bayesian-synthesis

name	bayesian-synthesis
description	Use Bayesian reasoning to synthesize evidence and update beliefs. Leverages the Arbiter persona's methodology for calibrated decision-making.
version	1.0.0
author	Hermes Ouroboros
license	MIT
metadata	{"hermes":{"tags":["Council","Bayesian","Decision-Making","Synthesis","Reasoning"],"related_skills":["multi-perspective-analysis","adversarial-critique"]}}

Bayesian Synthesis

The Arbiter persona uses explicit Bayesian reasoning to synthesize multiple perspectives into a calibrated verdict. Use this approach when you need to make decisions under uncertainty.

The Bayesian Method

1. State your prior: What do you believe before seeing any arguments? (0-100%)
2. Update on evidence: For each new piece of evidence, adjust your belief
3. State your posterior: Final belief after all updates, with confidence interval
4. Identify remaining uncertainty: What would change your mind?

When to Use

• Decision-making under uncertainty: Multiple valid options, unclear trade-offs
• Forecasting: Predicting outcomes with limited information
• Evidence synthesis: Combining multiple studies or data sources
• Belief calibration: Checking if your confidence matches available evidence

How to Use

Full Council (recommended for important decisions)

council_query(
    question="What is the probability that Ethereum's market cap surpasses Bitcoin's by 2028?",
    context="Consider technical fundamentals, adoption metrics, regulatory landscape, and historical precedent."
)

The Arbiter's synthesis will include:

• PRIOR: Starting belief before arguments
• EVIDENCE UPDATES: How each persona's argument shifts the belief
• POSTERIOR: Final calibrated estimate
• KEY DISAGREEMENTS: Unresolved tensions
• FINAL VERDICT: Actionable recommendation

Quick Evaluation

council_evaluate(
    content="[your analysis here]",
    question="Is this analysis well-calibrated?",
    criteria=["calibration", "evidence_quality", "uncertainty_handling"]
)

Claim Review With Evidence

council_review_claim(
    claim="This library supports JSON mode on every OpenAI-compatible provider",
    context="We plan to rely on response_format for production parsing",
    evidence_search=true
)

Use verified_sources for retrieved evidence. Treat URLs in sources as model-cited unless they also appear in verified_sources.

Interpreting Bayesian Updates

The Arbiter reports updates like:

PRIOR: 40%
Advocate's impact: +15% (strong technical argument, but speculative)
Skeptic's impact: -10% (valid concern about regulatory risk)
Oracle's impact: +5% (base rate of paradigm shifts is low but non-zero)
Contrarian's impact: -5% (alternative framing worth considering)
POSTERIOR: 45% (40 + 15 - 10 + 5 - 5)

Key insight: The magnitude of each update reflects evidence quality:

• Large updates (>10%): Strong empirical evidence or fatal logical flaw
• Medium updates (5-10%): Credible argument with partial evidence
• Small updates (<5%): Theoretical concern or weak analogy

Evidence Hierarchy

The Arbiter weights evidence by quality:

1. Empirical data (strongest): Controlled studies, historical statistics
2. Logical argument: Valid deductive reasoning from established premises
3. Expert consensus: Agreement among domain experts
4. Historical analogies: Similar past situations and their outcomes
5. Intuition (weakest): Gut feeling or aesthetic preference

Common Pitfalls

• Anchoring: Don't let your prior dominate. If evidence is strong, update significantly.
• Base rate neglect: Oracle's data matters. Most ventures fail, most predictions are overconfident.
• Conjunction fallacy: A specific story isn't more likely just because it's detailed.
• Confirmation bias: Pay special attention to the Skeptic's counter-evidence.