Layered SSOT data architecture: domain/marketdata (silver — Mbp10Snapshot/BidAskLevel, faithful port of
Rust data contracts incl. the load-bearing 1e-9 price scaling) -> domain/features (gold — MicropriceFeatures
Block N+V, a FAITHFUL port of Rust ml-features/microprice.rs, matching its test oracle bit-for-bit incl.
f64::signum(0)=1.0) -> ports/warehouse.FeatureStore (point-in-time) -> adapters/warehouse DuckDbFeatureStore
(as_of = no lookahead) -> application/feature_pipeline (silver->gold transform, causal). 11 new tests
(8 microprice vs Rust oracle + pipeline/as-of). 31/31 total.
Verified via omnisearch: the Rust formula is the weighted mid-price (commonly mis-called 'Stoikov
microprice'); Stoikov's TRUE micro-price (Quant Finance 2018) is a recursive Markov-chain estimator,
empirically a better future-price predictor — to be ADDED as a separate feature (not a silent oracle
divergence), so the gauntlet A/Bs them. Conservation-of-money ('every cent accounted for, follow the
money') adopted as a first-class principle: a warehouse/trading invariant AND the lens that prioritizes
the order-flow features (OFI/VPIN/book-deltas/Kyle-λ).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Local test with a capable model: no Claude API key in env, but OPENAI_API_KEY is set -> config-swap to
gpt-4.1 (FXHNT_AGENT_PROVIDER=openai), no architecture change (LLM-agnostic design pays off). gpt-4.1
reasons well (economic market-strategy fit, broad coverage) but first proposed SHORT trend windows that
whipsaw -> 0 specialists. Added the empirical lookback prior (trend ~100-250, mean_reversion ~10-40) to
the hunter prompt; re-run found 2 quality specialists from 6 candidates (mean_reversion GC trend_down|
high_vol, trend(120) NQ trend_up|low_vol) at global trials 6 not 18 — the discipline win of agentic
search with a capable model. 22/22 tests.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
application/agent/hunter.AgentHunter: an LLM proposes the next batch of (market, strategy, window)
candidates, reasoning about strategy-market fit and the gaps the current population leaves — clamped to
the real universe + known strategies so hallucinations can't enter. FactoryRuntime now takes a
territory_provider (callable): fixed grids OR the agent's per-cycle proposals. CLI run --hunters agent.
Verified on local qwen2.5:3b: proposed 3 FOCUSED candidates (vs 18 grid) -> global trials 3 not 18 (the
discipline win of agentic search: lower multiple-testing burden) -> found buy_hold on GC in
trend_down|high_vol (gold flight-to-safety in panic). fxhnt 1+2 complete: [hunt] 6 new candidates → 1 specialists | global hypotheses 9 | by regime {'trend_down|high_vol': 1}
[book] as of 2026-06-08 | 0 active, 2 dormant | weights (flat)
on cron = the standing agentic force that constantly seeks regime-specialists and runs them side by side.
22/22 tests.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
application/runtime.FactoryRuntime.cycle: hunt new regime-specialists (fleet) then assemble today's
regime-adaptive book (regime execution) in one pass. CLI run (single cycle = cron-friendly, like the
foxhunt paper trackers; --interval-minutes for a foreground daemon). Distinct-hypothesis DEDUP added to
the store (SeenCandidateRow + is_seen/mark_seen) and wired into the fleet, so re-running across cycles
does NOT re-inflate the global trial count or re-do work — honest multiple-testing accounting over time.
Verified: cycle 1 = 18 new candidates / global 18; cycle 2 = 0 new / global stays 18. 22/22 tests.
(Schedule [hunt] 0 new candidates → 0 specialists | global hypotheses 18 | by regime {}
[book] as of 2026-06-08 | 0 active, 3 dormant | weights (flat) via cron for the always-on force. Next: LLM-driven hunters = fxhnt 2.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
application/regime_execution.RegimeConditionalExecutor: reads specialists from the store, detects TODAY's
regime per market, activates only those whose regime is live, nets their latest positions into per-market
target weights (gross-capped). CLI book. The multi-strategy book is now regime-adaptive — trend specialists
run in calm uptrends, mean-reversion in panic, automatically in/out by regime. Live demo: as of 2026-06-08
all 3 discovered specialists are DORMANT (NQ in trend_up|HIGH_vol not its calm-trend regime; GC in calm
downtrend not its panic regime) -> book correctly FLAT. Discipline: don't trade out-of-regime. 21/21 tests.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
application/fleet: Territory (a hunter's mandate: family x universe x param grid) + FleetOrchestrator —
searches each territory, evaluates every candidate REGIME-conditionally (layer 1), writes the regime-
specialists it finds to the lifecycle store (layer 2) with honest global trial accounting. CLI hunt.
Hunters sequential now (local GPU serializes), interface ready for concurrent/agent-proposed later.
Live on cached futures: 2 hunters, 18 candidates -> 3 specialists (NQ trend in trend_up|low_vol;
GC mean_reversion in trend_down|high_vol — a non-trend specialist the full sample would miss). 20/20 tests.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
ADR 0002 designs the real vision: a standing fleet of agents continuously breeding regime-specialist
strategies into an adaptively-allocated book, with multiple-testing-at-fleet-scale as the constraint the
whole architecture is built around (structured search + regime-conditioning + forward-validation +
global-N accounting = the moat). domain/regime: VolTrendClassifier (causal trend x volatility state) +
evaluate_by_regime (judge a strategy WITHIN each regime). Demo on real NQ: trend(200) is MARGINAL on the
full sample but a clear OOS-confirmed SPECIALIST in trend_up|low_vol (DSR 0.984, OOS +0.93), correctly
rejected in range/high-vol regimes — the architecture's thesis proven. 16/16 tests.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
StructuredResearchAgent now collects every book it tests (dedup'd, returns kept via BookEvaluation.
is/oos_returns) and at the end re-judges ALL of them at the EFFECTIVE number of independent books
searched (effective_trials over their return streams) — the same multiple-testing discipline the grid
discovery uses, applied to the agent's OWN search. A book wins only if it clears the full-search bar.
