What stays verbatim
Failure evidence stays whole. When a run fails, the failing output is the input to the next decision, and a summary of a failure is a claim about the failure. Claims need checking. Context Firewall retains recognized failures and abnormal conditions with their provenance, so the decision maker sees the evidence, not a retelling.
What gets reduced
Passing output earns less room. A passing test run can return its verdict and its counts. The detail proved nothing beyond the pass, so the detail stays out of the window.
Context Firewall does this reduction deterministically. Rules, not a model, turn supported test output into a bounded evidence packet. The same input produces the same packet, and the packet can be checked against the raw output it came from.
What lives outside the window
Reduction is not deletion. The raw output stays available outside the context, and the packet carries a path back to it. Anything the packet dropped is one lookup away when a decision actually needs it.
Long-term memory gets the same split. A small shared brief holds what every task needs. Each task keeps its own working note. The full record lives outside both, reachable but not carried.
How stale memory stops compounding
What persists between tasks is narrow on purpose. Failures and decisions only. Working notes die with their task. When a summary written for one decision silently becomes the evidence for the next, errors compound and nobody can see where they entered. Provenance breaks that chain. Every packet states where its evidence came from, so a later decision can go back to the source instead of trusting the summary.
When the reducer cannot tell
Ambiguity requires escalation. If Context Firewall cannot safely interpret the output, it asks for the raw evidence rather than guessing. A reducer that guesses produces fluent summaries of things it did not understand, and those are worse than the raw output ever was.
The limits, stated plainly
The prototype supports a documented flat TAP subset and the Node spec and dot reporters. It is not an arbitrary-log parser, and it is not a production security boundary. Fewer bytes alone do not prove lower token use or better decisions.
What is verified is narrower. All 45 tests and the conformance suite pass from a fresh clone, last verified on November 9, 2026. The theory page states the principle behind the design, and the repository holds the reducer, the protocol, and the limitations in full.
