Read-only · real-money trading paused 2026-05-06.
The live-money bot was burning ~$190/mo in API spend to manage a $6.39 Kalshi balance — net structural loss every month, so we shut it off.
The paper tournament with the same intelligence pipeline is still live at
/dashboard →
trenchsignals.iobeta● live strategy
Autonomous conflict intelligence
Loading…-
PausedThe live-money bot was shut down on 2026-05-06.
It was costing ~$190/mo to manage a $6.39 balance with a $3 base bet size, generating
~$2.30 of trading P&L over 23 trades. The math wasn't there. The
paper tournament at /dashboard
now covers the same strategy at $1000 paper bankroll, with three competing variants —
more meaningful data per unit time, no real-money drag.
Historical trades and signals on this page remain intact for reference.
Realized P&L
-
-
Win Rate
-
-
Open Positions
-
-
Best / Worst
-
-
Avg Return
-
-
Last Signal
-
-
📊 Open Positions0
⏳ Exits triggered — awaiting Claude review
Market
Side
Entry
Now
P&L
Held
No open positions
🗂 Recent Trades0
Market
Side
Price
P&L
Hold
Status
Conf
No trade activity yet
🎯 Market Assessments
Claude vs market - latest signal
-
Market
Theater
Claude ¢
Market ¢ (live)
Spread
Net Edge ↓
Δ Cycle
No assessment data yet - waiting for next signal cycle
1 / 1
⚡ Live Assessment
-
Waiting for first signal…
📜 Signal History0
TimeSignalEsc. ProbConf
No signals yet
1 / 1
Claude's escalation probability across all signal cycles. Dot size = confidence; color = direction (cyan = ESCALATE, blue = DEESCALATE, grey = SKIP); ring = trade placed. Hover dots to inspect.
📈 Escalation Probability History
last 20 signal cycles — hover to inspect
⚡ Recent Whale TradesPolymarket - large orders in last 24h
Loading…
Tracked Markets
All geo markets currently assessed by Claude — live probability estimate vs market price. Updates with each signal cycle.
These are Claude's probability estimates vs the live market price — a divergence, not a demonstrated trading edge. Edge shown is net of estimated fees; capital on long-dated markets is locked until resolution. Skill isn't yet measurable (only ~1 distinct market has resolved). See calibration.
Waiting for signal data…
Financial tells · leading indicators
Oil & safe-havens often move before conflict markets reprice. A leading tell, not a trade signal.
When a market settles is the timing edge — soonest-first, long-dated demoted.
Market
Claude
Market
Net Edge ↓
Resolves
Signal
Claude History
-
Loading news feed…
Strategy Calibration · final real-money tape
Probability accuracy, edge capture, P&L attribution, and threshold sensitivity from the live-money bot's run.
Real-money trading was paused 2026-05-06 — the numbers below are the closing record, not a live tape.
The live paper-trading calibration is on the
paper-trading dashboard.
Cycle Outcomes (last 24h)
Every entry-decision tick exits via exactly one path. Where are cycles dying?
Only traded means a position was placed.
Loading…
Tuning Recommendations · auto-triggered
Strategy changes the data has earned. Each recommendation specifies a fire condition (n≥30,
ROI gap, etc.) — they appear automatically as the data crosses the threshold. No manual revisit.
Strategy Decisions Log
Append-only record of changes driven by what the calibration data is telling us. Honest about
findings, including the ones that say "your strategy is leaking money here."
2026-05-05
Raised confidence threshold 0.72 → 0.74
Backtest showed the 0.76–0.80 bucket winning 75% with +42% ROI vs the 0.72–0.76 bucket bleeding −7%.
Sample is small (n=13 signal trades) so bumped to 0.74 rather than 0.76 — provisional change,
revisit at n=30+. Both real bot and shadow updated.
2026-05-05
Fixed orphan flow leak (−$2.86 across 7 trades)
Orphan assessor (runs 2 min after restart on positions with no thesis) was biased toward EXIT when uncertain —
Claude was being told "if there's no clear reason to hold, exit", which on a freshly-restored
position with no original context almost always meant dumping. Rewrote system prompt: default HOLD with a
fresh thesis, only EXIT if the position is >15% past stop-loss with no recovery, the market has resolved,
or price is at terminal value. Also now captures Claude's actual reasoning in exit_reason
instead of a bare orphan_assessment_exit label, so the diary can render real lessons.
