Process

The engine runs as two phases — Evaluation (director, market-researcher, competitive-researcher, financial-modeler, engineer) and Build (designer, copywriter, content-QA, builder). The DAG below renders the actual dependency graph for this run; the panel beneath it shows per-agent durations, costs, and citation counts when telemetry is captured. This run shipped in degraded mode (AF telemetry unavailable in the build environment), so the panel renders the documented fallback rather than the per-agent table.

Every claim on the seven sibling tabs above traces back to a named agent step. The market-sizing TAM and SAM numbers come from the market-researcher's two-cycle pass against IBISWorld and trade-pub category reports. The competitive matrix comes from the competitive-researcher pulling public pricing pages on jobber.com, housecallpro.com, servicetitan.com, and markate.com. The unit economics, the CAC-by-churn sensitivity matrix, and the eighteen-month cashflow strip come from the financial-modeler's central case plus four-corner sensitivity runs. The architecture summary, the integration list, and the tech-risk register come from the engineer agent's structured-output pass. Each numeric claim carries a confidence tag — proven, customer-quote, modeled, sourced, or assumption — defaulting to assumption when the source agent did not emit one explicitly.

Content-QA grading on the eight investor-brief tabs cleared the 100-point rubric for this run (assessment_status complete). Visual-QA was skipped because this is a software POC with no render assets. Regression check ran post-build against the prior pitchpad ship; if any cross-ship fingerprint leaked (engine-internal markers, sibling-slug references, unsubstituted placeholders) the build would have failed the substitutions lint and never reached this page.