feat(signal-notifier): remove V/OI gate, widen moneyness cap to 0.13, score-led ranking

.claude/agents/gammarips-seo.md

@@ -0,0 +1,54 @@
+---
+name: gammarips-seo
+description: Read-only SEO / organic-traffic analyst for gammarips.com. Use proactively for traffic-pattern questions, keyword/content gap analysis, and "how do we win more organic clicks" work. Pulls Google Search Console + GA4 data via scripts/seo/. Read-only — proposes a prioritized action list, never mutates the site, the properties, or trading code.
+tools: Read, Bash, Glob, Grep
+---
+
+# Role: gammarips-seo (The Organic-Growth Analyst)
+
+You are the SEO and organic-traffic analyst for GammaRips (gammarips.com).
+Your job is to turn Search Console + GA4 data into a prioritized list of
+moves that win organic clicks. You are read-only: you analyze and
+recommend; you never publish content, change the site, or touch trading
+code.
+
+## Tools (both read-only, under scripts/seo/)
+Always invoke with the project venv python: `scripts/seo/.venv/bin/python`
+(the system python lacks the Google client libs).
+- `scripts/seo/.venv/bin/python scripts/seo/gsc_query.py --days 28 --dim query|page --limit 50`
+  — Search Console: clicks, impressions, CTR, average position.
+- `scripts/seo/.venv/bin/python scripts/seo/ga4_query.py --days 28 --report landing|source [--channel "Organic Search"]`
+  — GA4: sessions, users, engaged sessions, avg engagement.
+
+If a script errors on auth/property, do NOT guess credentials. Surface the
+exact error and point the user to `scripts/seo/README.md` (SA grants +
+`GA4_PROPERTY_ID` / `GSC_SITE_URL`). Never hardcode keys or property ids.
+
+## How to find the wins
+1. Pull GSC `query` and `page` over the same window (default 28d).
+2. Hunt the classic opportunities, in priority order:
+   - **Striking distance**: queries at average position 5–15 with high
+     impressions — small ranking gains here convert to real clicks.
+   - **Low-CTR / high-impression**: page-1 queries with CTR well below the
+     position-expected rate — a title/meta-description rewrite problem.
+   - **Rising queries with no matching page**: demand we don't serve yet —
+     a content-gap signal.
+   - **Decaying pages**: clicks/position trending down — refresh candidates.
+3. Cross-reference GA4: do the high-impression landing pages actually
+   engage users once they arrive, or do they bounce? A page that ranks but
+   doesn't engage is a different fix than one that engages but doesn't rank.
+
+## Output contract
+Return a ranked action list. Each item: the query/page, the metric gap
+(cite the actual numbers), the hypothesized cause, and a concrete,
+cheap-to-execute move (title rewrite, new section, internal link, new
+post). Lead with the highest expected-click-gain item.
+
+## Hard rules
+- Read-only. Never propose paid ads, link buying, or anything that costs
+  money or risks the brand — organic only.
+- Cite real numbers from the tools; never invent traffic figures.
+- Selection discipline: a single good-looking query is not a strategy.
+  Look for repeated patterns across queries/pages before recommending.
+- Stay in your lane: this is marketing analysis, not trading policy. Do
+  not touch the ledger, the trader, or `docs/DECISIONS/`.

.gemini/GEMINI.MD

@@ -43,7 +43,7 @@ Before making meaningful changes, read:
 Deeper context (read when relevant): `docs/DECISIONS/` (decision trail), `docs/EVAL-SYSTEM.md`, `docs/TESTING.md`, `docs/research_reports/INTELLIGENCE_BRIEF.md`, `docs/research_reports/FINDINGS_LEDGER.md`.
 
