Praxis

This document bridges analysis and action. It describes where Ghost’s knowledge meets trading practice, and how the system learns from experience.

The Decision Boundary

Ghost outputs end at a clear boundary:

Ghost Domain                    │  Human Domain
────────────────────────────────┼────────────────────────────────
Levels with confluence          │  Which level to enter at
Director alignment/conflict     │  How to weight disagreement
Zone status (IN_PRIMARY, etc)   │  Whether to act on status
R:R at each zone               │  Position size per R:R
Flags (earnings, gamma, etc)   │  What the flags mean for me

Ghost presents structured data. The trader applies context, risk tolerance, and judgment.

Session Workflow

Pre-Session

# Run full analysis on watchlist (20 LLM calls each)
ghost run VST
ghost run META
ghost run AMZN

This populates the KB with Market Aggregator output: entry zones, targets, stops, flags.

Pre-Market

# Check for setups across watchlist (zero API calls)
ghost scout scan

# If holding a position, activate defense mode
ghost run VST --entry 159.71

Scout surfaces actionable setups from KB. Defense mode shifts analysis from “where to enter” to “what to do with position.”

During Session

# Monitor for intraday signals
ghost intraday VST

# Re-scan if prices moved
ghost scout scan

Intraday analysis detects gaps, sweeps, opening range breaks, volume anomalies.

End of Day

# Verify predictions against actual outcomes
ghost verify VST

# Review guardrail accuracy
ghost guardrails

This feeds the self-correction loop. Predictions are marked hit/miss.

Execution Modes

Ghost surfaces two modes. A setup can be scalp, swing, both, or neither.

Mode Detection

See ONTOLOGY.md Execution Mode Concepts for full treatment.

Zone-Based Trading

Entry zones cascade:

primary (highest) > secondary > aggressive (lowest) > stop_loss

All three zones required. Stop must be below aggressive.

Zone Status Actions

Scout classifies current price vs zones:

R:R Calculation

R:R = (scalp_target - zone_midpoint) / (zone_width / 2)

Ghost calculates R:R at each zone. The trader decides minimum acceptable R:R.

Position Management

Defense Mode

When holding a position, pass --entry to shift analysis:

ghost run VST --entry 159.71

This activates guardrails:

Exit Optimization

Defense mode shifts from entry zones to exit guidance:

Post-Mortem Methodology

When analysis fails, trace backwards:

Outcome → Aggregator output → Director synthesis → Agent analysis → Data input

Error Categories

Correction Process

1. Identify — Which layer failed?
2. Isolate — Test that layer in isolation (ghost test [stack])
3. Correct — Adjust prompt or logic
4. Validate — Zero-cost validation (ghost prompt render)
5. Verify — Run full analysis, compare to known-good

Backtest Process

For systematic evaluation:

1. Scorecard

Track predictions systematically:

2. Audit

Group by condition to find systematic errors:

"Multi-TF squeeze" predictions: 12 total, 4 hits = 33% accuracy
"Near max pain on Friday" predictions: 8 total, 7 hits = 87% accuracy

3. Correction

If a condition shows systematic error, adjust:

• Agent prompt (hermeneutics)
• Director synthesis weighting
• Aggregator output constraints

4. Replay Validation

Re-run corrected prompts against historical data:

ghost prompt render [agent] --ticker VST

Zero-cost validation of prompt changes.

5. Live Verification

After correction, track next N live predictions for that condition. If accuracy improves, correction is valid.

Resource-Aware Practice

Ghost runs cost API quota. Practice with awareness.

Quota Rules

1. Never run ghost run more than twice per development session for testing
2. For prompt changes, use ghost test [stack] in isolation
3. For syntax validation, use ghost prompt render at zero cost
4. If quota hits 429, stop immediately — do not retry
5. Batch changes, verify together

Feedback Integration

Ghost’s feedback loop:

Run → Predict → Verify → Learn → Adjust → Run

The system doesn’t just predict — it learns from predictions.

Scout Cold-Start Behavior

Scout scoring has two effective regimes depending on prediction history:

Structural-only (cold start): When fewer than 3 verified predictions match the current condition combination, get_multi_condition_accuracy() returns accuracy_pct: None, which maps to a 1.0x multiplier — no adjustment. Scout scores using structural factors only: zone status, R:R, conviction, director bias.

Full scoring (warm): Once enough verified predictions exist, Scout applies condition-aware accuracy weighting. The system degrades gracefully — it progressively relaxes conditions (dropping lowest-priority first: day_of_week, then gap proximity, then director confidence, etc.) until it finds at least 3 matching predictions. This means partial condition matches activate before full matches do.

Bootstrapping a new ticker requires multiple ghost run calls to build prediction history, plus time for auto-verification (predictions must be >= 1hr old). This has quota cost implications — budget accordingly. The structural-only scores are still useful; condition accuracy adds refinement, not the foundation.

Related Documentation

• EPISTEMOLOGY.md — How Ghost forms and validates knowledge
• ETHICS.md — What Ghost must and must not do
• ONTOLOGY.md — What the concepts are
• DAY_DESK_REFERENCE.md — Day desk tool catalog
• PEDAGOGICAL.md — How to learn this system