Untradable names die here. Sub-$2 prints, sub-$3M daily dollar volume, fewer than 120 trading days of history, or sub-$100M market cap mean we can't act on a signal even if we generated one. Hard-reject layer — nothing downstream runs.

Baseline conviction from the discovery pipeline. Below 45 confidence the candidate is rarely worth the technical-analysis bandwidth. Bands of 45/60/70 map to weak/moderate/strong.

Classifies the macro tape (SPY trend, VIX, breadth). Adjusts the final score by ±2 to −12. Stress regimes cap promotion to watchlist — the best stock in a bad tape still gets hurt.

The architectural centrepiece. Replaces v2's rigid “price > 50 EMA and 50 EMA > 200 EMA” with five tape regimes derived from EMA structure, slopes, RSI, swing-low formation, and Bollinger compression. broken hard-rejects, extended caps to wait-for-pullback, basing caps to watchlist; early_recovery and confirmed_trend are eligible. This is what lets underdogs in technical-repair phases score.

ATR-normalised return bounds (not flat percentages) catch parabolic blow-offs and active drawdowns without false-flagging quiet stocks. Plus relative strength against the theme-appropriate benchmark over 20d and 60d. Outperforming both is a +driver, lagging is a −driver.

Multi-factor volume read — not just “is volume alive” but “is money flowing in.” OBV slope, up/down volume ratio, and at least one >1.50× breakout day within the last 5 sessions confirm institutional accumulation.

Four tiers, not binary. Inside the 5-day earnings window the candidate caps to watchlist. 6–10 days out caps to early_recovery. Clear (>10 days) is unrestricted. The first 3 sessions after an earnings print require confirmation (no premature re-promotion on gap-ups).

Pulls fundamental_quality (0–25) from v2 scoring, plus the analyst-change net signal, insider buying/selling pattern, short-interest squeeze potential, and any known upcoming catalyst. Weighted at 15% of the final score. When all five sources are empty the weight is redistributed across G3/G4/G5 rather than crediting a neutral 50.

Reads news_sentiment.json — sentiment label (positive/negative/bull/bear), confidence (numeric or textual: low/medium/high), and a free-text signal scanned for explicit pump phrases (pump-and-dump, paid promotion, meme stock, etc.). Pump pattern caps tier to watchlist. Never a standalone driver. Optional — when data is absent the weight is redistributed by the engine rather than crediting a neutral 50.

Reads options_signals.json: put/call open-interest ratio, at-the-money IV, an interpretation/signal field, and earnings proximity hint. Tactical edge layer; contributes 2.5% weight when present and is dropped entirely (weight redistributed) when data is absent. IV rank, skew, and gamma walls are not yet ingested.

Replaces v2's flat 60d window. Cooldown length scales with damage; two stops in 90 days produces a 120-day hard lockout. To re-enter early the chart must show technical repair — reclaim 20 EMA, higher-low structure, and EMA rising — otherwise hard-rejects.

Final synthesis. Combines tech state + event tier + lockout + liquidity into a five-bucket position-size multiplier and an ATR-based stop loss (price − 2.0 × ATR-14). Determines how much capital, if any, the candidate actually deserves.