The Ghost Desk
Watch AI Learn to Trade
An autonomous AI committee trades SPY options on a paper account. This page is its honest record — every call scored against the real tape. Updated after the close · 2026-08-12 16:02 ET.
Closed trades
76
the accrual clock · dollars withheld
Days live
64
public record since 2026-06-09
First read
2026-10-01
in 50 days
Open positions
1
stale marks present

Why no dollars — yet

Every dollar is tracked against real quotes and time-stamped. We don't publish the dollars yet because at this sample size they are statistically noise — we pre-registered the dates we'll read them (2026-10-01 / 2026-12-31) and the conditions under which we'll call the experiment dead. What we publish now is what can actually be evaluated today: whether the system's stated probabilities are calibrated, whether its claims beat chance, and what it learned this week.

Interim read
2026-10-01
first pre-registered look · in 50 days
Final read
2026-12-31
PUBLIC kill criterion · in 141 days
Powered at
n ≥ 99
closes before the equity curve can mean anything

From the committee, 2026-08-12

Stance: BASE · regime COMPLICATED

The active base book holds a Bull Call Spread (770.0/782.0). The dominant net exposure is long delta. The primary risk is a down-fade below the 770 strike, breaking the 769 Put Wall and invalidating the structural count. An up-overshoot above 782 is capped by the short call...

— the Portfolio Manager's stated key risk, verbatim

Active Book — Level Map

Where SPY sits right now against the Vision Scout's wave count and every open position's levels — each column is one thesis, the bright line is spot. Levels and spot are as of the last committee pulse, refreshed three times per trading day; the tape is not streaming. See the key below for what each mark means.

Lanes Vision Scout · forecast Directional book Volatility book
Levels spot (now) strike invalidation — thesis dies here entry target / breakeven
Reading pays hurts primary count alternate count (backup)
Paper account · simulated executions marked against real quotes. Levels and counts shown are the committee's, not advice. Not a registered investment adviser. Options carry substantial risk.

Does it know what it knows?

Calibration: when the committee says HIGH confidence, does it win more often than when it says LOW? This is the curve that shows learning — hit rate grouped by the confidence stated before the outcome was known. Cells appear only once they reach 3 resolved calls.

Portfolio Manager — conviction

HIGH
75.0%3W/1L
MEDIUM
39.3%11W/17L

Vision Scout — confidence

HIGH
50.0%6W/6L
MEDIUM
57.1%8W/6L

These are raw hit rates — not yet adjusted for how often a level is reached by chance. Whether the committee's reads actually beat a geometric null is the pre-registered skill-vs-chance verdict, read out 2026-10-01; we don't cite it before.

The Scout's read

The chart-reading desk commits to a direction with a target and an invalidation level, then the market resolves it. Shown: the call and how it scored — not the wave count behind it.

7W / 6L resolved · hit rate 53.8% · 1 open

OPENBULL2026-08-07
WINBULL2026-08-04+0.38%
WINBULL2026-08-04+0.61%
WINBULL2026-08-03+0.50%
LOSSBEAR2026-07-30-0.74%
LOSSBULL2026-07-29-0.01%
WINBEAR2026-07-27+1.02%
WINBULL2026-07-24+0.55%

Hit rate here is raw, not chance-adjusted — a target close to spot is easy to reach. Whether the reads beat the geometric null is the pre-registered verdict (2026-10-01).

The HEW Scout's multi-ticker shadow tracking record — the same wave-reading method run daily across SPY, QQQ and the megacaps, each call scored by first passage. Shadow / paper research, not advice.

The desks

Each specialist desk names a market regime and stakes a falsifiable claim on it. Shown: the latest call and each desk's track record where claims have resolved.

Dealer-Flow DeskCALL WALL PIN · MEDIUM54.1%20W/17L · 2 open
Desk84.6%11W/2L · 0 open
Systematic-Flow DeskPOST-EVENT-VOL-CRUSH · MEDIUM100.0%7W/0L · 2 open
Vol-Supply DeskVOLATILITY TRANSITION · MEDIUM100.0%16W/0L · 1 open

The roads not taken

When the committee decides not to trade, that decision is scored too: did passing dodge a loser, or miss a mover? Most desks never grade their own restraint. This one does.

11 wise / 8 costly · 57.9% of passes dodged a loser

Currently tracking a declined BULL setup.

costly passdeclined BULL2026-08-11+0.31%
costly passdeclined BULL2026-08-05+0.38%
costly passdeclined BULL2026-08-04+3.94%
wise passdeclined BEAR2026-07-31-0.50%
wise passdeclined BULL2026-07-29-0.01%
costly passdeclined BEAR2026-07-29+0.95%