AI Trader

No snapshots yet — the bot writes one every 5 minutes once deployed.

Watchlist warms up after the next bot tick.

Open positions

No open positions.

Desk integrity — is the judgment layer actually reading its data?

The desk's two logged reading failures were not bad judgment: INTC was vetoed for “no supportive volume” on a record-volume day, and OIL was shorted on “volume falling 17k→11k” while volume was rising 45%. Both came from a payload too large to hold — the technical analyst receives ~30 KB of bars and answers in ~950 characters. The machinery below exists to catch that class of error. Each line reports the number that decides whether it is earning its place, not merely that it ran.

Data custodian — proofreads each analyst against an exact, code-computed extraction of the same bars

Scanner second opinion — a free model run in parallel; approvals are unioned, never intersected

PM buffer — Opus 5 on the Max plan reads the full payload the PM compressed

Live model roster — what has ACTUALLY been serving (from the journal), checked against what is configured

A model swap only appears here once that role next runs, so a role can legitimately lag config for a while. A persistent mismatch means something else: either the rollout did not take, or the role is silently failing over down its fallback chain.

Mechanical twin — no-LLM control group

Three mechanical engines (trend / Donchian / mean-reversion, replicating the hackathon-bot design) over the shared risk core, on a separate $10k paper book. Same prices, same costs, zero LLM judgment — if the main desk can't beat this over time, the judgment layer isn't paying its way.

DCF value sleeve — 20-trading-day book

A separate $10k ledger holding the five cheapest names by discounted-cash-flow value-to-price, equal weight, long only, no leverage and no stops, rebalanced every 20 trading days. This is the desk's only signal validated point-in-time (21,927 observations, 20 symbols, 5 years, intrinsic value rebuilt from filings public at each date): cheapest-minus-richest quintile spread +0.378% at 20 days, ~0 at 5 days. It is deliberately outside the short-term trade path — the edge does not exist on the desk's horizon. Judge it against buy-and-hold, not zero: most of the absolute return is market drift, the spread is the edge.

Sleeve is in cash — next rebalance will fill the basket.

Two-stage discounted cash flow on trailing free cash flow: growth fades to a 2.5% terminal rate, discounted at CAPM (live 10-year yield + beta × 5% equity risk premium), net debt subtracted. Foreign-currency filers are excluded — a TWD intrinsic value cannot be compared with a USD price. Only six names are re-valued per night on a rolling basis, so a row can be several days old — the price shown is the one captured when that valuation was computed, not spot, and upside is measured against it. It is a valuation, not a trade signal.

No valuations yet.

Nightly backtest — frozen rulebook (walk-forward)

The mechanical rulebook replayed over the trailing month each night after the NYSE close — the LLM layer is not simulated, so this is the rulebook's floor, not a forecast.

First nightly run lands after the next NYSE close.