Methodology
How Predikt pressure-tests an idea, what goes into the score, and where the result stops being reliable.
The adversarial committee
An idea is never handed to a single model for a single opinion. Separate roles argue it out, and the report records where they disagreed rather than smoothing it over.
- SSkepticArgues the case against. Looks for the incumbent that already owns the market, the fatal flaw, and the assumption the pitch depends on.
- AAdvocateArgues the case for. Identifies the strongest version of the idea and the traction signals that would justify building it.
- RResearcherGathers competitors and public evidence. Reports link its claims to sources; anything it could not source is marked as model inference instead.
- JJudge & VerifierThe Judge weighs both sides into a verdict. A separate Verifier pass then checks the report against the evidence it cites before it is accepted.
Every report lists the exact model used for each role under Model Route, so you can see what produced the verdict.
What the score is made of
The 0–100 Predikt Score is a weighted assessment of four dimensions, each scored independently and shown separately in the report:
Who the buyer is, whether the demand is evidenced, and which competitors already serve it.
Whether it can actually be built as described, and whether anything about the build is defensible.
Team capability, scope realism, and whether the plan sequences into something shippable.
For AI products: model quality, data moat, and acquisition. A thin wrapper over a public API scores low here.
Above 75 reads as strong potential, 50–74 as a watchlist with real risks, and below 50 as a pass. A separate risk level is reported alongside the score — a high score with high risk is a different situation from a high score with low risk.
Checks applied after the models run
Model output is not accepted as-is. Code validates it before a report is issued:
- →Output structure and score ranges are validated; a malformed report is rejected rather than shown
- →Evidence coverage is measured and reported, so you can see how much of the verdict rests on linked sources
- →Committee disagreement is quantified as debate tension instead of being averaged away
- →The Verifier pass must return a pass before the report is released
Limitations & disclaimers
- !Not professional advice — Predikt is a decision-support tool. It is not financial, legal, or investment advice.
- !LLM variability — Underlying models may produce different outputs for similar inputs. Treat scores as directional, not absolute.
- !Public data only — Competitor and market research draws on publicly available sources, which can be incomplete, out of date, or wrong.
- !No backtesting — We do not yet have historical data correlating scores with real outcomes. This is an early-stage product.
How to use this
Predikt is designed to help you think critically about your idea — not to make decisions for you. Use the evaluation as a starting point for due diligence, not as the final word. High scores don't guarantee success, and low scores don't mean an idea is worthless. The real value is in the structured analysis and the questions it raises.