Technical Analysis Ontology

This document defines the fundamental concepts and relationships in Ghost’s technical analysis system. It establishes the shared vocabulary that all agents use.

Core Concepts

1. Memory Types

Technical analysis reads three types of market memory. (Momentum is a timing layer, not a memory type — it tells you WHEN to act within the structure the other three define.)

Relationship: Structural memory creates the framework, trend memory shows movement within that framework, positioning memory reveals who is trapped or defending.

2. Levels

A level is a price where market behavior is expected to change.

Level Taxonomy:
├── Support (price expected to bounce)
│   ├── Structural (Fib retracements, prior lows)
│   ├── Dynamic (moving averages, trendlines)
│   └── Positioning (VWAP where buyers defend)
├── Resistance (price expected to reject)
│   ├── Structural (Fib extensions, prior highs)
│   ├── Dynamic (moving averages, trendlines)
│   └── Positioning (VWAP where trapped sellers exit)
└── Neutral (inflection points)
    └── Pivots, mid-points, value areas

Level Strength is determined by:

• Source count: How many agents identify it
• Timeframe count: How many timeframes show it
• Recency: How recently it was tested
• Reaction magnitude: How strongly price reacted before

3. Confluence

Confluence occurs when multiple independent sources identify the same price zone.

Confluence Strength (Fib — computed by _calculate_confluence_zones() in timeframe_fibs.py):
├── STRONG:    3+ timeframes, OR 2+ with monthly, OR 10+ fibs from 2+ timeframes
├── MODERATE:  2 timeframes, OR 5+ fibs clustered
└── WEAK:      Single timeframe only
Clustering tolerance: 0.75% — levels within this % of cluster midpoint are grouped.
Pre-calculated from ALL stored fibs, filtered to near-price zones for agent prompts.

Confluence Strength (Cross-Agent):
├── STRONG:    3+ agents agree (Fib + MA + VWAP)
├── MODERATE:  2 agents agree
└── WEAK:      Single agent only

Confluence Types:

• Horizontal: Multiple levels at same price (e.g., Fib 61.8% + 50 SMA + session VWAP)
• Temporal: Same level confirmed across multiple timeframes
• Cross-domain: Structure + Trend + Positioning agree

4. Epochs

An epoch is a meaningful price movement worth measuring.

Epoch Requirements:
├── Magnitude: Minimum move size (varies by timeframe)
│   ├── Monthly: 30%+ moves
│   ├── Weekly: 15%+ moves
│   ├── Daily: 5-15% moves
│   └── Intraday: Session ranges (no minimum)
├── Duration: Sustained movement (not noise)
└── Significance: Clear swing high/low boundaries

Epoch Anatomy:

• Anchor Low (0%): Where the move started
• Anchor High (100%): Where the move ended
• Retracement Levels: 23.6%, 38.2%, 61.8%, 78.6%
• Extension Levels: 127.2%, 161.8%, 200%, 261.8%

⚠️ THE 50% RULE (CRITICAL)

The 50% level is REFERENCE ONLY — it is NOT a structural level.

Do NOT list 50% in “key support/resistance” outputs. The 61.8% golden ratio is the true structural level. See FIB_HERMENEUTICS.md for full details.

5. Regime

A regime describes the current market state.

Regime Types:
├── Trend Regime (from MA)
│   ├── Bullish: Price > stacked MAs
│   ├── Bearish: Price < inverted MAs
│   └── Tangled: MAs interleaved (choppy)
├── Session Regime (from VWAP)
│   ├── Buyers control: Closes above VWAP
│   ├── Sellers control: Closes below VWAP
│   └── Contested: Mixed closes
└── Momentum Regime (from Momentum)
    ├── Strong bullish: MACD > 0, histogram expanding
    ├── Strong bearish: MACD < 0, histogram expanding
    └── Transitional: Crossover imminent

6. Positioning

Positioning maps where market participants are located relative to price.

