Features

Pricing & strategy

Poisson + Dixon-Coles fair-value model, fixed-point bps math throughout
Five individually toggleable regimes (stoppage-time, mutual-danger, narrative, informed-flow, stale-quote decay) — see Core Concepts
Avellaneda-Stoikov-adapted inventory skew (reservation price shifts against net inventory)
Kelly-fraction position sizing with configured min/max stake bounds
Quote-odds sanity bounds — no markets made on near-certainties or deep longshots

Hash-chained decision ledger with CI-enforced replay(corpus) === ledger
Ed25519 signing of every decision record
Real Merkle tree + inclusion proofs (/ledger/proof)
Real Solana on-chain anchoring: pre-match commitment + periodic checkpoints
Signed policy config snapshots, appended only on real change
Source-proof verification (Merkle proof + on-chain validate_odds/validate_stat) required before any message is admitted

Feed-gap halt, unexplained-movement halt, informed-flow halt, model-divergence band
Per-fixture and portfolio-level exposure caps and drawdown kill (operator-restart-only)
Aggregate proof-failure-rate circuit breaker, distinct from per-message rejection
Network-consistency enforcement between pricing input and anchoring keypair

Strategy Arena: Tissue vs neutralized Baseline, same feed, same engine, same grader
N-way regime ablation matrix: each regime isolated individually against baseline
Deterministic sha256-bucketed calibration/holdout split, honest about being underpowered on small real samples
Human-reviewed tuning suggestions (evaluate:tuning), calibration-only, never auto-applied
evaluate:real — hard-fails on a clean checkout with no real TxLINE data, never falls back to synthetic

Dual independent TxLINE SSE streams with reconnect, Last-Event-ID, JWT renewal
Deterministic crash recovery: persisted ledger prefix reverified against a fresh corpus replay on restart, never trusting a cached receipt as source authority
Prometheus-format metrics: proof/stream/SSE counters plus real latency histograms for proof-verification time and end-to-end decision-loop time
Liveness/readiness endpoints (/health, /ready)
Read-only SSE evidence stream (/events) for external observers
Three-service Docker Compose (daemon, dashboard, analyst) with health checks and shared persistent evidence storage
Pruned, production-only compiled runtimes; CI exercises compiled entry points, not tsx

Live decision feed with regime badges (extra-time/penalties, stoppage, mutual-danger, narrative regime) shown only when non-default
Halt banner with distinct, reason-specific explanations per halt type
On-chain commitment timeline (pre-match + periodic checkpoints, real explorer links)
TxLINE input verification list (per-message proof evidence)
Strategy Arena view + regime ablation matrix table
Equity curve (realized PnL / drawdown per decision, from already-tracked exposure state)
Shareable, on-demand grade card (SVG export)
Loading, empty, error/retry, halt, and hash-verification states throughout

Isolated SQLite projection, readOnly: true at the connection level — no write tool exists
8 read-only MCP tools across 3 explicit skills: ledger forensics, Slip market intelligence, Slip settlement audit (see Reference)
Structured pattern recall (find_similar_decisions) — explicitly not a fabricated embeddings claim; every result is a real, citable ledger row matched on radar class, match minute, and edge magnitude
Live-model integration test proves the full model → MCP → Slip SDK → canonical market path without monkey-patching a response