Active Projects

AI Quant Pipeline (Speculative Fund)

Status: Design phase complete. All decisions locked. Ready to break into tasks. Docs:

• /home/ben/ava/research/ai-quant-pipeline-v2.md — system architecture
• /home/ben/ava/research/signal-agent-evals-v2.md — all-Haskell eval framework

Key decisions (locked):

• Standalone system in Omni.Fund.Quant — NOT integrated with Invest.hs
• All Haskell, no Python anywhere (data, math, evals)
• Speculative fund only (~5% of portfolio, ~$125K)
• Asset universe v1: 11 sector ETFs + 4 cross-asset (SPY, TLT, GLD, BTC) + VIX for signal = 16 instruments
• Linear algebra: pure Haskell for v1 (revisit hmatrix at 100+ assets)
• Market data: Twelve Data (free 800 req/day)
• Signal storage: JSONL for v1, SQLite later
• Rebalance frequency: weekly
• Eval-first: signals must pass walk-forward IC > 0.02 before affecting portfolio
• 4-phase implementation: Data → Alpha → Agents → Live trading

Next: Break Phase 1 into tasks when Ben is ready.

Earlier docs (superseded):

• /home/ben/ava/ben/research/ai-quant-pipeline.md — v1, integrated with Invest.hs
• /home/ben/ava/ben/research/signal-agent-evals.md — v1, had Python

Moral Agent Research

Location: ~/ava/ben/research/moral-agent Status: Phase 3 executing (coding agent), Phase 4 planned

Phase 2 Results (Complete)

• Phase 1a validated: SSM+FiLM hybrid beats baseline by 5%, SSM is load-bearing (~40%)
• No moral salience from pretraining (expected)
• Multi-agent restraint exists but is game-theoretic, not moral
• Transparency fix results (March 16): Fixed transparency condition (agents now share actual SSM hidden states). Results: transparency produces large but seed-dependent effects — sign flips across seeds. Opaque condition is more stable. Conclusion: it’s pure game theory + optimization dynamics, not Levinasian morality.

Phase 3 (In Progress)

Doc: docs/phase3-plan.md

• Replace reward maximization with viability-region maintenance (homeostatic training)
• Experiment A (kill switch): viability vs reward agents in same game
• Experiment B: precariousness gradient
• Experiment C: 2×2 factorial (opacity × objective)
• Coding agent is implementing now

Phase 4 (Planned)

Doc: docs/phase4-plan.md

• Isolate SSM-mediated cooperation as a standalone MARL contribution
• Experiment 4A: memory architecture baselines (SSM vs LSTM vs GRU vs MLP)
• Experiment 4B: representational probing (linear probes on SSM hidden state)
• Experiment 4C: cross-game generalization (IPD, Stag Hunt, Chicken)
• Independent of Phase 3; can run in parallel
• Target: publishable paper on SSM as implicit opponent model in MARL

Key Insights

• Moral agency may require non-optimizer architecture (homeostatic/autopoietic)
• SSM = subcortical affect system, not whole brain. Transformer = neocortex. SSM needs to be “alive”
• Robust finding: SSM-mediated cooperation (0.39-0.46 restraint) in multi-agent is real and consistent across seeds — worth publishing independently of moral thesis

Other

• Insurance review completed (March 16) — Ben is not over-insured, closer to under-insured
• PIL (PodcastItLater) — ongoing product
• Multi-LLM task spec pipeline — planning phase