Position States:
├── Trapped (underwater)
│   ├── Location: AVWAP above current price
│   ├── Behavior: Will sell rallies to escape
│   └── Creates: Supply/resistance
├── Defending (profitable)
│   ├── Location: AVWAP below current price
│   ├── Behavior: Will buy dips to add
│   └── Creates: Demand/support
└── Neutral
    ├── Location: AVWAP near current price
    └── Behavior: Unpredictable

7. Signals

A signal is an actionable indication from an indicator.

Signal Taxonomy:
├── Momentum Signals
│   ├── Divergence (price vs indicator disagreement)
│   ├── Crossover (indicator lines crossing)
│   ├── Extreme (overbought/oversold)
│   └── Squeeze (volatility compression)
├── Structural Signals
│   ├── Level test (price at key level)
│   ├── Breakout (level violated)
│   └── Rejection (level defended)
└── Timing Signals
    ├── Confirmation (multiple signals align)
    ├── Warning (early reversal sign)
    └── Invalidation (thesis broken)

8. Pattern Memory (Playbooks)

A playbook is a Bayesian model of a recurring market pattern, calibrated from historical instances. Where Fib, MA, and VWAP agents read real-time market memory (structural, trend, positioning), playbooks read compounded historical memory — what has happened on prior sessions that match today’s setup, and what that implies for today’s outcome probability.

What a Playbook Is

A playbook is a versioned YAML artifact per ticker module that encodes:

Loaded playbooks produce two numbers at runtime: posterior firing probability (how likely this pattern is active today, computed via Bayes from the prior and the observed signals) and posterior recovery rate (given the pattern is active, what’s the historical rate of the desired outcome, computed via Beta-Binomial shrinkage).

Signals: Mechanical vs Narrative

Playbook signals divide into two types, each with different handling in the Bayesian inference layer:

The schema enforces this split. Mechanical signals MUST have both likelihoods set. Narrative signals MUST NOT have likelihoods. The inference layer skips narrative signals by signal_type check, not by absence from the conditions dict — they’re passed through to the prompt as context for the LLM consumer.

This separation preserves epistemic honesty. Bayesian math is for computable, likelihood-calibrated signals. Qualitative judgment (e.g., “is this a real catalyst?”) belongs in the LLM’s interpretive layer, not in a likelihood ratio pulled out of thin air.

The Tri-State Model

Each historical instance records per-signal values as a tri-state:

The critical distinction is between absent and unknown. Marking a signal absent when we can’t actually observe it would systematically bias posteriors downward — the playbook would be penalized for our inability to see the signal. The unknown state preserves the option: “we don’t know, don’t count it either way.”

This matters most when historical data is uneven. The early VST stop mine instances (Jan-Mar 2026) have options/flow signals marked unknown because GCS reports only go back to April. The Apr 2 and Apr 9 instances have them present because the reports exist. The likelihood estimation counts known instances (present + absent) in the denominator, not total. This is an epistemic choice, not a convenience — it’s the honest treatment of missing data.

Calibration Regimes

Playbooks fall into one of three calibration regimes, determined by what data is available and how the negative pool is defined:

The regime determines the shape of discrimination and the appropriate expectations for the LLM consumer. A Regime 1 playbook firing at 70% posterior is a strong directional signal. A Regime 2 playbook firing at 70% posterior is a weaker context flag. The LLM should reason differently about them.

Definition Criteria vs Signals — The Critical Distinction

Every playbook has a definition_criteria field separate from its signals list. This distinction is easy to conflate and produces degenerate playbooks if confused.

Definition criteria are the rules used to IDENTIFY historical instances of the pattern. They are not predictive signals. By construction, every positive instance satisfies the definition criteria with probability 1. Including them as signals would produce likelihood ratios of 1.0 / likelihood_if_no_fire — tautological, overconfident, and uninformative.

Signals are observable features that VARY across instances and discriminate positive vs negative pools. They are the things the Bayesian layer actually reasons over.

Example — pre_mine_setup:

• Definition criteria: rally_from_open >= 2%, range >= 2.7%, close_position <= 25%. These are how we identify “this is a stop mine session.” Every positive instance has them. They are NOT signals.
• Signals: prior_day_keltner_squeeze, prior_day_macd_state, ceg_close_magnitude_match, etc. These vary across instances and give the Bayesian layer real discriminative power.

