Milestones — the project story

This is a factual history reconstructed from the repository's own records (taskboard, decision records, forensic reports). No invented narrative.

Phase I — Engine core (single-agent era)

The original engine: hexagonal architecture, MT5 adapters, 50D causal features, ScalpNet, policy matrix, risk engine, order manager, SQLite WAL ledger, FastAPI Control Center. Foundation invariants established: no sync DB on the tick path, no order authority for learning components.

Phase II — Multi-agent forensic era

Development moved to a multi-agent model (up to 56+ parallel AI agents) with a strict engineering contract: taskboard, locks, bug ledger, runtime invariants, change control. The bug ledger (BUG-001…) and forensic acceptance reports date from this era. Key products: database hygiene, migrations, incident response, observability architecture, release/update engine.

Phase III — The 60D/70D research series

The defining scientific arc. The base 50D contract was extended with news context (10D) and liquidity intelligence (10D) into the 70D canonical contract (scalp_v3). A complete parity/validation stack was built: schema hashing, inference validator with 10 rejection codes, replay anti-leakage, golden corpora, fair A/B/C benchmark protocol.

The honest result: the 70D candidate's real-data walk-forward and shadow benchmarks came back negative/inconclusive (OOS NOT_ELIGIBLE). The live contract deliberately stayed 50D. This rejection is preserved as a first-class result — it is the proof the validation gates actually gate.

Phase IV — Model Factory & runtime hardening

Artifact-first model governance: versioned datasets/experiments/models with manifests (inference needs no database), 10-gate model load gate, promotion transactions, rollback previews. Live-engine hardening: record-contract alignment (BUG-185), account-identity fail-safe (BUG-142), confidence semantics repair (CHG-0042).

Phase V — Research execution stack

Streaming replay, forward tests with frozen captures, canonical tick datasets with provenance, and the NO_TRADE counterfactual engine (CHG-0041): walking decisions the engine didn't take, with stratified evidence (CONFIDENCE_GATE covered N=393 FR 45.0% meanR −0.506 — a valid filter; SUPPORT-margin FR 60% meanR +1.35 — flagged for policy review).

Where this leaves us

The platform's differentiator is not a winning strategy claim — it is that

every claim is falsifiable and the falsification results are published. Current direction: see the Roadmap.

Milestone index (selected)