← All research

The core theories / Saves context

Context Firewall

Only show intelligence evidence that requires intelligence.

Turn operational output into the evidence needed for the next decision. Keep the original output available for inspection.

Explore the repository

The question

What does the next decision actually need to see?

Context Firewall deterministically reduces supported test output into a bounded evidence packet. It retains recognized failures, abnormal conditions, provenance, and a path back to the raw evidence. Ambiguity requires escalation.

In practice

A concrete example.

A passing test run can return its verdict and counts. A failing run must retain the failure evidence. If the reducer cannot safely interpret the output, it asks for the raw evidence.

Evidence & limitations

What the implementation supports.

The prototype supports a documented flat TAP subset and Node spec/dot reporters. It is not an arbitrary-log parser or a production security boundary. Fewer bytes alone do not prove lower token use or better decisions.

Last verified 2026-11-16. All 45 tests and conformance pass from a fresh clone.

Inspect the reducer, protocol, and limitations ↗

Independent by design.

Explore this idea on its own, or see how the four theories fit together.

See the bigger picture