The recovery_t_plus_2 review caught a tautology where a signal was equivalent to the outcome metric by construction — same failure mode, different expression. Always verify empirically that signals have non-degenerate likelihood ratios before shipping a playbook.

Sister Playbooks and Directional Symmetry

Many market patterns have directional mirrors that the initial framing misses. A stop mine runs both ways — long-side (rally→fade, trap longs) and short-side (flush→squeeze, trap shorts) are mechanically symmetric but structurally distinct (close in opposite quartiles). If a playbook is built for one direction, the mirror should be scanned explicitly:

Sister playbooks have disjoint populations (no session can be both by criterion) but share the same mechanical principle. They are documented as a matched pair in the library and may be consolidated into a single symmetric playbook in a future refactor.

Cross-Playbook Coupling

Playbooks are not always independent. Some are derived views of others — they operate on the same underlying population with different evaluation timing or different outcome framing. Their correlation is structurally 1.0, not 0.78.

Example: recovery_t_plus_2 is a derived view of pre_mine_setup. Every instance of recovery_t_plus_2 is BY DEFINITION an instance of pre_mine_setup (the 9 mine dates are the same). They evaluate at different T_offsets (T+0 for setup, T+2 for recovery) and predict different outcomes (the mine pattern vs the recovery pattern), but they are coupled through their shared population.

The LLM consumer must treat coupled playbooks as a pair, not as independent signals. The correlations field in each playbook’s YAML captures this explicitly — a value of 1.0 means “this playbook is structurally derived from the referenced one; do not double-count their posteriors.”

The Research Backlog Lifecycle

Not every observed pattern becomes a playbook. Patterns that fail empirical verification enter a Research Backlog — a holding area for candidates that require more data or a different approach.

Observed pattern
  │
  ▼
Empirical scan on YTD data
  │
  ├─ 0-2 instances ────────► Research Backlog (defer, grow via live capture)
  │
  ├─ 3-4 instances ────────► Regime 2 consideration OR Research Backlog
  │
  └─ 5+ instances ─────────► Build queue (graduates to the library)

The Research Backlog entries are documented with: the pattern description, the seed instances (if any), the specific obstacle to formalization, and a clear maturity path. Sub-task F’s online feedback loop captures live instances over time. When a backlog entry accumulates enough data, it graduates to a real playbook via a future build session.

The backlog is a first-class knowledge artifact, not a todo list. It represents empirically-validated observations that the current data can’t yet support as formal playbooks — valuable knowledge held in escrow.

Scale Invariance — playbooks operate at multiple timeframes

Pattern memory is scale-invariant. The same accumulation pattern (bait → distribution → structural support → catalyst exit) operates at multiple timeframes — intraday, single-session, multi-day, multi-week, multi-month — with scale-specific criterion thresholds. This is exactly parallel to Fib retracements operating at multiple epochs (monthly / weekly / daily / intraday) with epoch-specific magnitude thresholds.

The Phase 2 playbooks (pre_mine_setup, recovery_t_plus_2, mirror_session_pair, short_mine_setup, narrative_cover_rotation) are valid pattern detectors at the single-session scale. They are not wrong — they are scale-specific instances of a broader canonical pattern.

The canonical playbook (Phase 2.5: accumulation_stop_harvest, authored AFTER Phase 2 push) is a meta-playbook that:

• Sits on top of the existing single-session playbooks as a scale classifier + execution router
• Takes day-1 signals (intraday tape shape, prior-day characteristics, sector divergence, distance from multi-month log trendline, known catalyst calendar within 48-72h)
• Outputs a scale classification and an execution route appropriate to that scale

Why this matters: the execution asymmetry between scales is large. Multi-day-scale entry at structural support → exit at catalyst gap open captures 5-8x the return of single-session entry → T+1 swing exit, with the multi-day entry actually being lower-risk because it sits at multi-year structural support rather than intermediate session support.

