Territory Grounder · Field guide
Evidence you can inspect
Reviewed 2026-11-02. Governed remediation is implemented. Fresh installs start in Shadow: read-only, with no actuation targets. Three repairs demonstrated in an isolated lab; broad production effectiveness is still being evaluated.
These references distinguish implementation, recorded execution and comparative evidence. The source links below point to the public snapshot reviewed on November 2, 2026; they do not assert the current mode of any private deployment.
Implementation
- Prediction and action admission: the effect chokepoint and its tests.
- Deterministic verification: verdict logic and tests.
- Safety and default mode: the read-only default and safety controls.
- Executable specifications: requirements and their acceptance checks. Check the repository for the current test inventory; a test count alone is not an assurance result.
- Audit ledger and database migrations: hash chaining and database privileges.
Source presence is inspectable evidence of implementation. It is not a claim that every integration has passed a production trial in your environment.
Recorded lab results
The November 1 lab recording shows a disabled network link restored, a stopped service started, and a stopped virtual machine brought back. The VM case uses a deterministic liveness path. The recording also explains failed or held attempts.
Scope: a disposable test machine, selected faults and configured permissions. The film supports those bounded demonstrations; it does not establish fleet-wide reliability, a production success rate or superiority to another system. Read the walkthrough and chapter links.
The console gallery uses representative fixture data. Its screenshots illustrate the interface and are not additional live-run evidence.
Benchmark
The first campaign, reported July 31, reached 21 of 30 planned pairs. Its primary endpoint had only 17 eligible items across nine hosts. The frozen result did not meet its confirmatory population minimums. The README still summarizes that early run.
A later campaign, accepted August 26, 2026, analyzed 30 eligible pairs across 13 hosts and returned TG MATCHES (non-inferior) under its pre-registered rule. That is a specific result against one predecessor on that corpus, not a general claim of equal or superior performance.
The fault-injector endpoint favored Territory Grounder. The judged counter-comparisons significantly favored the predecessor on correct diagnosis and evidence grounding, which blocked an “EXCEEDS” verdict. The provenance also records that the predecessor recognized 31 of 32 submitted faults as synthetic, and that ground-truth assembly required manual interpretation.
Inspect the later frozen verdict, per-pair records, provenance and caveats, first campaign verdict, and pre-registration.
The remaining question is how consistently the complete prediction/action/verification loop improves outcomes across varied real incidents. The lab film and source checks do not answer that question by themselves.