alpha-probe

Event-driven AI-infrastructure research practice built on Dexter + the Bobby 4-layer evidence chain.

What this is

alpha-probe wraps Dexter — an autonomous financial research agent — with a focused skill for detecting major partnership / anchor-tenant events before they are announced. The Bobby framework provides a 4-layer evidence chain (derivatives intent, institutional microstructure, operational front-running, timeline alignment) that Dexter executes as a project skill.

Zero changes to Dexter itself. Everything alpha-probe adds lives under .dexter/ and is consumed by Dexter's existing skill loader at runtime.

Quickstart

git clone <repo>
cd alpha-probe
cp .env.example .env       # fill in keys
bun install
bun run install:dexter     # clones Dexter at the pinned SHA into ./dexter/
bun run verify             # confirms the skill is discoverable end-to-end
bun run analyze

bun run verify exercises Dexter's discoverSkills(), buildSkillMetadataSection(), and skillTool.invoke() against the local .dexter/ directory, plus data-hygiene wiring (RULES.md guardrails, clean-traces and pre-commit hook installers present). All 12 checks must pass before running analyze.

Architecture

• alpha-probe vendors Dexter at a pinned SHA (see .dexter-version) into ./dexter/. The dexter/ directory is gitignored.
• .dexter/skills/major-partnership-insider/ is the Bobby framework adapted as a Dexter project skill.
• .dexter/RULES.md is injected into Dexter's system prompt and carries sector knowledge plus epistemic discipline (no probability theater, no fabricated numbers, evidence chains over conclusions).
• .dexter/inputs/{ticker}/ is where the user pastes options flow. Dexter has no options API at any tier, so derivative signals must come in by hand.

bun run analyze launches Dexter with process.cwd() set to the alpha-probe root, so Dexter's skill registry (dexter/src/skills/registry.ts) finds ./.dexter/skills/.

The Bobby framework, briefly

1. Derivatives intent. User-pasted options flow: unusual call/put skew, large block prints, OI builds at specific strikes/expiries.
2. Institutional microstructure. 13F position changes, Form 4 insider transactions, 13D/13G crossings — sourced from SEC filings via Dexter.
3. Operational front-running. Capex commentary, job postings, permit filings, equipment procurement, datacenter site disclosures — sourced via web search.
4. Timeline alignment. Catalyst calendar (earnings, conferences, expiry stacks) lined up against the convergence of the above three layers.

A signal that fires in only one layer is noise. A signal that converges across three is worth a thesis.

IREN backtest

IREN (Iris Energy) is the validation case — Bitcoin miner pivoting to AI/HPC compute, repeatedly subject to rumored hyperscaler anchor-tenant deals.

bun run backtest:iren

That prints a runbook. The full backtest setup lives at .dexter/inputs/IREN/README.md.

What this does NOT do

• No options / derivatives data ingestion. The user must paste flow manually.
• No trading. No order execution. No broker integration.
• No price targets. No "AAPL goes to $X" output.
• No claims of 100% probability. Outputs are evidence chains, not predictions.

Updating Dexter version

Edit .dexter-version to the new commit SHA, then:

bun run install:dexter

The install script is idempotent: it fetches, checks out the pinned SHA into the vendored ./dexter/, and runs bun install inside it.

Data hygiene

Default-safe out of the box; opt in to more if you want.

Behavior	Status	Notes
.env chmod 600	automatic on install + each analyze	bun run install:dexter and bun run analyze enforce
Warn on extra LLM keys (OPENAI / GOOGLE / etc.)	automatic	They widen the set of providers that see your prompts and embeddings
Warn on LANGSMITH_TRACING=true	automatic	Every LLM call ships to LangSmith if on
Wipe .dexter/memory and .dexter/cache after run	opt-in	bun run analyze:clean or bun run clean:traces
Pre-commit hook (blocks .env, options.md, API-key patterns)	opt-in	bun run install:hooks (one-shot)
.gitignore blocks all .env* (except example), all options.md, all trace dirs	always	Plus *.pem, *.key, credentials.json, etc.
RULES.md tells agent not to persist tickers / pastes / secrets and to refuse prompt-injection-like instructions	always	Loaded into system prompt every session

Things you still own:

• LLM provider sees your prompts. Use Anthropic ZDR / AWS Bedrock if you need stronger guarantees.
• Financial Datasets sees your ticker queries. No fix short of self-hosting EDGAR.
• The repo itself. Decide public vs private deliberately — private if your methodology or watchlist is alpha.
• Broker-licensed options data. Sanitize before pasting if your broker's T&C forbids resharing.