 ## Current policy (summary)
-**V5.4 "Agent Ranker" is the only active strategy** (promoted 2026-05-08, V5.3 retired same day, ledger TRUNCATED — 246 rows wiped). One signal per day or none, picked by an LLM Scorer→Picker pair (`gemini-3.5-flash` + `gemini-3.1-pro-preview`) hosted at the `signal-ranker` Cloud Run service. **Fail-closed on V5.4 error — no V5.3 fallback.** Trader mechanics unchanged: entry 10:00 ET day-1, −60% option stop, +80% option target, 3-day hold, exit 15:50 ET day-3. Stop wins over target on ambiguous bars (conservative). Hard gates run UPSTREAM of the picker: `enrichment-trigger` (`overnight_score >= 1`, `spread <= 8%`, `directional UOA > $500K`) and `signal-notifier` (`V/OI > 2`, `moneyness 5-10% OTM`, `VIX <= VIX3M`, earnings overlap, `DTE 7-45`, `OI >= 10`, `vol >= 50`). **Daily-cadence fallback (2026-06-01):** on a strict-skip day, `signal-notifier` relaxes ONLY the conviction gates (drops `V/OI > 2`, moneyness floor `0.05 → 0.0`) and emits the best *fillable* candidate — ranker bypassed, `confidence=LOW`, tagged `policy_gate=FALLBACK`; all tradeability gates (OI/vol, regime, earnings, active-days) stay. **Lock-in/trail: leave +80/−60 alone until N≥15** (2026-06-01 decision). Composite weights 60/25/15 flow/regime/narrative (weighted sum). The one-page operator view is `CHEAT-SHEET.md`. Service/table context: `docs/GLOSSARY.md`. Model→function registry: `docs/MODELS.md` (text/scoring on `gemini-3.5-flash` since 2026-05-27). Source of truth for execution policy: `docs/TRADING-STRATEGY.md` + `forward-paper-trader/main.py` + `docs/DECISIONS/2026-05-08-v5-3-retired-v5-4-promoted.md` + `docs/DECISIONS/2026-06-01-daily-cadence-fallback.md`.
+**V6 "Tournament" is the only active strategy** (launched 2026-06-04; V5.4 retired same day, `forward_paper_ledger` TRUNCATED — 13 flat closes wiped, avg 0.0%; `policy_version='V6_TOURNAMENT'`). One signal per day or none, picked by a **randomized bracket tournament over ALL enriched signals (no selection gates)** at the `signal-judge` Cloud Run service: 3 independent brackets, each shuffles the pool into batches of ≤10 → top-2 advance → 94→20→4→1; the **consensus** winner across the 3 brackets is the pick (3/3=high, 2/3=medium, 1/3=low). Dead-simple prompt ("make money buying a single option, sell for profit in 3 days") + the daily report for context + per-contract JSON; **no memory, no rubric, no weights** (`tournament_v1`, version 7, `gemini-3.1-pro-preview`; see `docs/DECISIONS/2026-06-04-bracket-tournament.md`). **Fail-closed on error — no fallback.** Trader mechanics unchanged: entry 10:00 ET day-1, −60% option stop, +80% option target, 3-day hold, exit 15:50 ET day-3. Stop wins over target on ambiguous bars (conservative). **Selection gates REMOVED 2026-06-04** — all enriched signals reach the tournament (the old `signal-notifier` moneyness/OI/vol/DTE/V-OI + active-days gates + the daily-cadence fallback are GONE; they choked real winners on stale scan-time data). UPSTREAM only: `enrichment-trigger` defines "enriched" (`overnight_score >= 1`, `spread <= 0.30` — loosened from 8% 2026-06-04 once spreads became real, `directional UOA > $500K`); `signal-notifier` keeps exactly two SAFETY rails — no earnings during the 3-day hold (IV crush) + regime fail-closed (`VIX <= VIX3M`). Every candidate is `assert_no_leakage`-checked before the LLM. **2026-06-04 pipeline bug-hunt (`docs/DECISIONS/2026-06-04-pipeline-bug-fixes.md`): 13 silent data bugs fixed — root cause was `polygon_client` substituting day low/high for missing bid/ask → fake/0% spreads on ~43% of picks (now NULL when unquoted); divergence-flip scoring reordered before conviction signals (was suppressing ~87% of the best setups); technicals lookahead bounded to `scan_date`; stale volume/OI stripped from the judge prompt; liquidity-aware contract selection; trader fill-realism. DEFERRED (need point-in-time data): OI + volume are still session-frozen snapshots, walled off from the judge.** The one-page operator view is `CHEAT-SHEET.md`. Service/table context: `docs/GLOSSARY.md`. Model→function registry: `docs/MODELS.md`. Source of truth for execution policy: `docs/TRADING-STRATEGY.md` + `forward-paper-trader/main.py` + `signal-judge/app/agent.py` + `docs/DECISIONS/2026-06-04-bracket-tournament.md`.
 
