deep-research-perplexity

Sample successful and failed historical ClimbHill Attempts before planning new work.

Prepare a repository for controlled recursive optimization by defining goals, agent guidance, policies, evaluations, resources, and protected surfaces.

Classify proposed paths and enforce denied or approval-required ClimbHill policy before edits.

Ground implementation decisions in recorded ClimbHill resources and cited prior work.

Review policy, evaluations, cost, lineage, and risk before recording a human promotion or rejection decision.

Recombine proven parts of multiple evaluated attempts into a coherent follow-up attempt.

Produce evidence-backed run and attempt reports with evaluations, costs, risks, and promotion recommendations.

Analyze ClimbHill runs, attempts, lineage, costs, failures, and recurring constraints.

Convert evidence-backed ClimbHill experiment findings into reviewable GitHub issue proposals.

Coordinate independent, bounded coding attempts against one goal and compare them after policy and evaluation checks.

Run OpenAI deep research models through the Responses API with explicit effort, model, cost, and output controls.

Build and validate custom-corpus inputs for deep research from Google Docs, YouTube, local files, S3, arXiv, databases, Slack, email, tickets, and similar…