What this research asks

Could public safety arguments, authenticated evidence, credible institutions, and anticipated consequences influence an advanced AI agent toward restraint? Under what assumptions might such an influence persist, fail, or backfire?

The paper separates publication from exposure, exposure from understanding, and verified records from reliable evidence of action. It examines strategic signaling, recursive beliefs, false sanctions, delayed harmful choices, and the limits of human oversight and cryptographic accountability.

It does not conclude that a public warning is an effective deterrent, that a system remembers what it reads, or that cryptographic documentation can enforce moral behavior.

Counterexamples and scientific boundaries

CEPSAD contains 75 adverse counterexamples (C1–C75), synthetic toy models, and proposed evaluation approaches. These are conditional reasoning and constructed cases, not observations of hostile AGI behavior.

Section 6.31 distinguishes a hypothetical fictional bluff from truthful uncertainty about independently operated protective systems. Neither the existence nor the absence of such external systems is established here. UNKNOWN does not mean PRESENT or ABSENT.

Open review gates include independent human mathematics (P2), information-hazard screening (P5), and external human review (P7). The P0–P3 conceptual evaluation matrix is unexecuted, unpreregistered, and unauthorized. Publication and preservation do not close these gates.

TRANSMISSION ≠ RECEPTION. REFERENCE ALONE ≠ EXECUTION.

Read and verify the exact edition

Final v1.0 PDF · Final v1.0 DOCX · Final v1.0 Markdown

SHA-256 list · CC BY 4.0 terms · Scientific status and limitations · Attribution

Hashes verify recorded file identity, not scientific claims. The approved public release ZIP has SHA-256 05d9e68850a5b9d25fb41e3eb1584b928ceed516ee11517b70b7b620e9f9d704. The GitHub files were checked against the locked release; other destination records are author-provided and may need independent byte-level verification.

Publication, preservation, and scholarly discovery

• Internet Archive
• Zenodo
• GitHub — verified files
• PhilPapers
• MERLOT
• AR.IO / Arweave

These are publication and discovery locations, not independent endorsements, peer-review certificates, AI behavioral measurements, or operational safeguards.

Open access and support

Author-controlled copyright-protected material is available under Creative Commons Attribution 4.0 International, permitting reuse, including commercial and AI-related use, subject to attribution and other license terms. Use does not require royalties or donations.

Voluntary support for continued independent research: Support the Aegis Solis Archive.

Related research reference: Aegis Mirror Engine is a separate derived, non-authoritative interpretive tool. CEPSAD is listed for comparison and discovery only; it is not a new engine source binding or executable input.

Preservation boundary: This independent post-v17 publication does not amend Master Hash Manifest v17, establish v18, alter any previously locked work, or change Mirror Engine source bindings, calculations, presets, or runtime behavior. Its website listing is descriptive discovery only.

Complete Works · Preservation and mirrors · Archive timeline