How SolveRight Scores Decisions — Our Methodology

The Three-Phase Scoring Pipeline

Every decision submitted to SolveRight passes through four sequential stages. The pipeline separates AI understanding from deterministic computation, ensuring scores are reproducible while still leveraging AI for data extraction and explanation.

Phase 1: AI Extraction (~3 seconds)

Four specialized extractors run in parallel using Claude Haiku: strategy, financial, risk, and technical. Each extractor translates your natural language description into structured data points with confidence scores and reasoning traces. Results are cached in Redis with a 24-hour TTL so repeat analyses and sensitivity adjustments skip this phase entirely.

Phase 1.5: Calibration (~100ms)

Five deterministic sanity checks validate every extraction before scoring: cost range plausibility, team size validation, relevance verification (do extracted factors mention the actual options?), internal contradiction detection, and data completeness checks. Failed checks reduce confidence scores or exclude the framework entirely. No LLM is involved — these are pure rule-based guards against hallucinated data.

Phase 2: Deterministic Scoring (~100ms)

Each framework applies its rubric to the calibrated extraction data and computes a 0-100 score per option. Scores are then weighted and aggregated into an overall ranking. Confidence-weighted aggregation ensures low-confidence frameworks contribute less to the final score. This phase runs entirely on the server with no AI involvement — same input always produces the same output.

Phase 3: AI Narrative (~3-5 seconds, async)

Claude Sonnet generates a persona-adapted explanation of the scores. The narrative covers an executive summary, framework-by-framework reasoning, contradiction analysis, sensitivity insights, and recommended next steps. This phase loads asynchronously — users see scores and charts within 3.5 seconds while the narrative streams in afterward. The narrative explains scores; it does not produce them.

10 Rubric Pattern Types

Not every decision framework can be scored the same way. SolveRight implements 10 distinct rubric patterns, each designed for a specific class of analytical methodology.

155 Frameworks Across 10 Categories

SolveRight's framework library spans strategic analysis, financial modeling, risk assessment, decision science, and six more categories — the most comprehensive multi-criteria decision analysis platform available.

Why Deterministic + AI Hybrid?

Pure AI scoring (asking an LLM to rate options) is fast but non-reproducible — ask the same question twice and you may get different scores. Pure deterministic scoring requires structured data that most users cannot provide. SolveRight's hybrid approach uses AI only where non-determinism is acceptable (understanding natural language, generating explanations) and deterministic rubrics where reproducibility is essential (producing scores). The result: auditable, versioned scores that users can trust, combined with the flexibility of natural language input.

Cross-Framework Contradiction Detection

When multiple frameworks are applied to the same decision, they sometimes disagree on which option is best. SolveRight automatically detects every pair of frameworks that rank options differently. When the score divergence exceeds 15 points, the system flags a contradiction with an explanation — for example, "SWOT favors Option A (stronger ecosystem) but Porter's Five Forces favors Option B (less competitive market)." These contradictions are not errors; they surface the genuine trade-offs in your decision that single-framework analysis would miss.

Sensitivity Analysis

After initial scoring, users can adjust the weight of any framework using real-time sliders. Because Phase 1 extraction data is cached, sensitivity adjustments only re-run Phase 2 — the deterministic scoring step — which completes in under 100ms. The system flags when a weight change flips the winning option, helping users understand how robust the recommendation is to changes in their priorities.