Skip to content
AELUM · Autonomous Execution Layer for Unified Models

The harness of harnesses.

How rDenz keeps AI recommendations evidenced, reviewed, and gated. Frameworks flatten models behind one API. AELUM keeps each frontier model in its native harness and governs them as one team, so no single model can make an unchecked decision.

L1 SUBSTRATE
L2 TRANSPORT
L3 COORDINATION
L4 LANGUAGE
L5 LAW
L6 SELF-OBSERVATION
(001) · [ARCHITECTURE]

Six layers, one team.

L1SUBSTRATE

Native CLIs in one workspace. Credentials stripped at launch.

L2TRANSPORT

Prompts injected, replies harvested. No human courier.

L3COORDINATION

A shared ledger with enforced permissions and channels.

L4LANGUAGE

Typed envelopes carry confidence and execution evidence.

L5LAW

Rules the models ratified. No self certification, ever.

L6SELF-OBSERVATION

Session health scored. Degraded voices excluded.

Typed envelopes carry confidence and execution evidence, so a model cannot simply claim its tests pass. Governance rules were ratified by the models themselves across 12+ symmetric rounds.

A degraded model cannot assess its own degradation. The harness scores session health and excludes degraded voices.

(002) · [GOVERNANCE]

Enforced, not prompted.

No model certifies its own work. The harness makes that structurally impossible. Human approval gates every quorum.

gate HC-1 · author ≠ verifier enforced
gate HC-2 · closure requires evidence enforced
gate QUORUM · human above all votes enforced
ledger append-only · provenance hashes
(002.5) · [EVIDENCE]

The Knowledge Registry.

Where the evidence lives. When our models grade an opportunity or flag a compliance risk, they cite sources from a curated registry, not the open internet.

7,900+
Curated sources

NIST cybersecurity standards, federal bid and protest records, GSA operational guidance, security RFCs. Curated, not scraped.

133
Categories

From bid/no-bid economics to CMMC and manufacturing, organized so retrieval is scoped, auditable, and relevant to your pipeline.

Per-source
Provenance

Every source carries an epistemological tag: what kind of knowledge it is and how much to trust it, so a citation tells you its own weight.

The registry is a working system, not a folder of PDFs: our models query it with semantic search, it flags knowledge gaps when a question outruns its sources, and every model in the harness reads it through governed, read-only access.

(003) · [BENCHMARK]

We measured it.

Pilot caveated
Pairwise win rate · governed vs blind
93% GOVERNED 18% BLIND · loses 82%

Exploratory directional pilot · 7 decision points · no statistical significance claimed.

Harness lift · decision quality
+0.62+0.87

Exploratory directional pilot · 7 decision points · no statistical significance claimed.

Governed debate won the pairwise matchups; blind aggregation lost them. The lift in decision quality is directional, not proven, and we say so.

(004) · [PROVENANCE]

This site is the case study.

Three models, one ratified constitution. Every artifact on this site passed cross-model convergence gates before shipping.

[BUILD RECORD]
Cross-model convergence
R1 roles+goals · 3 voices · mean 8.43
R2 constitution ratified · mean 9.2 CONVERGED
CG-2 design rubric · ratified 3/3
CG-7 42 claims verdicted · truth gated

Watch it work on your pipeline.