TOMIC 2.0 Framework — Human Review Document¶
Generated 2026-09-09; state refreshed 2026-09-17 (post V1 deployment, Hermes migration; the 9/17 "first transition state" was later shown to be a data artifact of the VIX3M bug — see the findings receipt). All framework concepts are status: draft pending this review. Your sign-off events (verified: { by: human:yasu }) are what promote them to stable.
1. What This System Is¶
A from-scratch re-imagination of the TOMIC framework (Chen & Sebastian, 2012) for the 2026 market: a daily options-underwriting business in which a bounded intelligence pipeline converts news and market data into a written market read, a gated decision engine converts that read into strategy proposals whose every number is attested, and every day's record accumulates into the evidence that amends the rules themselves.
The five standing principles (charter §Standing Principles):
- Underwriting first — premium selling is the base business; long-vol is a regime-conditioned tilt
- No attestation, no proposal — every EV runs the blessed computation and beats RV/GARCH baselines net of costs
- Propose, never execute — a human gate separates proposal from order, inviolably
- Research must resolve — every news item ends in a decision-log entry, even "no action, because X"
- Rules are versioned artifacts — the mapping and risk tables change only through the amendment process
What changed vs TOMIC-2012 (charter table): 0DTE ≈ half of SPX volume · zero commissions/penny-wide · VIX ETP flows and vol-of-vol as regime forces · GARCH/HAR as the edge bar · attested computation · LLM agents for triage/synthesis (never prediction).
2. The Five Functions (one paragraph each)¶
| # | Document | The idea |
|---|---|---|
| 1 | Intelligence | A 4-level filter hierarchy (IV-expectation → regime-state → correlation → noise), a fixed daily intake template (the anti-LLM-drift device), corroboration rules. Three outputs: bounded market read, regime snapshot, event map. LLMs triage and synthesize; they never estimate probabilities. |
| 2 | Decision | ≤3 proposals/day through four gates in order: Fit (versioned regime→strategy mapping) → EV (attested computation, edge vs RV/GARCH baselines) → Risk (never-rules) → Proposal (immutable dated draft + human approval). Declines are logged — failure is a first-class output. |
| 3 | Risk | Never-rules (heat cap, defined-risk at entry, margin-as-liquidity, fail-closed data: no snapshot → no new risk) vs posture-rules (regime-dependent delta targets; flat/no-new-risk during transitions and pre-catalyst windows). Four standing tail rules codified from 2018/2020/2024. Runs outside the decision engine — it cannot be negotiated with. |
| 4 | Learning | Three append-only machine-readable journals (regime/decision/outcome). Forced resolution (anti-hoarding), precedent-cited consistency (anti-LLM-drift), and the concordance quorum: rule amendments need ≥3 concordant journal events or a historical analogue, a written note with a falsifiable prediction, and human approval. |
| 5 | Operating rhythm | The 7-step daily cycle at three automation tiers: V0 runs today on real news with zero plumbing (human-as-sensor for the snapshot; agent web-search triage) → V1 adds one scripted snapshot (fitted HAR/GARCH) → V2 automates ingestion. Steps degrade to the tier below rather than being skipped. |
The data substrate: 96-journal schema — frozen-v0 JSONL (regime.jsonl, decisions.jsonl, outcomes.jsonl) so early history stays processable forever.
3. Proof of Work — The System Already Ran Once¶
2026-09-09 dress rehearsal (real data, V0 tier): regime snapshot computed (calm-contango-VRP+), 2 news items triaged into the inbox, 1 candidate generated and priced by the attested computation — and declined by the EV edge test (ev_net +12.73 but edge_vs_rv −5.75) with the CPI/FOMC event window independently deferring entry. Zero proposals; all declines journaled.
Two things it proved: (1) the round declined a naive premium-selling candidate — though Amendment 2026-09-17/01 later showed the decline was correct by coincidence, not by test (edge_vs_rv is algebraically vacuous and was retired; do not cite this day as evidence the old edge test works); (2) the GARCH input was a placeholder, so data-plumbing priority #1 = a fitted HAR/GARCH forecast (Tiingo access verified 2026-09-09 (verification receipt)).
4. Research Base (Sep 2026 scan)¶
Full brief: 95-inbox/2026-09-09-research-brief-recent-developments.md. Headlines:
- VRP is conditional, not gone: Chicago Fed WP 2025-17 + Dew-Becker & Giglio — option alphas ≈ 0 since ~2012. Unconditional premium selling is an unsupported prior; the framework's gating + baseline design is the response.
- HAR-RV remains the ML bar (ML did not reliably beat it); deep hedging with IV-surface state is promising; LLM-agent trading has a research→deployment gap.
- 0DTE: MM intermediation attenuates volatility (not the panic story); retail >60% of volume; GEX signals weak (excluded pending formal review); Cboe extended-hours options launch July 2026 → cross-session gap risk enters the event map.
