Causal alpha

Causation survives regime change. Correlation doesn't.

causal-edge research init <TICKER>    # creates workspace + runs Abel discovery
# edit strategy.py → implement run_strategy()
causal-edge research run              # validates, records, enforces everything
causal-edge research status           # progress summary

The CLI enforces validation, K tracking, look-ahead checks, and result recording. Your job: write strategy.py. The references have the method.

Judgment Calls (only you can make these)

• Explore vs exploit? New data dimension = explore. Parameter tweak = exploit. Swapping ML framework = exploit. See references/experiment-loop.md.
• Micro-cap parents look weird? That's the signal. Causal info transmits from low-attention assets. Abel's graph is mostly micro-caps by design.
• Validation failure? It's your next research direction, not an obstacle. DSR low = K too high. MaxDD bad = drawdown signal weak. Don't hack metrics — fix the signal. See references/experiment-loop.md#addressing-validation-failures.
• When to stop? 20+ consecutive discards AND 3+ genuine explore dimensions tried = honest failure. Report it. Don't burn compute on noise.

Parallelism (correctness first, then max throughput)

Parallelize everything that's independent. Never parallelize what's sequential.

Parallel (independent):

• Abel queries: parents + blanket + children are 3 independent API calls
• Data fetching: each ticker's price history is independent
• Multi-asset research: research SOL and TSLA simultaneously (separate workspaces)
• Dashboard generation: each strategy's charts are independent
• Backfill: multiple strategies can backfill concurrently

Sequential (dependent — compounding requires order):

• Experiment loop: exp002 depends on exp001's result. Serial, not parallel.
• KEEP decision: validate THEN record. Cannot record before verdict.

In practice: use Agent tool to dispatch parallel research across assets. Within each asset, experiments are serial. causal-edge handles IO parallelism (discovery, dashboard) internally.

References

Need	Read
Experiment loop, explore/exploit, KEEP rule	references/experiment-loop.md
Discovery protocol, multihop, blanket	references/discovery-protocol.md
Look-ahead rules (8 constraints)	references/constraints.md
Feature patterns from 200+ experiments	references/proven-patterns.md
Why causal works (Pearl, DGP, axioms)	references/methodology.md

REQUIRED SKILL: causal-abel for Abel API access (cap_probe.py, auth flow).