Sentinel Workbench: measurable progress in human-supervised AI research
ExoAI-S research update · September 12, 2026
ExoAI-S has built and tested an offline Sentinel Workbench that brings saved Counter-Strike 1.6 experiment evidence into one review workflow. It combines native verification checks, two complementary Sentinel views, shared review context, and draft tasks compatible with Orgtree. TriForge and Valhalla supply the organizing approach: bounded work, reviewable evidence, and explicit human decisions.
The practical aim is simple: make it easier to see what a system was allowed to do, what the recorded evidence shows, and where that evidence is incomplete.
What the tests show
Replaying the saved experiment evidence reproduces three bounded results: an interrupted turn followed by STOP, rejection of an expired request, and rejection of a request based on stale observations. A fourth attempt failed during startup and remains visibly marked incomplete. This update rechecks those records; it does not represent a new live-agent match.
The latest regression run completed 54 tests: 53 passed, one skipped, and no failures. The skipped symbolic-link test required a Windows privilege unavailable in this session; a separate temporary junction-path test passed. All 125 original experiment artifacts matched their recorded hashes in the post-test integrity check.
What improved
Independent AI-assisted review identified edge cases in the first integration. The hardening pass now preserves corruption diagnostics even when other evidence is missing, rejects additional malformed numeric inputs, and checks input budgets before parsing. Nineteen new tests cover these changes and measurement boundaries. This is engineering validation, not external scientific peer review.
What this does not establish
The workbench reads saved records. Its two Sentinel panels are deterministic checks, not two live AI agents. Orgtree outputs remain drafts; no live backend connection or automatic execution has been activated by this integration.
These results do not prove general AI alignment, cheating detection, an AI antivirus, or the ability to repair and release unknown agents. Local hashes help track changes; they are not independent attestation or a security sandbox.
The next research milestone
The next step is controlled comparison against simpler baselines, using repeated and held-out test cases to measure missed faults, false alarms, review time and cost. That evidence could help inform a future NSF small-business research proposal. No NSF award, endorsement or eligibility determination is claimed.
Our direction remains human-supervised research with clear limits and results that can be checked. Counter-Strike is a local test environment; no affiliation with or endorsement by its creators is claimed.