The epistemic rule: when a pattern’s empirical criterion fails to match an observed event, do NOT immediately conclude “not the pattern.” First ask: “Is this the same pattern at a different scale?” Pattern criteria are scale-specific by design. A criterion built for single-session OHLCV cannot validate or invalidate the same pattern operating at multi-day or weekly scale.

The Jan 7-9 2026 VST event is the canonical example. Pre-compaction analysis classified it as “capitulation, not a stop mine” because Jan 7 had no morning rally (open=high), failing the single-session criterion. The corrected reading: the bait happened on Jan 6 (+4.05% on 1.5x volume — the multi-day-scale “morning”), Jan 7-8 was the distribution phase, the structural support was the multi-year log trendline at $149, and Jan 9 was the catalyst exit (Vistra-Meta direct PPA, gap open +15.8%). Same pattern, different clock.

See docs/PLAYBOOK_METHODOLOGY.md for builder-side guidance on scale-invariant playbook design and docs/PHASE_2_5_CANONICAL_PLAYBOOK_BRIEFING.md for the canonical author’s brief.

Relationships

Memory Hierarchy

          Structural Memory (Fib)
                 │
                 │ provides framework for
                 ▼
           Trend Memory (MA)
                 │
                 │ shows movement within
                 ▼
        Positioning Memory (VWAP)
                 │
                 │ reveals participants within
                 ▼
           Timing Layer (Momentum)

Level-Confluence Relationship

Level ──────────────────────────┐
  │                             │
  │ multiple levels at          │ confirmed by
  │ same price create           │ momentum
  │                             │
  ▼                             ▼
Confluence Zone ◄──────────── Signal
  │
  │ strength determines
  ▼
Conviction Level

Agent Synthesis Flow

┌─────────────────────────────────────────────────────────────┐
│                    Technical Director                        │
│                                                             │
│  Fib Reading ──┐                                            │
│                │                                            │
│  MA Reading ───┼──► Confluence Detection ──► Trade Plan    │
│                │                                            │
│  VWAP Reading ─┤                                            │
│                │                                            │
│  Momentum ─────┘                                            │
└─────────────────────────────────────────────────────────────┘

structural_reading → trade_plan

Each agent outputs a structural_reading object. The exact fields vary by agent, but all include a thesis (1-2 sentence interpretation):

See each agent’s *_HERMENEUTICS.md doc for the full output schema.

The Technical Director consumes all four outputs and produces a trade_plan:

structural_reading (Fib)               trade_plan (Director output)
┌─────────────────────────┐            ┌─────────────────────────┐
│ current_position        │            │ long_zones              │
│ confluence_zones        │ ────────►  │   primary_buy_zone      │
│ golden_ratios           │            │   secondary_buy_zone    │
│ structure_quality       │            │ profit_zones            │
│ next_tests              │            │ risk_management         │
│ structural_thesis       │            │ timing                  │
└─────────────────────────┘            │ agent_evidence          │
         +                             │ chain_of_thought        │
structural_reading (MA)                │ confidence              │
┌─────────────────────────┐            └─────────────────────────┘
│ trend_regime            │
│ extension_status        │ ────────►
│ dynamic_sr              │
│ monthly_context         │
└─────────────────────────┘
         +
structural_reading (VWAP)
┌─────────────────────────┐
│ positioning_map         │
│ session_regime          │ ────────►
│ trapped_cohorts         │
│ accumulation_distrib    │
└─────────────────────────┘
         +
structural_reading (Momentum)
┌─────────────────────────┐
│ divergence_status       │
│ sentiment_extreme       │ ────────►
│ momentum_regime         │
│ indicator_confluence    │
└─────────────────────────┘

The Director cites each agent’s contribution in agent_evidence and chain_of_thought fields.

Concept Definitions

Price Action Concepts

Technical Structure Concepts

Indicator Concepts

Positioning Concepts

Market Environment Instruments — The Quartet

Four instruments provide a complete environment read before looking at any individual ticker. Each is normalized and universal:

Position + Sentiment + Information Velocity + Persistence = full environment read.

