← Trading Hub

AI features — model cards

Per Execution Bible §15.61: every AI feature ships a public model card. Purpose, prompt context, input/output examples, limitations, bias, eval results. Reviewed quarterly. Linked from /trust scoreboard.

AI principles (apply to every feature below)

Live AI features

Scoring engine
v2.7
Multi-factor weighted scoring across 14 asset classes. 25-30 factors per asset; layer-conditional weights via LS6 regime detection. Tier output: Strong Buy / Buy / Hold / Sell / Strong Sell.
Tier 1-4 Live
deterministic · no LLM · <5ms per ticker
Prescience (forecast)
v1.4
Closed-form GBM forecast with 14-asset-class profile table; produces p10/p50/p90 cones, regime persistence, lead indicators, catalyst countdowns. No Monte Carlo client-side.
Tier 1-2 Live
deterministic · no LLM · ~10ms per security
Playbook (action plan)
v1.2
Generates entry zone, target, stop, R:R, expected hold, abort triggers, and "when NOT to act" disqualifiers. Quarter-Kelly sizing with 15% hard cap per position.
Tier 2 Live
deterministic · no LLM · per-security
Options Strategy
v1.0
21 canonical options strategies + decision rubric (direction × IV × confidence × event proximity). Black-Scholes pricing, terminal-payoff POP, log-normal EV. Ranked by composite score. Powers /options.
Tier 2 Beta
21 strategies · BS · POP · EV · cap-tier eligibility
AI commentary (/ask)
v1.1
Natural-language commentary on scores, regimes, and tickers. Uses Anthropic Claude. Cites factor IDs from the scoring engine. Hallucination eval pending.
claude-sonnet-4-6 Live
~$0.005-0.02/call · 95% groundedness target

Foundational engines (factor + regime + math)

Factor inventory
v2.7
30+ factor library powering scoring. Tier 1-4 hierarchy (macro / sector / instrument / sentiment). New factors added proactively per Bible §6.6.
Tier 1-4Live
deterministic · <1ms per ticker
Regime engine
v1.0
Cross-asset regime classification with AR(1) persistence half-life. 4 states: risk-on / risk-off / transition / recovery.
Tier 1Live
deterministic · ~5ms per pass
Layer-conditional weighting (LS6)
v1.0
Factor weights vary with regime + asset class. Risk-on rewards momentum; risk-off rewards quality. Crypto rewards funding signals.
OrchestratorLive
deterministic · per-score modulation
Walk-forward validation
v1.0
Rolling in-sample → out-of-sample factor-weight evaluation. Reports OOS Sharpe + IS/OOS ratio for overfitting detection.
ValidationLive
deterministic · cron-friendly
News velocity (v95)
v1.0
Story-rate × sentiment composite. Storm flag fires when story-rate >5σ above baseline AND |sentiment| ≥ 0.4.
claude-haiku-4-5Live
~$0.003/headline · 30s/24h windows
Calibration scorecard (v104)
v1.0
Brier + reliability tracker for every score / forecast / probability output. The third leg of the named dimension. 30d-warmup before populates.
Tier 130d-warmup
closed-form · weekly aggregation
Alerts pipeline
v1.1
Real-time delta watcher: score jumps, tier transitions, lead-indicator firings, regime flips. 30-min dedup. Audited 2026-04-26 (silent no-op fix).
Tier 1Live
30s tick · in-memory ring + store.events log

Beta / scaffolded

Counter-thesis
v0.3
Generates the strongest argument AGAINST the current recommendation. Surfaces blind spots in factor coverage. Designed to disagree with itself on purpose.
claude-sonnet-4-6 Beta
~$0.01-0.03/call · groundedness eval pending
Decision replay
v0.2
Replays "what would the engine have said on date X" against historical bars. Compares model output to realised outcome. Inputs to calibration scorecard.
claude-haiku-4-5 Beta
~$0.005/call · auditable trail
Daily newsletter
v0.4
13:00 UTC weekday synthesis: top movers, regime, calendar, risk-off triggers. Cron-driven, single LLM call, stored as static artifact. Read at /now.
claude-sonnet-4-6 Beta
~$0.05/send · 5 sends/week

Versioning + eval policy (Bible §15.111 + §15.112)

Every prompt is a versioned artifact at prompts/{feature}-{version}.md with frontmatter (version, created_at, model, eval_score, notes). Code references prompts by (feature, version) — never inline strings. Prompt changes ship behind feature flags, A/B tested before promotion (Bible §15.66).

Offline eval set ≥50 input/output pairs per feature. Weekly groundedness sample of 50 outputs per feature; LLM-judge + heuristic checks. Regression >5% blocks merge. Eval results stored alongside prompt version. Public per-feature trends at /trust.

Last reviewed: 2026-04-26 · Next review: 2026-07-26