Verified on local qwen2.5:3b: tested 4 books (~4.0 effective), NQ+6E trend SURVIVES at n_trials=4
(DSR 0.971) while weaker combos correctly reject. LLM proposes; gauntlet + multiple-testing dispose.
14/14 tests.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
application/agent/structured.StructuredResearchAgent: a LangGraph StateGraph that constrains the LLM to
one job (emit a structured ProposalBatch of books) while the graph runs each through the portfolio
gauntlet deterministically and loops. Robust where free-ReAct failed on the weak 3B model. CLI agent
--mode structured (default) | react. VERIFIED on local qwen2.5:3b: reliably drove the loop (6.5s) and
found the platform's FIRST passing book -> trend on 6E+NQ (EUR/USD x Nasdaq, uncorrelated): Sharpe 0.82,
OOS 0.86, maxDD -10%, DSR 0.971 PASS; correctly rejected weak combos (ZN+CL DSR 0.68). LLM proposes,
gauntlet disposes. 14/14 tests. (Caveat: n_book_trials fixed at 10 — should track the agent's full
search breadth for complete trial-accounting honesty.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
application/agent: ResearchAgent (LangGraph create_react_agent) + build_tools (list_strategies,
evaluate_candidate, evaluate_book wrapping the services) | adapters/llm.build_chat_model (config-
swappable: local Ollama now, cluster/API later) | AgentSettings | CLI agent. The LLM proposes; the
deterministic gauntlet inside the tools disposes. Verified: tools work + the model emits tool-calls
(architecture correct). Local qwen2.5:3b is too weak for the reliable multi-step loop (model bottleneck,
swappable via config). Also fixed flaky test seeds (hash()->deterministic). 14/14 tests. Bonus: tools
surfaced GC+NQ+ES trend book = Sharpe 1.06 OOS 1.47 DSR 0.949.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
domain/portfolio.combine_returns (date-aligned weighted sum of sleeve returns) + application
PortfolioEvaluationService (backtest each sleeve, combine, gauntlet on the book) + CLI validate-book
+ DRY _data_provider helper. The unit of alpha is the diversified book: live futures-trend book
(6 markets) = Sharpe 0.81, OOS 0.91, maxDD -9%, DSR 0.947 (n=1) vs individual sleeves' DSR 0.02-0.04
-> the book is vastly more robust, borderline-significant (exactly what a real CTA book is). 14/14 tests.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Blind grid deflated by RAW N over-penalises: trend-150/200/250 on the same asset + buy-hold on
correlated markets are NOT independent tests. Added gauntlet.effective_trials (design-effect:
N_eff = N/(1+(N-1)*rho_bar), rho_bar = mean off-diagonal correlation of candidate returns). Discovery
now deflates by the effective N. Live: 48 futures candidates -> 11.2 effective trials (bar ~1.7 -> ~1.0
Sharpe). Genuine edges (NQ/GC trend, OOS 1.2-1.35) now near-but-still-below the individual bar -> the
PORTFOLIO is the right unit (next). 12/12 tests green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Databento historical fetches are slow (~30-90s/symbol for continuous futures over 8y) — so ingestion
must be decoupled from search: fetch+cache a universe once, search it many times fast. Added fxhnt
ingest (pre-fetch into the analytical store), --data-source on discover (yahoo|databento), and the
missing Settings.databento field. This is the data-layer-value lesson made concrete (the user intuited
it). 11/11 tests green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
application/discovery: CandidateGenerator (seam for an LLM agent later) + GridGenerator + DiscoveryService.
The discipline that makes it honest at scale: every candidate judged with n_trials = total searched and
sr_variance = variance across ALL candidates' Sharpes -> the deflated-Sharpe bar accounts for the whole
search. Shared run_id (domain), per_period_sharpe helper. CLI . Verified on real data: 32
candidates -> 0 survivors (CORRECT: trend-on-SPY passes at n_trials=1 but not at 32 -> multiple-testing
correction working, the anti-false-positive core proven at scale). Perf: persistent DuckDB connection
+ only survivors persist timeseries (was connect-per-op, 87s I/O). 11/11 tests green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
3 strategy templates now (trend, mean_reversion contrarian z-score, buy_hold baseline) — gauntlet
correctly rejects mean_reversion on trending SPY (strategy-market fit working). DatabentoDataProvider
behind DataProvider port: survivorship-free daily bars across futures/equities (GLBX 2010+, equities
2018+), with a COST GUARD (estimate via metadata + refuse above cap before download — the $3.5k lesson
as code). Verified live: guard blocks when cap too low; ES.c.0 fetch ok (620 bars @ $4812). data extra.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
adapters/broker/ibkr (IbkrBroker implements Broker via ib_async: managed connection, paper detection
via DU* prefix, whole-share orders, gross-position-value for the data-consistency gate) | IbkrSettings
config | domain/portfolio.book_from_runs (survivor library -> executable Book) | CLI
(composition root wiring IBKR + ExecutionService, dry-run default). Verified live against IBKR paper
(NLV $1.04M): research -> gauntlet -> survivors -> book -> netting -> plan. 10/10 tests still green.
ib_async is an optional extra (ibkr).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
