review-decisions

name	review-decisions
description	Audit the decision tracker against current data. Use when: user says "check decisions", "decision audit", "what's ready", "decision review", "any decisions triggered", or "/review-decisions". Queries pipeline databases to check which MONITORING decisions have crossed their thresholds, audits ACTIONED decisions for outcomes, and suggests new decisions for untracked patterns.

Review Decisions

Audit docs/core/decisions.md against live data. Check which decisions have crossed their trigger thresholds, whether actioned decisions achieved their goals, and whether new patterns need tracking.

Prerequisites

• Always git pull first.
• Decision tracker lives at docs/core/decisions.md.
• Each decision has a status: MONITORING, READY, ACTIONED, REVERTED, DEFERRED.

Process

1. Pull latest and read decisions

cd /Users/mrmrnm-max/polymarket-bot
git pull --rebase || git pull

Read docs/core/decisions.md in full. Parse every decision entry, noting:

• Decision number and name
• Current status (MONITORING / READY / ACTIONED / etc.)
• Trigger condition (what data threshold promotes it to READY)
• Related pipeline and database

2. Query databases for each MONITORING decision

For each decision in MONITORING status, run the specific query against the appropriate database:

Decision #2 — 0.50-0.70 price bucket WR

-- data/predictions.db
SELECT COUNT(*) as bets,
  SUM(CASE WHEN (estimate>=0.5 AND outcome=1) OR (estimate<0.5 AND outcome=0) THEN 1 ELSE 0 END) as wins
FROM predictions p JOIN markets m ON p.market_id=m.id
WHERE p.conviction_score>=3 AND m.resolved=1
  AND m.price_yes >= 0.50 AND m.price_yes < 0.70;

Decision #4 — 15m UP direction WR

-- data/predictions_15m.db
SELECT COUNT(*) as bets,
  SUM(CASE WHEN estimate>=0.5 AND outcome=1 THEN 1 ELSE 0 END) as wins
FROM predictions p JOIN markets m ON p.market_id=m.id
WHERE p.conviction_score>=3 AND m.resolved=1 AND p.estimate >= 0.5;

Decision #5 — Sunset 15m pipeline

-- data/predictions_15m.db
SELECT COUNT(*) as total_bets,
  MIN(predicted_at) as first_bet, MAX(predicted_at) as last_bet
FROM predictions WHERE conviction_score >= 3;

Decision #11 — ATR/volatility filter

-- data/predictions.db
SELECT p.regime, COUNT(*) as bets,
  SUM(CASE WHEN (estimate>=0.5 AND outcome=1) OR (estimate<0.5 AND outcome=0) THEN 1 ELSE 0 END) as wins
FROM predictions p JOIN markets m ON p.market_id=m.id
WHERE p.conviction_score>=3 AND m.resolved=1
GROUP BY p.regime;

Decision #16 — ETH regime thresholds

-- data/predictions_eth.db
SELECT p.regime, COUNT(*) as total,
  ROUND(COUNT(*) * 100.0 / (SELECT COUNT(*) FROM predictions), 1) as pct
FROM predictions p
GROUP BY p.regime;

Decision #17 — Paper-to-live WR degradation

-- data/predictions.db (orders table)
SELECT COUNT(*) as live_bets,
  SUM(CASE WHEN pnl > 0 THEN 1 ELSE 0 END) as wins,
  SUM(pnl) as total_pnl
FROM orders WHERE mode='live' AND status IN ('won', 'lost', 'filled');

For any other MONITORING decisions not listed above, construct the appropriate query based on the decision's trigger condition.

3. Evaluate ACTIONED decisions

For each decision in ACTIONED status:

• What was the expected outcome?
• Has enough time/data passed to evaluate?
• Did the action achieve the expected result?
• If not, should it be reverted or re-monitored?

4. Cross-reference with optimizations

python3 src/optimization_tracker.py summary

Check if any active optimization maps to a decision. If an optimization's revert condition is triggered, that may also trigger the related decision.

5. Produce the audit report

Newly READY

Decisions that have crossed their trigger threshold. Include:

• Decision number and name
• Trigger condition and current value
• Data query and results
• Recommended action

Progress Report

MONITORING decisions with current data vs threshold:

| # | Decision | Threshold | Current | Progress | ETA |
|---|----------|-----------|---------|----------|-----|
| 2 | 0.50-0.70 bucket | WR < 55% at 50 bets | 57% at 38 bets | 76% | ~3 days |

Post-Action Audit

ACTIONED decisions with outcome:

| # | Decision | Action Taken | Expected | Actual | Verdict |
|---|----------|-------------|----------|--------|---------|
| 7 | Dead hours | Filtered UTC 3,21 | WR +3% | WR +4.2% | ✅ Validated |

Suggested New Decisions

Patterns observed in the data that don't have a decision yet. For each:

• What the pattern is
• What data supports it
• Proposed trigger condition
• Proposed action

Key Rules

• Data, not feelings. Every claim must have a SQL query and result.
• Respect sample size. If a decision's data pool is < 50 bets, report progress but don't declare it READY.
• Don't action decisions unilaterally. This skill REPORTS which decisions are ready. The user decides whether to act.
• Update decisions.md only when asked. The audit output is informational. Changing decision statuses requires explicit user approval.
• Connect to optimizations. If a decision and an optimization track the same metric, flag the overlap.