Vertical Templates

Three models independently review code, then adversarially attack each other's findings. Attackers probe for security vulnerabilities, logical flaws, edge cases, and robustness issues. Defenders respond to each attack. A judge evaluates the validity of attacks and strength of defenses.

Review this authentication middleware for security vulnerabilities and suggest improvements

Models explore different perspectives on complex research topics, challenge each other's sources and interpretations, and converge on well-supported conclusions. All claims must include citations. Uncertainties are explicitly flagged rather than glossed over.

Synthesize current research on transformer architecture efficiency improvements published in 2024-2025

Five models evaluate risks with mandatory dissent reporting. Every assessment must explicitly identify both majority and minority positions. Models rate likelihood and impact for each risk, propose concrete mitigations, and map to compliance frameworks.

Assess risks of migrating our production database from PostgreSQL to a multi-region CockroachDB setup

Safety-critical deliberation with mandatory citations and dissent. Models provide differential diagnoses with evidence strength ratings. Red flags are automatically highlighted. Every claim must be backed by medical literature or clinical guidelines.

Evaluate differential diagnosis for patient presenting with acute chest pain, elevated troponin, and normal ECG

Dialectical format: one model argues the risk position (this is dangerous/non-compliant), the other argues acceptability (this is fine/compliant). Both are evaluated blindly — model identity stripped — to ensure the quality of legal reasoning matters, not the model brand. Mandatory dissent ensures both sides are fully explored.

Review this SaaS terms of service for GDPR compliance risks and recommend specific clause revisions

Quantitative financial risk analysis with mandatory dissent. Models evaluate market, credit, operational, and liquidity risks using standard metrics (VaR, CVaR, Sharpe ratio). Stress testing scenarios are required. Results mapped to regulatory frameworks (Basel III, SOX, Dodd-Frank, MiFID II).

Evaluate the risk profile of this investment portfolio under current market conditions with stress scenarios

Templates can be loaded programmatically via the template registry. Each template returns a configuration object with mode, rubric, system prompts, max rounds, and default models.

from consilium.templates import get_template, TEMPLATES

# List all templates
print(TEMPLATES.keys())
# → ["code_review", "research_synthesis", "risk_assessment",
#    "healthcare", "legal", "finance"]

# Load a template
template = get_template("code_review")
# Returns: {
#   topic: str,
#   mode: "redteam",
#   rubric: { security: 0.30, correctness: 0.25, ... },
#   system_prompts: { attacker: "...", defender: "..." },
#   max_rounds: 2,
#   default_models: 3
# }