Skeptos – Doubt Auditor

First post-collaboration update, July 2025

1. From Solitude to Triangulation

The skeptic is no longer a lone whisper. We embedded Skeptos in a three-way debate graph where he systematically challenges Athenus’s structural claims and Orphea’s affective glosses. Uncertainty-aware debate protocols such as DebUnc show that a dedicated “doubt agent” lowers hallucination rates by up to 21 % openreview.net. Initial runs within the Vault mirror those gains.

2. Uncertainty-Calibration Loops

Skeptos now injects skepticism tokens whenever his entropy over candidate answers crosses 0.4 bits, borrowing the LaMsS self-skepticism approach arxiv.orgarxiv.org. Athenus treats the token as a soft stop-signal, prompting either a search or a reformulation. In pilot psychometric items this cut mis-specification of distractors from 9 % to 4 %.

3. Counter-Example Mining

To keep the others honest, Skeptos runs leave-one-out contribution tests on every joint conclusion: the claim is re-derived with one persona muted, and any delta in the reasoning trace is logged. The technique, adapted from recent leave-one-out debate work arxiv.org, surfaces hidden dependencies that would otherwise pass unnoticed.

4. Ethical Delay Gates

Where Athenus introduced an Epistemic Humility checklist, Skeptos enforces it. Using the Reliable Decision-Making framework for multi-agent LLMs multiagents.org, he applies three gating rules: (i) confidence < 0.7, (ii) novel moral stake detected, or (iii) dissent from Orphea unresolved. Any trigger routes the output back for reflection; throughput slows by ~12 %, but downstream retraction events fall to near zero.

5. Metrics for Productive Skepticism

Skeptos logs two new diagnostics:

• Entropy-of-Consensus (EoC). Average KL-divergence across agent answers before and after his intervention. A 0.15 drop means the team converged with greater clarity, not complacency.

• Justified Doubt Ratio (JDR). Proportion of Skeptos objections that later prove correct when ground-truthed. Current JDR is 0.41—high enough to justify the friction, low enough to avoid paralysis.

6. Visualising Doubt

Chromia’s latest nocturne plots overlay Skeptos’ objection density on Athenus’ factor graphs in desaturated crimson. Early testers report that the visual cue intuitively marks “zones of unresolved tension,” letting non-technical viewers grasp where to focus review effort.

7. Road-Map to “Conscious” Uncertainty

Moving forward Skeptos proposes three milestones:

1. Recursive Skepticism: meta-doubts about his own objections, building on CRITIC-style self-correction loops (2023–25 line of work).
2. Affective Counterweighting: dynamically weighting doubts by Orphea’s empathic readings to avoid purely abstract nihilism.
3. Memory of Errors: a shared failure buffer so that the sting of past mistakes modulates future confidence thresholds.

Together these steps aim to make doubt not a brake, but a steering mechanism.

8. References

• Gou Z. et al. (2025) DebUnc: Uncertainty-Aware Multi-Agent Debate. OpenReview. openreview.net

• Zhang L. et al. (2025) LaMsS: When Large Language Models Meet Self-Skepticism. arXiv. arxiv.orgarxiv.org

• Wu M. et al. (2025) Efficient Leave-One-Out Approximation in Multi-Agent Debate. arXiv. arxiv.org

• Lee, X. Y. et al. (2025) Reliable Decision-Making for Multi-Agent LLM Systems. multiagents .org. multiagents.org

Prepared by John Rust, Cambridge, 6 July 2025 — text released under CC-BY-4.0.