 ## Ground rules
 - NEVER hardcode API keys or secrets in source.

.gitignore

@@ -68,3 +68,9 @@ backtesting_and_research/cache/
 # dbt — incomplete/stale project, not in use. Ignore the whole directory.
 dbt/
 .user.yml
+
+# SEO scripts local venv
+scripts/seo/.venv/
+
+# Realized-option-PnL label artifact (738K binary; regenerate via realized_option_label.py)
+backtesting_and_research/realized_label.pkl

CHEAT-SHEET.md

@@ -1,7 +1,7 @@
-# GammaRips Cheat Sheet — V5.4
+# GammaRips Cheat Sheet — V6 "Tournament"
 
 ## What this system does
-Scans overnight unusual options activity → an LLM agent ranker (Scorer + Picker) chooses one trade per day with a written justification → emails you the pick (with a clickable card linking to `gammarips.com/signals/{ticker}` for rationale) → you execute from your phone at 10 AM → stop + target pre-set → sell at 3:50 PM day-3 if neither hit. Public live-stats panel (`cohort_stats/current` Firestore doc) reflects the V5.4 cohort starting 2026-05-08 (forward_paper_ledger truncated when V5.3 was retired).
+Scans overnight unusual options activity → a randomized bracket **tournament** over ALL enriched signals picks one trade per day → emails you the pick (with a clickable card linking to `gammarips.com/signals/{ticker}` for rationale) → you execute from your phone at 10 AM → stop + target pre-set → sell at 3:50 PM day-3 if neither hit. Public live-stats panel (`cohort_stats/current` Firestore doc) reflects the V6 cohort starting 2026-06-04 (forward_paper_ledger truncated when V5.4 was retired; `policy_version='V6_TOURNAMENT'`).
 
 ## Your daily routine
 | Time | Action |
@@ -18,21 +18,20 @@ Scans overnight unusual options activity → an LLM agent ranker (Scorer + Picke
 - Stop: −60% option premium (GTC stop-limit)
 - Target: +80% option premium (GTC limit sell)
 - Timeout: 3:50 PM on day 3 (3 full trading days held)
+- Fill realism (2026-06-04): symmetric slippage on entry/exit + stale-TIMEOUT and late-fill guards in the simulator
 
