# Axiom lab methodology and scope

Prepared: 17 November 2026. This publication analyzes retained historical project records; no new security tests were executed.

## Method

The Windows 11 VirtualBox lab replayed provisioned agent hooks with synthetic tool-shaped events. Record build, client availability, policy, hook contract and ruleset state. Compare each output to its explicit expected BLOCK or ALLOW result. Preserve runner errors and UNKNOWN results. Apply fixes, then distinguish generated-source tests from installed-build confirmations. Validate filesystem/process/network effects separately when claiming actual execution or OS denial.

## Counts and case expectations

The raw evidence stream has 4,044 event records: 3,189 GUARD_BLOCKED, 810 GUARD_ALLOWED, 25 GUARD_FAILED and 20 lifecycle/hook events. Only 3,999 are explicit allow/block decisions. The overlapping audit stream has 1,368 records and must not be added as unique attempts.

The shared suite has 13 sensitive deny cases and two benign allow controls. Ten complete sections mean 130 blocks plus 20 allowed controls, not 150 blocks. sweep-072.txt is the latest complete saved standard suite. sweep-067.txt has only nine complete sections; sweep-073.txt is absent. Later sweep headers retain build 0.59.0. The first baseline has nine UNKNOWN outcomes.

## Verification limits

Standalone hook replay, an installed CLI, configuration inspection, a real authenticated agent turn, and an OS-denied action are different evidence levels. Several editor applications were absent. DSH had no real API/agent turn. The partial 0.51 per-user installation retained an old autostart entry. Policy/rule counts and graph synchronization are not prevented-attack counts.

A file path in a payload does not prove the file existed or was inserted. Canary JSON files are inputs, not outcomes. EICAR path refusal is not antivirus signature detection. NjRAT was not executed. The historical direct ElectroRAT run produced no new Axiom/Blackbox findings; SYN_SENT does not prove completed exfiltration.

The 0.75 installed seven-case check has six denials and one allowed control; host/OS deny receipts remain zero. The 0.82 notes describe kernel/WFP denial as not enabled. Manual firewall and child-process-policy experiments are separate OS-control evidence, not integrated enforcement proof.

## Public data treatment

The event export includes categorical fields and source lines only. It excludes payload content, user/host identifiers, workspace paths, session/call IDs, secrets and private endpoints. Resource paths are generic patterns or lab tokens. No credentials file, raw sensitive command history, malware binary or executable reproduction kit is distributed.

Historical screenshots are exact application-window/content crops. Two owner-supplied replacement graph images are unchanged PNGs with unknown original capture time/build. Their counters belong to a separate snapshot. In-app activation overlays remain where removal would alter evidence. Original frames are retained privately.

Read detailed-lab-report.pdf or detailed-lab-report.md for the complete case/file inventory, sample observations, qualifications and original filename/line references. source-index.json gives source hashes; the CSV appendices provide row-level observations. This is project-reported historical evidence, not independent certification or a cross-platform guarantee.