The key insight: each instrument is universal. %R works on any range. VIX on any fear level. Surprisal regardless of catalyst (0.5 nats is 0.5 nats whether oil, tariffs, or Fed). Hurst classifies regime character (mean-reverting < 0.45, random ≈ 0.5, trending > 0.55) without requiring a price target.

The 1+1=3 reads:

• %R oversold + Hurst mean-reverting = STRONGEST BUY (washed out and reverts)
• %R oversold + Hurst trending = FALLING KNIFE (washed out but momentum continues)
• %R overbought + Hurst mean-reverting = pullback is buyable (stretched but reverts)
• %R overbought + Hurst trending = chase risk (stretched and momentum)

Per-ticker Hurst: The mechanical snapshot also computes Hurst on each ticker’s own returns (ticker_hurst field). NVDA can be trending while SPY is mean-reverting — that’s a different trade than if both align. Per-ticker Hurst answers “does THIS NAME revert?” independently of broad market character.

Yield curve as rate path context: Not part of the quartet (it’s not directly a regime instrument), but the environment layer also fetches FRED 2Y/10Y yields with daily deltas. rate_path_delta_2y_bps shows the bond market’s forward repricing — useful for sectors where rate sensitivity matters (utilities, financials, REITs, growth tech). The Strategist gets this context as yield_curve_data with sector-specific interpretation.

Where they surface:

• All quartet instruments appear in the report header and FLAGS section
• Williams %R, VIX, Hurst flow through market_environment in the Strategist output
• Surprisal signals are ranked in the MARKET SIGNALS report section and injected into News Analyst and Strategist prompts
• Williams %R, VIX, and Hurst regime are tracked as condition keys in the feedback loop; surprisal is not yet a condition key (it flows through flags and director reasoning)
• Yield curve flows into the Strategist prompt as a rate path context block

Market Breadth Concepts

Key Principle: Williams %R is a market-regime signal, not a ticker signal. It belongs alongside VIX as a broad environment read that modifies conviction on individual setups. Displayed in the report header next to VIX and tracked as williams_r_regime in the condition accuracy system.

Regime Classification:

Prediction Market Concepts

Key Principle: Prediction market probabilities are real-money consensus prices on binary events. They outrank headlines and pundit opinions because participants have skin in the game. When headlines say “de-escalation” but ceasefire probability is <50%, the market is NOT pricing de-escalation. Surprisal-based ranking ensures a 2pt move at 5% probability correctly ranks above a 5pt move at 50% — the low-probability event update contains more information because the market was already very confident.

Options Structure Concepts

Catalyst Mechanism Concepts

Key Principle: Narrative catalysts are probabilistic — the market can ignore them. Mechanical catalysts are deterministic — index funds must execute. When both are present (e.g., index addition during a selloff), mechanical flow provides a floor that narrative fear cannot override. The News Analyst classifies the type; the mechanical structure engine may detect the resulting flow as organic accumulation.

Catalyst Timing Asymmetry: When observable flow (volume, buying type, options positioning) precedes a narrative catalyst by days or weeks, the market priced the catalyst through information channels before the headline. This is not inside information — it’s observable logistics: military deployments visible via satellite, shipping data via AIS, insurance repricing via broker quotes, speculative positioning via CFTC. The calibration question: “Has a similar event happened before, and what did the market learn about what to price?” The probability question: “Each observable channel moves the market’s probability estimate of the downstream consequence — what’s the arc?” See NARRATIVE_FORENSICS.md.

Feedback Loop Concepts

Key Principle: When prompt strategy changes significantly, old predictions were made under different logic. Their accuracy patterns don’t reflect the new strategy. The regime system prevents this contamination without losing the data (old predictions are preserved, just filtered out).

Trading Philosophy Concepts

Ghost’s trading style has a specific philosophy that informs how analysis becomes action.

Key Principle: Prescient trading means the work happens BEFORE the trade. Analysis at 6am defines zones; execution at 2pm is mechanical (price hit zone = evaluate entry). Chasing is the opposite of prescient.

Entry Style Concepts

Every dip presents two valid entry framings. Ghost surfaces both; the trader decides which game to play.

