Skip to content

Summary

The five functions run on a fixed daily heartbeat. This concept designs the cycle and, critically, its automation tiers: the framework is honest that no market-data plumbing or fetch automation exists yet, so it defines what a fully legitimate day looks like at each tier — V0 needs nothing but a human with a browser and the agent. The heartbeat, gates, and records are identical across tiers; only the plumbing matures.

The Daily Cycle (Seven Steps)

# Step Function Output Role
1 Regime snapshot Intelligence Regime label + variables + data-missing flag Deterministic (V1+) / manual (V0)
2 News intake Intelligence Dated inbox concepts LLM triage, fixed template
3 Market read + event map Intelligence Bounded read, catalyst calendar LLM synthesis, precedent-cited
4 Candidate generation + EV Decision Candidates with attested receipts Deterministic compute, LLM framing
5 Risk gates Risk Pass/decline with reasons Deterministic checks
6 Proposal filing Decision Playbook drafts (≤3) LLM drafting
7 Journal write + review Learning JSONL records, resolutions Deterministic + human review

Steps 1–7 run in order; failed gates end candidates, not the day. A zero-proposal day is valid. Timebox: the cycle is designed to complete within one focused hour at any tier.

Automation Tiers

Tier Regime snapshot News intake EV gate Journal Prerequisites
V0 (day one) Human reads VIX/term data (or cached values), agent records Agent triages user-flagged or agent-searched news via web search Fires only when inputs exist; else proposal stays draft unpriced Manual JSONL append by agent None — this is the honest "consumes news daily" starting point
V1 One script computes the snapshot from cached/ThetaData pulls, fail-closed Agent web-search triage Scripted calls to the attested computations Scripted append One snapshot script; frozen JSONL schema
V2 Scheduled, automated Automated source polling with dedup/corroboration Live EV on live quotes Automated Fetch plumbing, scheduler, monitoring — deliberately deferred until ~2 weeks of V0/V1 reveals the real data needs

The tier rule: a step degrades to the tier below rather than being skipped. V0 is genuinely the framework — the human becomes the sensor until scripts exist, which is exactly TOMIC's original operating mode.

Fail-Closed Rules

  • Snapshot inputs missing → regime = data-missing; no new risk (a never-rule); the journal records it.
  • Computation inputs missing → candidate unpriced; it may be filed as draft-hypothesis only, never as a proposal.
  • Intake inconsistency (today's read contradicts cited precedent) → divergence is stated in the journal, and the human review note at step 7 addresses it.

Periodic Review Cadence

The daily cycle is step one; the framework itself is reviewed on three fixed cadences (all append to the journals; all outputs are evidence for the learning function):

Cadence When Scope Output
Weekly First trading day of the week, appended to the daily round Week's regime series vs decisions (did labels and outcomes concord?); amendment queue triage; ops review (fail-closed events, tool bugs, env gaps) # Weekly Review section in the round report + decision-log record
Monthly First trading day of the month Rule-amendment window: score tagged AMENDMENT-PROPOSALs against the quorum; recalibration checks (VRP series growth, model-version drift); REVIEW.md drift check Decision-log record with amendment dispositions
Quarterly Calendar quarter start Charter-level review: research-brief refresh (new papers/market structure), mapping performance vs the accumulating labeled series, tier promotion assessment (V1→V2) Dated review concept in 90-playbooks

Skipped reviews are journaled (verdict: deferred) — the cadence itself is fail-closed, like everything else.

Design Status

Nothing in this concept is implemented. Implementation milestones: (1) freeze 96-journal JSONL schemas, (2) one V1 snapshot script, (3) scripted EV calls, (4) fetch/scheduler. The automation architecture holds the pipeline implementation view; this concept holds the rhythm contract.

Links

References