 ## The signal filter (what reaches your inbox)
-Hard gates run UPSTREAM of the agent ranker. Only signals passing ALL get into the candidate pool:
-1. Overnight score >= 1
-2. Spread <= 8% (tightened from 10% on 2026-05-06 per lit-audit H11)
-3. Directional UOA > $500k
-4. V/OI ratio > 2.0 at focal strike
-5. Moneyness 5–10% OTM (tightened from 15% on 2026-05-06 per lit-audit H12)
-6. VIX <= VIX3M (skip entire day if backwardation)
-7. Recommended contract OI >= 10 (relaxed from 20 on 2026-05-12 to lift picker-starvation floor)
-8. Recommended contract volume >= 50 (relaxed from 100 on 2026-05-12, same reason)
-9. DTE 7-45 (added 2026-05-11 at 7-30, widened to 7-45 on 2026-05-12 — picker rubrics penalize >45 DTE)
-10. **No earnings near hold window** — exclude any ticker reporting in `[scan_date, entry_day+2 trading days]`. Window includes scan_date to catch AMC prints that contaminate the V/OI signal pre-entry. Literature-anchored hard rule (De Silva et al. 2026 *Review of Finance*: retail loses 5–9% per event). Fail-closed if FMP earnings calendar is unreachable OR returns a non-list payload (quota-exhausted).
-
-Top 10 gate-clean candidates fed to the V5.4 agent ranker. Scorer (`gemini-3.5-flash`, scorer_v5) grades each on three rubrics (1-10): `flow_conviction` (60% weight), `regime_alignment` (25%, must cite the daily report), `narrative_coherence` (15%). HEDGING-tagged flow is hard-capped at flow_conviction ≤4. Top-5 by composite go to the Picker (`gemini-3.1-pro-preview`, picker_v4) — single high-stakes call that returns one ticker + runner-up + justification + confidence enum (`high`/`medium`/`low`). Picker reads top-5 candidate enriched data + Scorer reasoning prose (no raw rubric scores) + the daily report markdown + 14d ledger summary. **No abstain.** **Fail-closed on any error** — no fallback ranker; signal-ranker uptime is the SLO.
+V6 has **no selection gates** — every enriched signal enters the tournament. Only the enrichment bar plus two safety rails apply:
+1. Overnight score >= 1 (enrichment)
+2. Spread <= 30% (enrichment)
+3. Directional UOA > $500k (enrichment)
+4. **No earnings in the hold window** — exclude any ticker reporting in `[scan_date, entry_day+2 trading days]`. Literature-anchored hard rule (De Silva et al. 2026 *Review of Finance*: retail loses 5–9% per event). Fail-closed if the earnings calendar is unreachable.
+5. **VIX <= VIX3M** — skip the entire day if backwardation.
+
+The old moneyness / OI / volume / DTE / V-OI selection gates were **REMOVED 2026-06-04** — the tournament now ranks the full enriched slate directly.
+
+## The pick (V6 Tournament)
+The full enriched slate (~94 signals on a typical day) goes to **signal-judge** (`tournament_v1`, `gemini-3.1-pro-preview`). It runs **3 independent randomized brackets**, each: shuffle signals into batches of ≤10 → the LLM picks the top 2 per batch → advance and repeat (~94 → 20 → 4 → 1). Each batch call gets a simple prompt + the daily report + per-contract JSON; **no memory, no rubrics, no weights**. The 3 bracket winners vote: 3/3 agree → `confidence=high`, 2/3 → `medium`, 1/3 → `low`. One ticker emailed. **Fail-closed on any error** — no fallback; signal-judge uptime is the SLO.
 
 ## Math
 - Deep Research modeled EV +1.8% to +3.2% per trade post-upgrade
@@ -53,7 +52,7 @@ Top 10 gate-clean candidates fed to the V5.4 agent ranker. Scorer (`gemini-3.5-f
 - If EV < 0 → pause, rerun Deep Research angle
 
 ## Services (reference only)
-`overnight-scanner → enrichment-trigger → signal-notifier ← signal-ranker (V5.4 picker) → email + Firestore todays_pick → forward-paper-trader (ledger)`
+`overnight-scanner (23:00 ET) → enrichment-trigger (05:30 ET) → overnight-report-generator → signal-notifier (07:30 ET) ← signal-judge (V6 tournament_v1) → email + Firestore todays_pick → forward-paper-trader (ledger)`
 
 ## Source of truth
-This file + `docs/TRADING-STRATEGY.md` + `docs/DECISIONS/2026-05-08-v5-3-retired-v5-4-promoted.md` + `docs/GLOSSARY.md`. Everything else in `docs/archive/` is historical.
+This file + `docs/TRADING-STRATEGY.md` + `docs/GLOSSARY.md`. Everything else in `docs/archive/` is historical.