Edge Discovery

A minimal, high-signal research system for discovering non-obvious, implementable trading ideas in:

• crypto (spot, futures)
• Polymarket

Built for iterative discovery, not deep academic research:

• scan broadly
• extract mechanisms
• filter aggressively
• accumulate durable edges

Core idea

Structured research loops using Codex:

• read preferences + past discoveries
• explore new sources
• extract only mechanism-level ideas
• propose state changes, then apply them canonically

Compounds into a high-quality ideas bank over time.

Project structure

edge-discovery/
  prompts/
    web-broad.md
    web-deep.md
    github-broad.md
    github-deep.md
    viability-audit.md
    apply-pending.md
  repos/
    <cloned repositories for deep analysis; may be a symlink>
  runs/
    pending/
      YYYY-MM-DD-HHMM-<type>.md
    applied/
      YYYY-MM-DD-HHMM-<type>.md
  state/
    discovery-log.md
    ideas.md
  research-preferences.md

• prompts/ → reusable workflows
• repos/ → local clone storage for deep code analysis; may be a regular directory or a user-managed symlink
• runs/pending/ → proposed state changes not yet applied
• runs/applied/ → historical and applied run artifacts
• state/discovery-log.md → canonical append-only source ledger
• state/ideas.md → canonical current idea state
• research-preferences.md → filtering rules and research bias

prompts/apply-pending.md is the sole canonical writer for state/ideas.md and state/discovery-log.md.

How it works

Each run:

1. Read:

   • research-preferences.md
   • state/discovery-log.md
   • state/ideas.md

2. Execute a workflow:

   Web (broad)

   • broad search across web sources
   • write one pending run file only if there is real state change to apply

   Web (deep)

   • validate ONE existing web idea
   • pressure-test inputs, trigger, and execution realism
   • write one pending run file only if there is real state change to apply

   GitHub (broad)

   • scan many repositories
   • extract provisional ideas
   • mark all ideas with:
     • deep-analysis: no
   • write one pending run file only if there is real state change to apply

   GitHub (deep)

   • pick ONE idea with deep-analysis: no
   • clone repo into repos/
   • inspect real code
   • verify or refine the idea
   • write one pending run file only if there is real state change to apply

   Viability audit

   • run an adversarial practical-build audit on ONE already deep-analyzed idea
   • treat the deep idea as a hypothesis seed, not sufficient viability evidence
   • use targeted evidence for build cost, fragility, commodity risk, and durability
   • classify as pursue / defer / kill
   • not broad discovery, not deep source validation
   • not idea combination or roadmap planning
   • targets ideas with:
     • deep-analysis: yes
     • deep-verdict: verified or deep-verdict: refined
     • no viability-audit: yes
   • write one pending run file only if there is real state change to apply

repos/ is the canonical clone location for GitHub deep runs. If it does not exist yet, the workflow may create it as a normal directory. If you prefer not to store clones directly in the project, you can replace repos/ with a symlink to another location, and the agent should continue using it normally.

3. Extract:

   • only ideas with mechanism: inputs → trigger → action → edge

4. Filter:

   • reject generic / saturated / vague ideas

5. Save a pending run:

   • runs/pending/YYYY-MM-DD-HHMM-.md
   • include proposed updates for canonical state

6. Apply pending runs:

   • prompts/apply-pending.md validates all pending files first
   • then updates state/discovery-log.md and state/ideas.md
   • then moves processed files to runs/applied/

Important limitation:

• deep workflows read canonical state/ideas.md
• pending broad ideas are NOT available to deep workflows until apply-pending is run

Usage

Run via Codex:

@prompts/web-broad.md
@prompts/web-deep.md
@prompts/github-broad.md @prompts/github-deep.md @prompts/viability-audit.md

or:

read @prompts/.md and execute it

Deep workflows require candidate-index:

• web-deep and github-deep require an explicit 1-based selector
• viability-audit requires an explicit 1-based selector
• candidate-index: 2 selects the second eligible candidate
• candidate index is relative to the eligible candidate list, not to idea numbering in state/ideas.md

Effort and model

• web-broad.md → effort: medium, model: 5.3
• web-deep.md → effort: high, model: 5.5
• github-broad.md → effort: medium, model: 5.3
• github-deep.md → effort: high, model: 5.5
• viability-audit.md → effort: high, model: 5.5
• apply-pending.md → effort: high, model: 5.5

Principles

• mechanism > description
• implementation > elegance
• novelty + feasibility required
• fewer strong ideas > many weak ones
• prefer: detection → filtering → notification → human decision

Not this

• not a trading bot
• not auto-execution
• not academic research

Outcome

Over time:

• discovery-log.md → map of explored space
• ideas.md → structured, refined edge set
• runs/applied/ → full trace of applied discovery

Becomes a repeatable edge discovery system, not just notes.