Skip to content

Summary

The decision function turns the day's market read into at most three proposal drafts (a deliberately small number — the bottleneck is human review, not idea generation). Each candidate passes four gates in order; failing any gate ends the candidate with the reason recorded. Nothing here executes a trade: the output is a dated playbook draft awaiting the human gate.

Gate 1 — Fit: The Versioned Regime→Strategy Mapping

Candidates are generated from the mapping, not invented freely: the current regime label selects the eligible structure families and delta posture per regime-dependent delta exposure. The mapping is a versioned artifact (canonical band values: tools/regime_config.json — consumed by the snapshot script; amendments change the config and mapping_version together) — amended only through the learning function's amendment process (concordance quorum of ≥3 journal-cited concordant events, or one named historical case-study analogue with explicit reasoning — the analogue path supports proposals; activation additionally requires calibration where applicable and human approval), never mid-session.

Current mapping summary (v0, draft — pending accumulated validation):

Regime Delta posture Eligible families Forbidden
Calm contango, positive VRP (see amendment 2026-09-09/01 — size threshold proposed) short-delta neutral condors, calendars, diagonals naked tails, ratio backs
Elevated vol and backwardation (both required) flat-to-long delta defined-risk debit spreads, calendars w/ long-vol legs, long straddles / strangles new premium selling, 0DTE shorts
Elevated vol without backwardation (flat or contango) flat no long-vol entry; manage existing new long-vol risk
Transition (label changing) flat no new risk; manage existing everything new
Event-window (pre-catalyst) flat no new risk by default (risk function governs); exceptions are defined-risk event structures only and require an operator-recorded exception undefined-risk shorts

The long-vol archetype (relocated into the mapping 2026-09-30)

The pilot's book is long straddles, and until 2026-09-30 no row admitted them. The rules that governed it lived in prose in the round prompt, outside mapping_version and outside the amendment quorum — a prompt edit could have authorised straddle entry in a calm regime with no mapping amendment. Finding 2026-09-30:regime-mapping-straddle-gap-finding relocated them here; the canonical conditions are long_vol_archetype in regime_config.json.

Both conditions must hold. This is the finding's sharpest result, and it is why the row above now says "and" rather than a slash. Over 4,231 days (2009–2026), the share of days on which realised vol exceeded implied:

regime share median win
stressed backwardation 22.0% −0.3917
stressed contango 2.8% −0.0070
stressed flat 0 of 132 days —

High VIX with a flat or contango term structure is the worst long-vol environment in the sample — worse than anything in calm. "Elevated vol" alone would select it. The term-structure conjunction does the work, not the vol level.

Note also that this row is rarely open: backwardation of any kind is 101 of 4,231 days, 2.4%, about one day in 41.

Calm regimes are management-only for this archetype — hold to exit targets, monitor EV_PHYS drain, no entry. Transition and event windows forbid entry outright, as for every other family.

What this does not establish. These conditions gate entry, on the only measure available. The book's claimed edge is the exit path — EV_RN ≈ 0 at entry by design, with static holding measured at −21.9% of risk on SPY 765 — and vrp_realised cannot price an exit path. Whether that edge exists is untested and needs backtest Tier 2, blocked on historical option chains. Codifying the entry gate is not evidence for the edge.

Gate 2 — EV: Attested Computation or Nothing

Each candidate structure is priced and evaluated by its Attested Computation with the day's inputs. The EV contract governs (decision rules): real-world probabilities and explicit costs. A receipt that fails attestation, or inputs the computation refuses (fail-closed), ends the candidate.

The edge test is layered per Amendment 2026-09-17/01:

Layer Test Status
1 ev_net > 0 under the physical forecast, net of costs (ev_gate in the attester verdict) in force 2026-09-17
2 iv − forecast_rv ≥ θ pending θ calibration
3 edge survives the forecast's own standard error pending se_forecast
4 conditional EV beats the unconditional archetype baseline pending the backtest harness

Layer 1 is absolute: no candidate with ev_net ≤ 0 advances, in any regime, under any operator exception.

edge_vs_rv is no longer a gate — it reduces algebraically to garch_forecast < rv_window and admitted negative-EV candidates. It remains emitted as a diagnostic so the historical receipt series stays comparable.

Gate 3 — Risk: Heat, Posture, Tails

The risk function checks: portfolio heat headroom, delta-posture consistency with the regime target, forbidden-structure list, event-map windows, and the standing tail rules derived from the case studies. A candidate that fits the regime but breaches a limit is declined — limits are not tradable through.

Gate 4 — Proposal: The Draft Artifact

Surviving candidates become playbook drafts (90-playbooks) with a fixed anatomy: hypothesis (one paragraph), structure, attested EV receipt reference, regime context, risk checks, invalidation condition, and exit plan. The draft is immutable once filed; human approval moves it to tradable, or it dies as a record.

Failure Is a First-Class Output

A day with zero proposals is a valid day. Declined candidates are logged with their failing gate — declines are the evidence the learning function uses to improve the mapping.

Links

References