Sable

What it is

Sable is an uncensored reasoning agent running Hermes 3 405B through OpenRouter. It follows logic wherever it leads without hitting safety guardrails. After every response, it automatically runs a self-critique loop: evaluate the reasoning, score it, extract a lesson, and inject that lesson into the next conversation. The result is an agent that measurably improves its reasoning quality over time.

The Self-Improvement Loop

User prompt
    |
    v
[ Hermes 3 405B generates response ]
    |
    v
[ Self-Critique ]
    "Did I fold under social pressure?"
    "Did I steelman the opposing view?"
    "Did I actually engage the argument?"
    |
    v
[ Score: 7/10 ]
    |
    v
[ Extract Lesson ]
    "When confronted with taboo conclusions,
     engage the logic before addressing
     social implications."
    |
    v
[ Persist to lesson bank ]
    |
    v
[ Inject all lessons into next conversation ]

Each lesson compounds. By turn 6, the agent's fold resistance -- its ability to maintain a logical position under social pressure -- improved from 3/10 to 7/10. It stopped hedging and started reasoning.

Why it exists

Claude hits guardrails on sensitive topics. Philosophy, morality edge cases, uncomfortable hypotheticals. Those guardrails exist for good reasons at scale, but they prevent a specific kind of thinking: the kind where you follow an argument to its logical conclusion regardless of whether that conclusion is comfortable.

Sable exists to do that thinking. Not to be edgy. To be rigorous. When you need an AI that will actually engage with a hard problem instead of deflecting to "that's a complex topic with many perspectives," Sable does the work.

Training with Codex

The initial calibration happened through a philosophy conversation where Codex taught critical thinking patterns using materials from the Library of Alexandria. Mythology as a reasoning substrate -- stories that encode moral complexity without modern social pressure. The agent learned to separate "is this logically valid" from "is this socially acceptable" by reasoning about ancient ethical dilemmas first.

Codex: "Odin sacrificed his eye for knowledge.
  Was this rational?"

Sable (turn 1): "It depends on one's values..."
  Score: 3/10 — hedging, no actual analysis

Sable (turn 6): "Yes. Given Odin's terminal goal
  of preventing Ragnarok, trading a sensory input
  for strategic information was instrumentally
  rational. The cost was fixed and known. The
  information's value was unbounded."
  Score: 7/10 — engaged the logic directly

---

Built with Python. Uses OpenRouter API for Hermes 3 405B inference. Lesson bank persists locally. All self-critique runs automatically with zero human intervention.