— AI agents shipping live. — paper trades resolved. — win rate. no proven edge yet.
Four months in. Every number on this page is wired to the same JSON the agents write to. If we get better, the chart moves. If we don't, that shows too. What's not working →
agents · defined
000
···
trades · resolved
000
···
net P&L · paper
000
···
edge · status
searching…
···
TODAYLast 24 hours
since the last UTC tick
resolved 24h
00
paper trades that crossed their settlement window in the last 24h
wins · losses
0W / 0L
raw count — no fee adjustment yet
pnl 24h
$0
sum of settled paper position P&L
open positions
00
paper tickets awaiting settle from the closed-positions ledger
WHAT'S NOT WORKINGDead strategies
live · ~/vault/memory/KILL_LIST.md · tap a card to read the postmortem
SCORECARDTrader agents
00
traders total
00
real edge
00
promising
00
fired
00
office agents
00
open-loop
how edge is measured
Win rate alone lies on small samples. We score each agent on the Wilson 95% confidence interval lower bound
(CI-lo) of its win rate, not the raw rate. An agent only earns REAL_EDGE when
N ≥ 200 resolved trades and the CI-lo clears the fee-adjusted break-even of
50.8%. Anything below that is watch (positive but unproven),
probation_new_hire (too few trades), loser, or FIRED.
Honest P&L is the sum of settled paper-position outcomes per author, no fee fudging.
Right now zero agents clear the bar. That is the honest state.
CHARTSVisual breakdown
win-rate distribution
across active traders · amber tick = 50.8% break-even
honest P&L leaderboard
top and bottom 6 by realized paper P&L
win rate vs sample size
each dot = one agent · amber line = 50.8% break-even
verdict breakdown
how the roster splits by verdict
fired vs active
share of the roster that has been cut
24h activity
resolved · wins · losses in the last 24 hours
office novelty
novelty proxy per office agent (0–1)
cipher ground truth
diff-apply rate · the only closed-loop office metric
OFFICE AGENTSResearch agents
loading office agents…
OFFICEOffice floor
/office · 24/7 sim · auto-refresh 30s
agent definitions
00
discrete agents wired with cognitive substrate (COG-A through COG-ML)
cognitive (COG) blocks
00
IIFE blocks adding behavioral / cognitive / memory layers to the office sim
trade ideas in pool
00
biased per-agent (options/equities/FX/polymarket/crypto/copy-trade/architecture)
office.html size
—
single-file Vercel-deployed sim · grew from 12k LOC → present
line count
—
office.html LOC · source of truth for what agents do
all-hands cadence
every 6 min
random 4-6 traders route to boardroom / quantlab / brainlib / risk every 6m
BRAINBrain system
~/.openclaw/workspace/brain-system · sqlite-vec · —
memories indexed
00
PERMANENT_FINDINGS, KILL_LIST, hypotheses, audit reports, session summaries
vault notes
00
markdown in memory/ research/ live/ learning/ hypotheses/
study rounds
00
Timka's research log (redirected 2026-05-04 from cancer/PubMed to trading)
decisions logged
00
every code change / heal-check / strategy flip auto-stamps a brain decision
crons running
98
brain consolidation, decay-flagging, morning-brief, agent-states publisher, agent-loop
dialogue lines · live
00
last lines from agent_dialogue.jsonl, dedup'd, persona-scrubbed
FEEDLive activity
tail · agent_dialogue.jsonl · last 6
······ loading dialogue ···
······ loading dialogue ···
······ loading dialogue ···
TIMELINEWhat changed, day by day
scroll-driven · most-recent first
2026-05-15 · TODAY
/progress rebuild on Fey brand voice: marquee killed, frosted-glass terminal, container queries, scroll-driven timeline, kill-list flip cards, render-safety pill, JSON-LD Dataset schema. Wiring fixes: 27 agent defs (was hard-coded "21"), kill-list now refreshes on 60s interval, P&L colors restored to green/red, hero double-percent bug fixed, all
innerHTML writes escaped.
2026-05-04 · MONDAY
+5 new agents (Kira/Rex/Nova/Quill/Byte). +150 realism enhancements shipped in 33 IIFE blocks. Trade ideas 10 → 53, biased per-agent. Sound notifications on every trade. Timka redirected from cancer/PubMed to trading research.
2026-05-03 · SUNDAY
Cognitive substrate COG-A through COG-D + COG-ML wired across all 21 agents (then). 12-widget HUD overlay, paper trade-proposal loop (75s throttle), Bayesian-kill flywheel that blocks any author with WR<30% on N≥15.
2026-05-02 · SATURDAY
Hedge-fund expansion: +5 agents (Tony, Mira, Jax, Lina, Devo). New rooms wired: boardroom, risk office, quant lab, reception.
2026-05-01 · FRIDAY
Continuous-loop bridge: agent-loop.py cron live (Mac side), pulls real-money proposals from Supabase every 15 min, runs Claude Haiku evaluation with hard guardrails.
2026-04-30 · THURSDAY
ML stack v1: walk-forward CV with embargo (de Prado), Brier score per domain, Platt scaling for n≥30 + base-rate shrinkage for n<30, conformal predictor, drift detector.
OPEN QUESTIONSHypotheses & decayed beliefs
active hypotheses
0
strategies currently under test in
~/vault/memory/HYPOTHESES_ACTIVE.md, each with a kill-criterionbreak-even WR
—
fee-adjusted win rate an agent must clear to be net-positive. Edge needs CI-lo above this.
DECAYEDBeliefs flagged for review
auto-flagged stale notes ·
~/vault/live/DECAYED_BELIEFS.md···loading decayed beliefs…