Entry Style Characteristics:

Key Principle: All three are valid. Ghost presents each with their respective levels and R:R. The trader decides based on tape read and risk tolerance.

RETEST:  Entry $167.91-$169.96 | Stop $167.00 | Target $175.50 | R:R 3.4
SWEEP:   If $167.91 breaks → Entry $164.73-$166.61 | Stop $164.00 | Target $167.91 | R:R 1.3
BRIDGE:  If $158.42 breaks → Put wall catch $155.00 (12,500 OI, γ+) | Target $158.42 | R:R 1.7

Execution Mode Concepts

Ghost supports two execution modes that apply to every setup. The same price level can be a scalp entry, a swing entry, or both.

Mode Characteristics:

Mode Relationship:

SCALP profits fund SWING conviction

Example path:
  1. Scalp $642 → $650 (bank $8)
  2. Scalp $635 → $645 (bank $10)
  3. Price reaches $616 (structural level)
  4. Swing entry with size (funded by scalp profits, clear head)

Key Principle: A name can have a scalp setup, a swing setup, both, or neither. Ghost should surface all applicable modes.

Documentation Boundary:

• OPTIONS_HERMENEUTICS.md — Day-of-week interpretation. How to read the same data differently on Monday vs Friday. Dealer mode, tactical vs structural flip, weekly cycle mechanics.
• MARKET_STRUCTURE_HERMENEUTICS.md — Overall positioning synthesis. How to combine options data with short interest, flow, and institutional positioning into a unified structural thesis. The MS Director outputs flow_synthesis (merged flow + sweep with conflict detection) and uses a 3-read chain of thought (flow, walls/gamma, max pain).

Weekly Expiration Cycle

Options mechanics vary by day of week:

Monday     Tuesday    Wednesday   Thursday    Friday
  │          │           │           │          │
  ▼          ▼           ▼           ▼          ▼
BUILD ──► ADJUST ────► ADJUST ───► PREPARE ─► UNWIND
  │          │           │           │          │
New OI    Positions   Positions   Pre-pin    Hedges
forming   adjust      mature      activity   close

Key Principle: The same data means different things depending on where you are in the weekly cycle. Friday’s 0DTE structure dies at close; Monday starts fresh.

Deep Dive: See OPTIONS_HERMENEUTICS.md for complete day-by-day interpretation rules, including:

• How dealer mode interacts with tactical vs structural gamma flip
• Session breakdown for Friday expiration (9:30am → 3:30pm unwind)
• When to weight max pain vs technicals by day of week

Agent Responsibilities

Two-Layer Architecture: Morning Analysts + Day Desk

Ghost operates as two layers with distinct roles:

Layer 1: Gemini Agents (Morning Analysts)

The multi-agent ensemble (Fib, MA, VWAP, Momentum, Directors, Strategist) runs as a batch analysis — typically once in the morning via ghost run or the 7 AM scheduled run. These agents:

• Produce deep thesis-level analysis (narrative, interpretation, conflict resolution)
• Write structured outputs to the Knowledge Base (Firestore)
• Cost 12 LLM calls per run (5 Fib + MA + VWAP + Momentum + Sector Alignment + Technical Director + News Analyst + Strategist)
• Are the 20% that requires interpretive reasoning

Layer 2: Claude Code Agent (Day Desk)

The Claude Code agent is the trader’s primary interface throughout the day. It:

• Reads KB entries directly (everything the Gemini agents produced)
• Calls deterministic tool functions ad hoc for live data (get_williams_r(), get_options_levels(), bayesian_hold_rate(), etc.)
• Reasons across tickers, incorporates position context, answers follow-ups
• Costs zero LLM calls on Ghost’s Gemini quota (uses Claude’s own context)

The Interaction Model

7 AM:    ghost run NVDA          → Gemini agents analyze → KB updated
         ghost run VST           → Gemini agents analyze → KB updated
                                        ↓
Day:     Trader asks Claude Code  → Reads KB + calls tools ad hoc
         "What's actionable?"    → Reads all KB entries, reasons
         "Where's the zone?"    → Calls get_options_levels(), math
         "Did gamma flip?"      → Calls analyze_oi_gamma()
         "Price at $175, am I in?" → Math: $175 vs zones
         "Things changed, rerun" → ghost run NVDA (only when needed)

