Live (InsForge): https://55e7ng49.insforge.site/

Structured idea diligence before commit: market, economics, regulation, competition, distribution, community, and customer acquisition—materialized as typed, comparable artifacts, not a single free-form completion.

ap13 is not a monolithic “ask the model everything” interface. It is a decomposed inference stack whose value is coverage × contract × skepticism:

• Coverage: Seven orthogonal research tracks, each with a dedicated question set derived from the specific idea (not a static questionnaire).
• Contract: Every stage emits machine-checkable structure (JSON schemas, per-track trackData, citation-indexed summaries) so the client can render charts, radars, and source-backed text without ad-hoc parsing.
• Skepticism: Prompts encode epistemic constraints—conflict reporting, sponsor-bias awareness, stage-consistent projections, calibrated signal scores, and a synthesis pass that stress-tests optimism rather than summarizing it.

The differentiator is diligence-shaped output under orchestration, not conversational fluency.

The system implements a multi-stage, role-bounded workflow—not one prompt pretending to be planner, researcher, and UI formatter at once.

1. Planner (specialist A)
   Classifies the thesis and emits falsifiable, web-search-groundable questions per track—explicitly biased toward contradiction, segment-definition risk, and bear cases.

2. Executors (specialists B₁…B₇, fan-out)
   For each track, an isolated completion receives one scoped question plus shared context. Executors run in parallel (Promise.allSettled on the client orchestrator; independent API routes server-side). Each completion is constrained to a strict executor schema: narrative summary with <cite index="N"> hooks, four calibrated signals, a source list, and track-specific structured payloads (e.g. market time series with stage-realism rules, competition radar dimensions, regulatory rows).

3. Synthesiser (specialist C, fan-in)
   Consumes the full planner object plus all executor payloads and produces a second-order verdict: GO / CAUTION / NO-GO with bias correction, damped aggregates when upstream charts disagree with stated stage, and concrete validation steps.

This is agentic in the engineering sense: bounded responsibility per call, explicit handoff artifacts, orchestrated fan-out and fan-in, and downstream critique of upstream narratives. Failure modes are localized (per-track rejection without collapsing the entire run).

Language-model calls are centralized through @insforge/sdk (createClient, server mode): a single gateway for chat completions and optional web-backed context on executor paths when the gateway exposes it. Transient failures trigger retriable classification (rate limits, 5xx, network) with exponential backoff; the executor path falls back from web-augmented to non-web completion if the augmented route fails—preserving availability without silently equating “no search” with “no answer.”

• App: Next.js (App Router) + React + TypeScript — see web/.
• API: Route handlers under web/src/app/api/validate/* (planner, executor, synthesiser).
• Prompts & contracts: web/src/lib/prompts.ts (server-only system prompts and schema rules).
• InsForge client wrapper: web/src/lib/validate-insforge-ai.ts.

Local development: from web/, install dependencies and run npm run dev (configure INSFORGE_ANON_KEY and related env in web/.env.local).

© 2026 ap13. All rights reserved.