Atlas

Atlas is a cross-market prediction-market research system. It watches Kalshi and Polymarket markets, normalizes their contract language, identifies possible economic equivalents, verifies those relationships with deterministic rules, simulates executable order-book pricing, and records what a paper trade would have done.

The product thesis is:

AI finds relationships. Deterministic rules decide whether a relationship is real. Order-book math decides whether the price is actionable. Settlement evidence decides whether the learning label can be trusted.

The long-term product may become a mispricing terminal, alerting service, or data product even if the trading strategy does not survive live execution.

Current operating boundary

Atlas is paper-only. No real order-placement path is enabled.

This is an intentional safety invariant, not a temporary UI setting:

• ATLAS_TRADING_ENABLED remains disabled.
• The API health endpoint reports trading_enabled=false.
• atlas.paper records simulated executions only.
• atlas.shadow and atlas.live_monitor observe live markets but never submit orders.
• Relayer and execution credentials are not used by v0.1.
• Model or semantic proposals cannot bypass catalog-ID validation or deterministic verification.
• REVIEW_REQUIRED, inconclusive, unknown, or non-guaranteed pairs cannot become trusted approval labels. Evidence-backed REJECTED labels may derive only from same-canonical-subject review pairs whose terminal settled outcomes diverge on both venues (owner-signed decision, 2026-08-13).
• Public venue requests use bounded retry budgets so a stalled catalog cannot block the monitor indefinitely.
• The continuous monitor catches read-only venue failures, reports NEVER_EXECUTED, and retries on the next interval.

Do not add a live order client casually. Any future execution capability must be a separate module with an explicit promotion gate, default-off feature flag, independent risk controls, audit records, and a deliberate human authorization step.

The research phase is complete — read this first

Five pre-registered hypotheses, five negatives, no principal spent. The one-page record, with every verdict linked to its charter and evidence: docs/FINDINGS.md. The system keeps running as a paper-only instrument; the sections below describe how.

Where we stand

2026-08-20 — the study was measuring an untradeable venue, and the answer changed. Every one of the first 18,650 gap observations priced a Polymarket Global leg: offshore, no published order book, and declared non-executable by its own adapter. The radar now also watches Polymarket US — the venue a US account can reach, which publishes a two-sided book so both legs can be sized for real. Against that venue the picture is different: 36 twin-shaped pairs, 2 executable, both GDP Q3 at under a cent; a hand-run snapshot across FOMC Sep/Oct, CPI YoY, and unemployment found 27 pairs and 0 executable, with baskets costing $1.00–$1.02. The two US venues are arbitrage-linked and agree on price. Every tradeable executable gap on 2026-08-20 was below the floors — a 7.8c GDP gap was backed by 0.06 contracts of Polymarket depth.

The study's go/no-go now charges for capital lock-up (the phase-2 metric, pulled forward from day 31) and flips GO → NO-GO on unchanged data: 3.38% annualized on locked capital overall, 1.66% on the tradeable venue, median basket $459. On that evidence the owner dropped the execution track and pivoted to contract intelligence. The staged path in "Beyond paper-only" below is now dormant, not pending — paper-only stands indefinitely. Details: docs/NINETY_DAY_STUDY.md (Amendments, 2026-08-20).

Last verified: 2026-08-20 (full suite + ruff check + live API/dashboard read).

以上為 README 的開頭部分。其餘內容請閱讀原文。 檢視完整 README ↗