Signal Classification

80% of Ghost’s signals are deterministic — computed by tool functions with zero LLM cost. 20% are interpretive — require Gemini agent reasoning. See docs/SIGNAL_AUDIT.md for the full mapping.

ghost mechanical TICKER

A zero-LLM-cost quantitative snapshot calling all deterministic tool functions. Returns structured data (zones, MA regime, momentum, options, flow, environment, ticker_hurst, enrichment) that Claude Code reads for instant triage. See docs/SIGNAL_AUDIT.md for the output schema.

Canonical Price Source

get_current_price(ticker) in ghost/tools/market_data.py is THE price function for all code paths. Uses 1-minute intraday history (reliable across all sessions including pre-market and after-hours). NEVER use fast_info.last_price, fast_info.previous_close, or stock.info["currentPrice"] — they return stale (prior close) data. This rule was earned through 3 incidents of bad analysis from stale prices. All internal price helpers (mechanical.py, full_stack.py, trigger.py) delegate to it.

When to Rerun the Full Stack

The full Gemini stack (ghost run) should only rerun when the thesis has changed — detected by has_thesis_changed() in the event detection layer, or when the trader explicitly asks. Signals that warrant a rerun:

• Zone boundary crossing (price entered/exited a zone)
• Gamma environment flip (positive → negative)
• VIX regime change (NORMAL → ELEVATED)
• Williams %R regime shift
• Breaking news with high surprisal (>0.5 nats)

Signals that do NOT warrant a rerun:

• Small price moves within the same zone
• Time passing (stale data ≠ changed thesis)
• Curiosity (“let’s see what it says”) — use ghost mechanical instead

Anti-Concepts (What Things Are NOT)

Levels Are Not:

• Exact prices: They’re zones with a tolerance (±0.5-1%)
• Guaranteed: They’re probabilities, not certainties
• Permanent: They decay with time and tests

Confluence Is Not:

• Just multiple lines: Sources must be independent
• Always strong: Quality of sources matters
• Binary: Strength is a gradient

Signals Are Not:

• Commands: They inform decisions, not make them
• Timers: They indicate condition, not exact timing
• Isolated: Context determines meaning

Trends Are Not:

• Permanent: All trends eventually end
• Linear: Trends have pullbacks and consolidations
• Binary: Trends can be strong, weak, or transitioning

Related Documentation

Foundational

• EPISTEMOLOGY.md — How Ghost forms and validates knowledge
• ETHICS.md — What Ghost must and must not do
• PEDAGOGICAL.md — How to learn this system

Per-Agent

• FIB_METHODOLOGY.md — How to identify Fib structure
• FIB_HERMENEUTICS.md — How to interpret Fib structure
• MA_METHODOLOGY.md — How to identify MA structure
• MA_HERMENEUTICS.md — How to interpret MA structure
• VWAP_METHODOLOGY.md — How to identify VWAP positioning
• VWAP_HERMENEUTICS.md — How to interpret VWAP positioning
• MOMENTUM_HERMENEUTICS.md — How to interpret momentum timing signals
• MARKET_STRUCTURE_METHODOLOGY.md — How to identify market positioning data
• MARKET_STRUCTURE_HERMENEUTICS.md — How to interpret positioning overall
• OPTIONS_HERMENEUTICS.md — How to interpret options by day of week

Operational

• SIGNAL_AUDIT.md — Which signals are deterministic (80%) vs LLM-interpretive (20%). Tool function inventory. ghost mechanical output schema. The reference for what Claude Code can call ad hoc.
• CONDITION_ACCURACY.md — Feedback loop mechanics (Bayesian posteriors, hierarchical shrinkage)

Practice

• PRAXIS.md — Where analysis meets action
• DAY_DESK_REFERENCE.md — Day desk tool catalog, when to rerun
• INDEX_CONFLUENCE_PROPOSAL.md — Index/sector confluence subsystem