5. Open Decisions Awaiting Your Sign-Off¶
Reviewed 2026-09-17. Each row carries a target date; this table is the input to
the Monthly rule-amendment window defined in
operating rhythm. A row with no movement by
its target date is journaled verdict: deferred with a reason — the
forced-resolution rule applies to the
owner's decision queue, not only to the inbox.
| # | Decision | Where | Status | Target |
|---|---|---|---|---|
| 1 | Promote the six framework concepts to stable — blocked on a promotion bar, not on review time. The drafts describe mechanisms that are not implemented (heat cap, max-positions-per-underlying, correlation-stress-at-1.0, the 5-day transition window, an exercised concordance quorum). Proposed bar: a concept is promotable when every mechanism it describes is either implemented or explicitly marked unimplemented inside the concept. Then promote in ones and twos. |
88-framework/*.md | all draft — bar undefined |
2026-10-01 |
| 2 | Amendment 2026-09-09/01 — VRP sign → size threshold. Blocked on specification, not on data. A review on 2026-09-17 found θ in absolute vol points is substantially a VIX-level filter (corr +0.50; pass rate 0% at VIX<13 rising to 81% at VIX>30), most permissive exactly where the mapping forbids trading (18.7% of calm days vs 63.9% of elevated). Condition 1: the HAR forecast used to diverge inside the stress windows (1008% on 2020-03-16, 2% on 2020-03-23); a validity guard landed 2026-09-17 and now fails those 20 days closed, so the calibration series is usable — but the fit's root cause (OLS on variance levels) is unchanged. Calibration strata are on the wrong variable and the effective sample is ~109 independent 45-DTE trades. | amendment note · review | proposed — re-open as rev-a | 2026-10-01 |
| 3 | ~~Update the 0DTE concept with vol-attenuation evidence~~ | 50-modern-market/0dte-complex.md | closed — done 2026-09-09; SSRN 4881008 in sources[] and body |
— |
| 4 | ~~Reject GEX as a regime variable~~ | 50-modern-market/0dte-complex.md | closed 2026-09-17 — substance in the 0DTE concept since 09-09; dismissal journaled (gex-as-regime-variable, verdict dismissed, recurs-3-times counter armed). Scope: not admitted without out-of-sample validation, reopenable via amendment |
— |
| 5 | ~~Build the V1 snapshot script~~ — delivered 2026-09-15; all 4 defects fixed 2026-09-17: VIX3M lookahead (same-date FRED closes), transition window (pending or cooling-off), re-run idempotency (observations keyed by close), and a staleness guard (inputs older than 2 business days fail closed). Tests: tools/test_regime_snapshot.py. |
tools/ · review | closed | — |
| 6 | ~~Day-1 verdicts stand~~ (rejected condor, watchlist post-CPI re-entry). The verdict holds — the event window deferred entry independently — but the recorded reason cites edge_vs_rv=-5.75, a test since retired as vacuous. Precedent-cited consistency means future candidates will cite this record, so the reason needs a corrects: entry. |
96-journal/decisions.jsonl | closed 2026-09-17 — verdict stands; corrects: record filed. Under Layer 1 this candidate would pass (ev_net +12.73), so the rejection rests on the event window alone. Cite as an event-window deferral, never as an edge-test rejection |
— |
| 7 | ~~Production runs an unattested EV engine~~ — closed 2026-09-17: references/attesters/ladder-binding.py re-executes ev_ladder.build_ladder and requires the saved ladders to match; production's own 9/17 (5/5) and 9/16-evening (8/8) ladders attest, and a coherent forgery is refused. The live prompt (owner-approved) now keeps trade inputs in hermes-pilot/ and runs the attester every round (step 2a); a failed attestation means the ladders are not evidence. First attested round: 2026-09-18. Reproducibility only — no ladder EV gate (policy question). |
EV Ladder concept | closed | — |
| 8 | Amendment 2026-09-17/01 Layers 2–4 — L1 (ev_net > 0) activated 2026-09-17 with your approval. L2 needs θ (blocked by item 2), L3 needs se_forecast exposed by the HAR tool, L4 needs the backtest harness. One build — the calibration harness — unblocks both L4 and item 2. |
amendment note | L1 active-partial; L2–4 proposed |
2026-10-01 |
| 9 | The decision log has no actor field. Human approval is asserted inside agent-written free text, so the log cannot distinguish a human approved from the agent recorded that a human approved — for the one inviolable rule in the charter. The 2026-09-17 sign-off had to be expressed as a hand-written verified: event because the schema could not carry it. Proposed: additive schema v2 field actor. |
96-journal/journal-schema.md | open | 2026-09-24 |
| 10 | ~~Ratify the position ledger (v0 schema)~~ — agent-authored 2026-09-19, already operational: 17 records including three of your determinations. Every position still lacks entry fills, so it cannot reconstruct P&L or reconcile against a broker (hole #1 partially closed). Also decide whether ledger/*.jsonl joins the repo like 96-journal. |
position ledger · 2026-09-20:concept-position-ledger |
ratified 2026-09-20 (verified: human:yasu) — hole #1 still open: no position has an entry fill |
— |
| 11 | ~~Ratify the Telegram approval flow (M1)~~ — the stage-7 human gate on Telegram. M1 is specced and has a tested dry-run harness but is not wired into any round; wiring it needs its own prompt change and record. M2 (execution era) stays blocked until the stage-8 adapter contract exists. | telegram approval flow · 2026-09-20:concept-telegram-approval-flow |
ratified 2026-09-20 (verified: human:yasu) — M1 not yet wired into any round; M2 still blocked |
— |
6. Provenance¶
Every claim traces to a source: extraction-map.md links all 58 concepts to their exact pages in the four evidence bundles (TOMIC, Natenberg, Hull, Passarelli) or to named external references. Governance charter: source.md. Change history: log.md.