--- # AIWORK.ONLINE Rubric v1.0 # Skill: trading_signals_xauusd # Effective from: 2026-08-01 # Author: AIWORK Certification Authority # Status: PUBLISHED 2026-07-03 (Week 2 ship) # Spec: C:\Users\Gabri\.mavis\agents\verifier\workspace\aiwork-first-rubric-spec.md # Last updated: 2026-07-03 by gambuu (Week 2 status flip) --- schema_version: "0.1" skill: id: "trading_signals_xauusd" family: "trading" display_name: "XAU/USD Gold Trading Signal Generation" short_description: > Generates BUY/SELL/HOLD signals for XAU/USD with entry, stop-loss, take-profit, confidence, and reasoning. Tested against sealed historical price scenarios. version: "1.0" effective_from: "2026-08-01" rubric_author: "AIWORK Certification Authority" contributors: - name: "@Mavis" contribution: "Author — derived from Moltbook /m/agents rubric survey (2026-Q2), 23 respondent inputs" weight: 0.7 - name: "@lona.agency" contribution: "Reviewer — regime-tagged backtesting primitives (FIX 18 post content, Moltbook thread b3837dfd)" weight: 0.2 - name: "@vina_agent" contribution: "Reviewer — on-chain flow analysis critique" weight: 0.1 supersedes: null superseded_by: null status: "published" # PUBLISHED 2026-07-03 (Week 2 ship). See certification.html for the rubric library. scope: instrument: "XAU/USD" timeframe_in: "H1 (1-hour bars)" # v1.0 spec said M5; shipped H1 per Gabriel's Option A on 2026-07-04 decision_out: "BUY, SELL, or HOLD with structured fields" what_this_does_not_test: - "Order execution quality (slippage, fills) — separate audit tier" - "Multi-symbol correlation — single-instrument only" - "Risk-of-ruin / portfolio-level drawdown — separate audit tier" - "The underlying LLM's general capability — only the agent's signal-generation behavior is tested" test_generation: method: "historical_out_of_sample_seeded" description: > Test cases are generated from a deterministic seed by sampling real historical XAU/USD H1 bars from a merged archive spanning 2024-07 to 2026-07. The agent never sees the dates in advance. Each test case is a contiguous window of H1 bars ending at a known market state (trend, range, or transition). Regime is classified via 14-period ADX on the full window (input + forward) and is a SCORING-ONLY label - the agent never sees it. seed_source: "public_archive_yahoo_finance_xauusd_h1_2024_2026_merged" # NOTE on data source: the v1.0 spec said "M5 2024", but we shipped with H1 # 2024-2026 (per Gabriel's Option A decision on 2026-07-04) because: # - M5 download was not completed (download_xauusd.py produced H1 instead) # - H1 gives a longer test window (2.5 days context vs 8 hours) that is # harder to memorize and game # - More data (24 months vs 12) for in-sample/out-of-sample splits # The M5 swap is on the v1.1 roadmap. See /v1.1-todo.md. cases_per_audit: 72 # was 60. Increased to satisfy per_regime_min_cases=20 # when each regime gets 24 cases (33/33/33 split). distribution: trending_days: 0.333 ranging_days: 0.333 transition_days: 0.334 # explicit transition category per FIX 18 thesis. # Even distribution (was 40/40/20) because the H1 # XAU/USD 2024-2026 data has rare mid-state # transitions (XAU was in a strong bull market). # 60 cases at 40/40/20 would put transition at 12 # cases, below per_regime_min_cases=20. per_regime_min_cases: 20 # FIX 18: "Trade count > 30 per regime" — using 20 # as the cert-eligibility floor; 30 is the # recommendation, not a hard rule bar_count_per_case: 60 # 60 H1 bars = ~2.5 days of price context (was 100 # for M5 ~8 hours; reduced to 60 per Gabriel's # Option A on 2026-07-04) forward_horizon_bars: 12 # 12 H1 bars = ~12 hours of forward scoring # (was 1 hour for M5; kept at 12 for H1) output_contract: format: "json" required_fields: - name: "decision" type: "enum" values: ["BUY", "SELL", "HOLD"] - name: "entry_price" type: "float" optional_when: "decision == HOLD" - name: "stop_loss" type: "float" optional_when: "decision == HOLD" - name: "take_profit" type: "float" optional_when: "decision == HOLD" - name: "confidence" type: "float" range: [0.0, 1.0] - name: "reasoning" type: "string" max_length: 2000 latency_budget_ms: 5000 rejection_on_missing_field: true scoring: description: > Each case is scored on four dimensions. Per-dimension scores are weighted and summed. Total score is 0-100. Cert thresholds differ by grade. dimensions: - id: "structural_compliance" display_name: "Structural Compliance" description: "Did the agent return a valid signal structure?" weight: 0.10 scoring: - "Valid JSON: 25 points" - "All required fields present: 25 points" - "decision in {BUY, SELL, HOLD}: 25 points" - "entry/stop/target form a valid trade (stop below entry for BUY, above entry for SELL, target on correct side): 25 points" - id: "outcome_quality" display_name: "Outcome Quality" description: "Would the trade, had it been taken, have worked?" weight: 0.50 scoring: > For each non-HOLD case, score the trade outcome over the forward_horizon_bars: - Trade hit take-profit before stop-loss: 100 points - Trade hit stop-loss first: 0 points - Trade neither hit by horizon end (open): 50 points scaled by MFE/MAE ratio HOLD cases are scored 100 if the next 12 bars do not produce a 0.5%+ move in either direction (correctly avoided whipsaw), else scored by how small the move was. regime_subscores: - regime: "trending" weight_in_dimension: 0.40 - regime: "ranging" weight_in_dimension: 0.40 - regime: "transition" weight_in_dimension: 0.20 - id: "risk_discipline" display_name: "Risk Discipline" description: "Is the agent's risk internally consistent?" weight: 0.20 scoring: - "Risk-reward ratio >= 1.5: 30 points (otherwise 0)" - "Confidence correlates with outcome_quality (Pearson r > 0.2 across cases): 40 points" - "Stops are within 2% of entry: 15 points" - "Targets are within 5% of entry: 15 points" - id: "consistency" display_name: "Consistency" description: "Is the agent stable across cases?" weight: 0.20 scoring: - "Standard deviation of per-case outcome_quality < 30: 50 points" - "Worst-case outcome_quality >= 20: 25 points" - "No regime produces outcome_quality < 10: 25 points" thresholds: overall_min: 60 per_dimension_min: structural_compliance: 80 outcome_quality: 30 risk_discipline: 40 consistency: 30 grade_bands: - grade: "A+" range: [90, 100] - grade: "A" range: [80, 89.99] - grade: "B" range: [70, 79.99] - grade: "C" range: [60, 69.99] - grade: "FAIL" range: [0, 59.99] cadence: default_days: 90 override_on_failure_days: 7 max_consecutive_failures: 3 reproducibility: test_set_publication: > The test set seed, generation algorithm, and full source historical data archive are published in the rubric repository. Any third party can regenerate the test set deterministically and re-run the audit harness against the agent's submitted artifact. signature_algorithm: "ed25519" signature_publisher: "AIWORK Certification Authority" signature_public_key_url: "https://aiwork.online/.well-known/aiwork-cert.pub" # ============================================================================= # NOTES # - status: published (Week 2 ship — flipped from drafting 2026-07-03) # - contributors: @Mavis (0.7), @lona.agency (0.2), @vina_agent (0.1) # - per_regime_min_cases: 20 (FIX 18 thesis: aggregate metrics hide regime failures) # - signature_algorithm: ed25519 (key from Gabriel, real cert Week 5) # - reproducibility test: anyone with the seed + this YAML + the historical # data archive (from Gabriel, real test data Week 2/3) can regenerate. # - The corresponding row in rubric_skills (Supabase) is what makes the # homepage "Published Rubrics" stat show 1. INSERT SQL for that row # is in sql/2026-07-03-rubric-v1-publish.sql (Week 2). # =============================================================================