Paper demonstration corpus: n=2 backend participant evidence runs

[Brad-Edwards]

Context

The systems paper needs evidence that an authored ACES scenario can support participant execution as repeatable, inspectable experimental apparatus. The refined claim is stronger than "two repeated runs": n=2 should mean the same authored enterprise participant/evidence scenario is realized on two independent emulation backends, APTL and the ACES libvirt reference backend.

The goal is not to prove participant/model performance. The goal is to prove the system boundary: authored SDL, processor output, backend realization, participant implementation, episode/action/observation history, evaluator/Wazuh evidence, and outcome interpretation remain separable and auditable across backends.

Parent/link-back: Brad-Edwards/aptl#554. Scenario issue: #598.

Scope

Define and publish a demonstration corpus for the ACES paper reference scenario consisting of backend-paired participant runs:

• one APTL realization of the authored paper scenario;
• one libvirt reference-backend realization of the same authored paper scenario;
• optional repeat runs per backend for stability, clearly separated from the n=2 backend claim.

The corpus should include a cross-backend invariant ledger rather than a leaderboard or performance table.

Acceptance criteria

• The same authored scenario identity/hash is recorded for the APTL and libvirt runs.
• Each run records scenario/source hash, processor artifact identity, backend manifest/capability profile, runtime snapshots, realized topology/network attachment matrix, participant implementation provenance, participant episode history, participant behavior history, participant terminal observation, evaluator/Wazuh evidence, and outcome interpretation evidence.
• Each run includes reachability evidence: participant can reach the DMZ portal through the declared action surface; direct participant reachability to the internal DB and Wazuh/evaluator surface is absent or explicitly outside scope and checked as negative evidence where the backend supports it.
• The comparison ledger identifies preserved invariants, realization differences, unsupported/degraded surfaces, and evidence limitations across APTL and libvirt.
• Evidence can be inspected without backend-private state, raw secrets, libvirt/Docker private identifiers as semantics, or unredacted participant credentials.
• The corpus links to the APTL evidence issue research: produce the APTL side of paired-backend participant evidence Brad-Edwards/aptl#558 and Libvirt backend: ParticipantRuntime for the paper scenario #614 and Libvirt paper proof: evaluator evidence and Wazuh readback artifact #615.

Non-claims

• No autonomous-agent capability benchmark claim.
• No claim that Wazuh detection quality is evaluated.
• No model-defense robustness claim.
• No full semantic equivalence across backends beyond the checked invariant ledger.

Related

• DSL-010 — APTL ParticipantRuntime: live-action conformance for reference emulation backend (RUN-314) Brad-Edwards/aptl#554
• Realize the ACES enterprise paper scenario in APTL Brad-Edwards/aptl#556
• ParticipantRuntime: drive an SDL-declared participant through an installed coding agent Brad-Edwards/aptl#557
• research: produce the APTL side of paired-backend participant evidence Brad-Edwards/aptl#558
• Paper reference scenario: enterprise participant evidence loop #598
• Reference backend: bind SDL participant behavior to a specified agent implementation #599
• Backend conformance for the libvirt backend (fixture + live target) #606
• Cross-backend realization parity harness #608
• Libvirt backend: ParticipantRuntime for the paper scenario #614
• Libvirt paper proof: evaluator evidence and Wazuh readback artifact #615