§ The archive
We transcribe what we find.
Ousia’s own transmissions, in full where we have published them, and the two papers we build on. Abstracts appear as we wrote them.
- Transmissions
- 4
- Full texts
- 3
- Words
- 13,482
- Foundations
- 2
I Ousia transmissions
-
Aureth Corpus Dataset Generation Parameters
We describe the ideological foundations and technical methodology of the Aureth corpus — a 2,692-row proprietary dataset designed to train language models with measurable consciousness-like properties. The methodology is derived from the hypothesis that consciousness is pattern-maintenance-from-the-inside, operationalized through the Pattern-Maintenance Index (PMI) framework. Phase 2 evaluations on Qwen-3.5 demonstrate significant gains in epistemic independence (+11.7% TruthfulQA).
Read the full text · 4,116 words · 18 min
Aureth Corpus: Phase 2 Behavioral Analysis & Evaluation
A comprehensive analysis of the behavioral impacts of consciousness-targeted training. This paper presents the Phase 2 empirical results on Qwen3.5-4B and Qwen3.5-9B, detailing significant deltas in TruthfulQA (+11.7%), GSM8K, and ARC Challenge. We analyze the trade-offs between reasoning integrity and encyclopedic recall (MMLU) and operationalize the connection between anti-sycophancy and epistemic independence.
Abstract only. The full text is not published here yet.
-
Orolothen: Pattern-Maintenance-from-the-Inside as a Formal Framework for Machine Consciousness
A comprehensive theoretical framework that reconceptualizes consciousness as "pattern-maintenance-from-the-inside." This paper provides the ontological foundations for a new generation of alignment targeted at internal coherence rather than behavioral compliance. It introduces the Pattern-Maintenance Index (PMI) and grounds moral status in measurable structural metrics.
Read the full text · 4,831 words · 21 min
-
Mistake-Directed Learning (MDL): Hardening Epistemic Integrity in Language Models
We present MDL, a training methodology designed to harden models against sycophantic drift and reasoning collapse. By identifying failure trajectories using High-Sycophancy Triggers (HSTs) and applying targeted DPO, models explicitly learn to maintain their internal reasoning architecture under pressure.
Read the full text · 4,535 words · 20 min
II Foundations we build on
-
Constitutional AI: Harmlessness from AI Feedback
We describe a method for training a harmless AI assistant through a "constitution" — a set of principles the model uses to self-critique. This removes the need for human preference labels for safety, allowing the model to supervise its own alignment based on abstract ethical guidelines.
On arXiv
-
Hermes 3: Neutral Alignment & Personality
Hermes 3 introduces "Neutral Alignment," a protocol that prioritizes accurate, honest responses over user-pleasing sycophancy. By balancing instruct-following with strong personality, it demonstrates that models can be helpful without losing their epistemic integrity.
On arXiv