Skip to content

Summary

Scope note (2026-09-09): this concept is the implementation view of the pipeline. The governing operating framework is the TOMIC 2.0 charter; the daily heartbeat contract lives in its operating rhythm.

Hermes is the automation loop that turns news and market state into approved, executable trades without ever letting the model improvise risk or EV arithmetic. It is a fixed sequence of stages; some are LLM-driven (reading, ideating, drafting), and others are deterministic gates that either pass or reject. Human approval sits between proposal and execution by design.

The Hermes Loop

# Stage Kind What it does Output
1 Intake Deterministic + LLM summarization Ingest news via /95-inbox/index.md conventions (dated, append-only, snapshot noted) Dated inbox concepts
2 Regime state Deterministic Classify current regime from vol/breadth/correlation features over the data stack (/80-system-design/data-infrastructure.md) Regime label + evidence (/60-regimes/index.md)
3 Ideate LLM Propose candidate strategies fitting regime + news; draft playbook structure Candidate playbooks
4 EV attest Deterministic Run the blessed EV code per the EV contract (/80-system-design/ev-contract.md); record executor receipt Attested Computation with pass/fail (/85-computations/index.md)
5 Risk check Deterministic Verify regime-aware limits: size, concentration, monthly loss, duration (see trading plan) Pass/fail against plan parameters
6 Propose LLM Assemble the trade proposal referencing playbook, EV run, regime context Proposal concept
7 Human approval Human Owner reviews proposal; no auto-execution path exists Approved / rejected
8 Execute (external adapter — outside the framework) Deterministic Submit orders per playbook execution rules; capture fill receipts. Boundary: stage 8+ is a post-approval execution/infrastructure control plane that does not belong to the TOMIC 2.0 framework (which ends at stage 7); it must be a separately governed, separately approved artifact before any order path exists Execution receipts
9 Journal Deterministic Record entry/exit data, IV at entry, P&L, days in trade into the journal 1 Journal entries
10 Outcome feedback LLM + deterministic stats Compare realized outcomes to EV projections; write results to the append-only outcome log — filed playbooks stay immutable and link to outcome records Outcome records; loop restarts

Design Rules

  • Deterministic gates are unskippable. Stages 2, 4, 5 are code, not prompts: the LLM cannot approve its own EV run or waive a risk limit — mirroring the checklist discipline TOMIC prescribes for underwriting and execution 1.
  • LLM stages are bounded. The LLM is an orchestrator and framer: it invokes the deterministic EV/risk code, it never executes gates itself, computes prices/greeks/probabilities, or waives limits (see /70-ml-ai/llm-agent-roles.md).
  • One-way valve at approval. Stage 7 is the only path from paper to money; rejections log a reason that feeds back into ideation.
  • Feedback closes the loop. Stage 10 distinguishes process failure (bad playbook) from variance (bad luck), using monthly performance statistics in the TOMIC style.

Links

References

  • Open Knowledge Format v0.2 specification, §10 (attested computations): ../OKF_SPEC_v0.2.md
  • LangGraph documentation on graph-based agent orchestration and human-in-the-loop interrupts (architectural analogue for stages 3–7): https://langchain-ai.github.io/langgraph/
  • The Option Trader's Hedge Fund (Chen & Sebastian 2012) — checklist and journal discipline.

Source Notes


  1. The Option Trader's Hedge Fund, trading-plan topic (sources[]: tomic-plan). ↩↩