The 12 questions
Q1 vendor archetype, Q2 engagement type, Q3 budget band, Q4 who is leading the conversation, Q5 working example, Q6 pricing model, Q7 underlying AI model, Q8 data handling, Q9 ongoing costs, Q10 IP ownership, Q11 timeline, Q12 references. Wording adapts for six vendor archetypes (consultancy, agency, SaaS, freelancer, staff-aug, mixed) so the question feels native; the option codes stay constant so the engine is variant-stable.
Each option encodes a discrete declarative answer (e.g. live_demo_client, slides_only). The engine reads option codes only - we never let an LLM interpret free-form vendor answers, so reproducibility is total.
How flags fire
Each flag is a deterministic predicate over the 12 answer codes. Predicates are pure - no randomness, no LLM call, no time dependency. A flag has three properties: severity (red or yellow), the categories it touches, and a list of references back to the procurement-failure case study that justifies the rule.
Flags do not stack - each flag fires at most once. A pitch can trigger zero, one, or many flags. The full list (currently 25) is below.
How we score: the verdict cascade
After flags are evaluated, we tally red and yellow counts. The verdict is a deterministic function of those two integers:
if red >= 3 -> high_risk (walk away unless amended) if red >= 1 -> concerns (resolve before signing) if red == 0 && yellow >= 4 -> mixed (ask the recommended questions) if red == 0 && yellow <= 1 -> looks_legit (final due diligence) otherwise -> worth_proceeding (with the open questions)
The 7-category scorecard tracks per-category red/yellow tallies. A category with zero of both is marked clean. Categories never aggregate into a numeric score - they are diagnostic, not diagnostic + summative.
How we recommend clauses
Every triggered flag maps to one or more contract clauses in the 21-clause library. The mapping is many-to-many: token_cost_trap recommends monthly_cost_cap, quarterly_cost_review, and 30day_cost_notice. unclear_ip_red recommends ip_ownership_clause, derivative_works_clause, and source_code_escrow.
The PDF report ships every recommended clause with US-default sample copy. The clause library is below for quick reference; have your counsel adapt the language to your jurisdiction before signing.