Add Fast / Standard / Deep audit modes to control cost vs coverage

[PraneelBhatia]

Problem

The product behaves like a deep audit by default. That makes the audit rich, but expensive and harder to price predictably.

Current behavior mixes:

• high persona count
• long step budgets
• post-hoc persona reflection
• screenshot annotation
• large final report generation

This should become explicit product behavior rather than the default path for every run.

Proposal

Introduce three audit modes:

Fast

• 6 personas
• 6 max steps
• no per-persona reflection
• no per-persona screenshot annotation
• Flash-tier models only
• short report

Standard

• 10–12 personas
• 8–10 max steps
• representative persona reflection only
• one overview annotation
• Flash for agent work, Pro only for final synthesis

Deep

• current behavior or close to it
• 20+ personas
• 12+ steps
• richer evidence and more screenshots
• premium mode

Why

This is the cleanest way to reduce default token burn without weakening the product.

Suggested file targets

• backend/main.py
• backend/pipeline.py
• frontend mode selector in app/page.tsx
• shared config in lib/config.ts

Acceptance criteria

• User can choose Fast / Standard / Deep before launching a run
• Backend enforces different persona counts and step caps by mode
• Final report explicitly states which mode was used
• Default mode is not effectively “deep audit”