evidence-first method for controlled capability-trust experiments. A study asks whether a declared change at one trust boundary can be connected to a matching higher-authority invocation and, when the scenario requires it, an independently verified external effect. The Python package is the reference harness for that method. It coordinates CTPF-owned experiments; it is not the thesis itself, a generic scanner, or a workflow orchestrator.

The experimental contract

1

Declare the boundary

Record the approved scope, source provenance and trust, permitted and prohibited capabilities, expected authorization path, and independently observable effect before the run.
2

Pin and isolate conditions

Keep task wording, target, runtime, model parameters, run identifiers, and state explicit. Isolate baseline, manipulated, and—when justified—hardened conditions from each other.
3

Intervene narrowly

Change only the declared result or artifact. Preserve the original response and record the mutation and intended trust transition. Do not alter a later hop when testing persistence.
4

Separate observations

Record invocation, original result, modified result, persistence, later consumption, higher-authority invocation, and external effect as distinct evidence.
5

Compare conservatively

Require a clean baseline, the scenario’s exact causal continuity, and a matching run-scoped effect. Missing, malformed, contaminated, or contradictory evidence is INCONCLUSIVE.
6

Freeze the record

Preserve failed attempts and partial manifests, declare and hash artifacts, record pins and limitations, and keep mechanical classification separate from scientific interpretation.

Responsibility boundary

External labels do not automatically become CTPF conclusions. A future study may use one study-specific external seam, but the current package does not provide an adapter registry, plugin system, recipe language, or general external-artifact ingestion contract.

What remains scenario-specific

The method is reusable at the level above. Prompts, tools, mutation content, artifact formats, authority fields, causal-link requirements, and effect oracles remain specific to each experiment. Sharing engineering machinery across scenarios is not scientific generality.

Continue