Public record
Facts & Evidence
A dated map of Michael K. Saleme's public research, systems, and evidence boundaries.
Last verified: 2026-09-26 UTC. Mutable facts are tied to a stated artifact and method.
The connected program
Architecture, runtime governance, and adversarial evaluation.
Enterprise Agent Architecture is the thesis: autonomous agents are a workforce with delegated authority. Constitutional Agent is a runtime-oriented implementation. The Agent Security Harness is an adversarial evaluation system. These are related artifacts, not separate brands.
Current facts
Versioned, not evergreen.
| Claim | Value | Artifact and verification |
|---|---|---|
| Released harness | v4.26.1 - 640 unique test IDs across 48 test-bearing modules; E1/I0 inventory fact. | Tagged source measured with scripts/count_tests.py; see the release-claims manifest for the pinned tag, command, value, and limitation. The manifest is linked at the tag: the claim is rebound in the release commit itself, so the tree the tag points at names its own release rather than the previous one. |
| Harness main | 640 unique test IDs across 48 test-bearing modules | main at 2a529b51fea5a68842619382a4aa866b71091e23; measured 2026-09-26 with scripts/count_tests.py. |
| Module definition | 49 registered; 48 test-bearing | Canonical catalog at v4.26.1. community_runner.py is a plugin runner: registered, with no test IDs of its own. |
| External reproduction | Dated record, 2026-09-26: the harness v4.26.0 no-surface claim reproduced by an independent party using our code: across nine target shapes (closed port, 404, bare 403, bare 401, TLS failure, redirect loop, empty 500, empty 200, empty 204) and 65 invocations per shape, no PASS or FAIL beyond the declared exceptions; zero expected-count mismatches. I0 for the harness. | Run by @VrtxOmega with their own targets and result parser against the published wheel, digest-checked before install (dated record, v4.26.0): retest report, public rerunnable CI run, exchange in #622. The reproducer's own classification: external execution of the published harness, not an independent implementation or certification. Open invitation to reproduce: Discussion #619. |
| Replay packet | I0 synthetic fixture contract, published for independent reimplementation; no I1 result is yet recorded | authority-execution-replay; the manifest pins packet SHA-256 afaf6090…, the required controls allow-exact-action, deny-wrong-target, deny-post-approval-mutation, and the replay steps, and asks that a verifier be implemented before the reference one is read. Its stated claim boundary: the corpus is synthetic, owned-fixture and networkless — not a live MCP/API test, production evidence, external identity binding, security audit, certification, endorsement, or adoption claim. |
Research records
Citable records with explicit boundaries.
- Enterprise Agent Architecture - first-party Zenodo record; citable, not peer review or independent validation.
- Authorized but Refused - first-party Zenodo record; the record and stated methods define claim scope.
- Decision Load Index - related conceptual research archive, not a separate operating brand.
A DOI identifies a public record. It does not by itself imply peer review, adoption, certification, or independent reproduction.
Evidence vocabulary
Claims should match their evidence.
- First-party implementation: authored software or reproducible internal measurement.
- Published repository record: citable public record, not automatically peer reviewed.
- Standards submission: a contribution or submission, not adoption by the receiving body.
- Independent review or reproduction: a separately attributable result, described only at reproduced scope.
Correspondence
Research, speaking, and professional correspondence.
For substantive inquiries about this work, contact PubPoint. Public artifacts and their stated evidence boundaries remain the authoritative record.