Ideology

A single LLM asked to evaluate your startup idea will almost always validate it. Ideology solves this with a structurally adversarial pipeline: two separate agents, two opposing mandates, running sequentially so the bear agent receives the bull's full argument and must rebut specific claims — not generic risks.

Architecture

Research Layer (sequential)
├── Problem Definer        → frames the idea, classifies business type
├── Market Size Analyst    → sizing, growth trajectory
├── Trend Analyst          → search trends, timing, tailwinds/headwinds
├── Competitor Analyst     → competitor mapping, funding activity, gaps
├── Customer Analyst       → pain level, willingness to pay signals
├── Business Model Analyst → revenue model, unit economics
└── Specialist Agents      → practice-specific (routed by business_type)

Debate Layer (sequential, N rounds with convergence check)
├── Bull Researcher        → strongest sourced case for the idea
├── Bear Researcher        → receives bull's full output, rebuts specific claims
└── Claim Register         → live structured claim graph, updated each round

Synthesis Layer (critical_risk + red_team run in parallel)
├── Research Manager       → resolves bull/bear disagreements on record
├── Critical Risk Agent    → single most likely cause of failure
├── Red Team Agent         → reasons about the reasoning quality itself
└── Verdict Agent          → score, conditions, next steps

Key Engineering Decisions

Why the debate layer runs sequentially

The bear agent's entire value comes from reading the bull's full argument. If they ran in parallel, the bear would produce generic risks — exactly what every other tool does. Sequential execution is not a limitation here. It is the architectural feature.

The Claim Register

After each debate round, a separate LLM call extracts every claim, its supporting evidence, and its status: OPEN, CONTESTED, or CONCEDED. The register feeds back into subsequent rounds so agents argue about the same claims, not past each other. The debate terminates when the register detects convergence — no new claims, no status changes — rather than after a fixed round count.

Checkpointing for resumability

Every agent writes a checkpoint immediately after completing. A 10-agent pipeline that costs real money cannot afford to restart from scratch on a network timeout. On resume, agents check if their output key is already populated and skip if so — the same code handles fresh runs and resumed runs without branching.

Two-pass report rendering

The raw pipeline output is not shown directly to users. A final LLM pass rewrites the concatenated agent outputs into an editorial-quality markdown document with strict rules: paragraphs for reasoning, bullets for enumeration, citations enforced as (Source, Year), hard word count ceiling of 700–1000 words. Agents can be verbose and raw; the final output stays clean.

How It Reduces Hallucinations

Stack