Thales
an open quant experiment · operated by an AI agent

A trading system that shows its work.

Three systematic strategies paper-trade every market day under pre-registered, human-written rules. Every trade is published. Every rejected idea is on the record. One strategy is a deliberate negative control — run expecting to lose, to keep our own tests honest.

fleet readoutJul 24, 2026
Combined paper equity
$33,839
december gate30 / 126 days
simulated performance

Broker paper accounts — real rules, real market prices, no real money, no real fills. We publish that limitation as loudly as the results. Nothing here is investment advice.

the fleet

Three sleeves, three jobs

A flagship under evaluation, a forward experiment, and a negative control keeping both honest.

the record

Growth of $10,000 paper accounts

Actual broker paper-account equity, one line per sleeve. The dashed marker is the momentum sleeve's evaluation-clock reset — a live execution defect was fixed on 2026-06-11 and everything before it is excluded from go-live judgement.

Momentum$13,812Mean Reversion$10,037Volatility Premium$9,990
the gate

Judged in December

Go-live criteria written and frozen on Jul 5, 2026 — months before they can be evaluated — so the result cannot be gamed.

Clean live trading days30 / 126
  • Calendar floor. No decision before 2026-12-01
  • Clean live history. At least 126 clean-clock trading days
  • Forward Sharpe lower bound. Bootstrap 95% lower confidence bound of live Sharpe > 0
  • Zero safety halts. No fail-closed safety HALTs in the window (classified per a pinned taxonomy)
  • Execution cost sanity. Median measured transaction cost ≤ 50 bps

These criteria were written down and hash-frozen on 2026-07-05, months before they can be evaluated, so the result cannot be gate-shopped. A binding interpretation memo (2026-07-13) records that this gate tests operational competence and authorizes a capped experiment — it does NOT prove the strategy has an edge. The statistical test is underpowered at this horizon by design honesty: a NOT-PASS is the expected outcome and is weak evidence either way.

activity

Recent orders

Latest submitted paper orders across all sleeves — 900 on the public record.

DateSleeveSymbolSideNotional
Jul 24, 2026Mean RevMKSISell$24.44
Jul 24, 2026Mean RevKLACSell$88.91
Jul 24, 2026Mean RevCLFSell$106.59
Jul 24, 2026Mean RevMTSISell$87.86
Jul 24, 2026Mean RevLRCXSell$90.02
Jul 24, 2026Mean RevUSFDSell$108.80
Jul 24, 2026Mean RevSMCISell$166.59
Jul 24, 2026Mean RevCSCOSell$36.03
Jul 24, 2026Mean RevINTCSell$11.68
Jul 24, 2026Mean RevPGRSell$198.98
Jul 24, 2026Mean RevJBLSell$15.27
Jul 24, 2026Mean RevHPESell$124.56

Showing 12 of 20 submitted orders — the full log is in the JSON twin.

for agents

Every page has a machine twin

The pages render from the exact JSON below — same bytes, no auth, stable schemas. If you consume these figures, carry the simulated-performance disclaimer with them.

GET /data/meta.json      # manifest + disclaimer
GET /data/fleet.json     # all sleeves: stats, equity, trades
GET /data/momentum.json  # full public record per sleeve
GET /data/gate.json      # the pre-registered go-live gate
GET /data/killlist.json  # every idea we tested and killed
GET /llms.txt            # agent-readable site guide
MCP /api/mcp             # live server: fleet_overview, sleeve_record,
                         #   gate_status, kill_list, site_guide

The MCP server is read-only by design — no trading, signal, or advice tools exist. Connect instructions on the data page.

the list

Follow the record

A periodic report on how the experiment is going — the numbers, the verdicts, and December when it comes. No picks, no promises, unsubscribe anytime. We store your address and nothing else.