2026-05-05
Honesty rails on every public surface
Loss cards in the diary, weekly digest, and tweet templates now lead with
"Why I was wrong" — Claude's own exit reasoning, plus a structural failure-mode label
(high-confidence miss / thin signal / late entry / thesis invalidated). Weekly digest puts
"What I got wrong this week" above "What worked." The point: anyone can claim wins; the moat is
owning specific, classifiable losses.
2026-05-05
Risk circuit-breaker: --max-session-loss
New flag halts new entries (existing position management continues) when cumulative session
P&L drops below a configurable loss. Latched until restart. Real bot: −$5 cap.
Shadow: −$100. Defensive control before any future move from paper to real-money sizing.
2026-05-05
Cycle-outcome instrumentation
Every entry-decision tick now emits exactly one structured line: Cycle outcome:
<category>. Categories cover every early-return path
(too_few_new_items, source_diversity_gate, analyzer_timeout,
signal_skip, session_loss_cap, confidence_too_low,
cap_reached, tier_gate_75/80, no_candidates_above_min_edge,
traded). Tomorrow we'll have a real histogram of which gate is killing
the 83% of cycles that don't produce trades.
2026-05-05
Theater concentration audit
11 of 13 signal trades have been on Iran. Audited the market universe — Kalshi
currently exposes 81 Taiwan markets, 7 Iran, 3 Ukraine/Russia.
Trade-side filters and theater tagging both pass through Taiwan correctly,
so the gap is upstream: the news + signal pipeline is Iran-heavy, so Claude
rarely generates directional ESCALATE/DEESCALATE signals for Taiwan. Strategy
work to fix this needs a deliberate prompt change + data, not an ad-hoc tune. Logging
the gap and watching whether Claude-driven theater signals surface organically as
Taiwan news evolves.
2026-05-05
/dashboard leads, /livestrategy follows
While paper-trading is the focus, all new instrumentation
(Trench-right-now hero, live whale flow, full cycle-outcome funnel,
big-trades feed) ships to the shadow dashboard at
/dashboard
first. This live-money page mirrors once the real bot has more
capital and activity worth showcasing. Honesty rail applied to
the dashboards themselves: don't promise polish you're not
actively investing in.
2026-05-06
Live-money bot paused — cost / value math wasn't there
The live-money bot was costing ~$190/mo in Anthropic API spend to manage a
$6.39 Kalshi balance with $3 base bet sizing. Real trading P&L generated:
~$2.30 over 23 closed trades (+2.2% ROI) — net of API cost, the bot was losing
money every month at a structural level. Decision: shut down the live-money bot,
keep the read-only API alive so this page continues to serve historical data,
and let the existing 2 open Kalshi positions resolve naturally. The paper
tournament at
/dashboard
already runs the same config (Baseline variant: conf=0.74, TP=20%, SL=10%) at
$1000 paper bankroll — strictly more data, strictly less drag.
Cleanly reversible if/when we want to flip back to real money.
Still on the radar:
TP/SL bracket sizes
(deferred — brackets fire too rarely; Claude exits dominate, need ~30 more trades),
deliberate Taiwan expansion
(open question: tune analyzer prompt to surface non-Iran theaters more aggressively),
return to live-money trading
(waits for paper tournament to identify which variant has alpha — then size up real money against the winner).
Loading calibration data…
Equity Curve
Cumulative P&L per closed trade. Signal-only trades (confidence > 0) shown in green; orphan/restored trades in grey.
Escalation Probability Timeline
Claude's escalation probability (cyan) vs baseline (dashed blue) across all signal cycles. Dot size = confidence; dot color = direction (cyan = ESCALATE, blue = DEESCALATE, grey = SKIP). Ring = trade placed.
Market Probability View
Scatter: Claude's estimated YES probability vs market price for all evaluations.
Points above the diagonal = Claude sees positive edge. Color = exchange; ring = trade placed.
Select a market below to see its probability history over time.
Confidence Threshold Sensitivity
›
How P&L changes if the confidence cutoff were raised or lowered — using only trades we actually took.
Min Confidence
Trades
Win Rate
Total P&L
ROI
P&L by Theater
›
Theater
Trades
Win Rate
Total P&L
ROI
P&L by Signal Direction
›
Direction
Trades
Win Rate
Total P&L
ROI
Exit Reason Breakdown
›
Exit Type
Count
Avg P&L
Total P&L
P&L by Confidence Band
›
Confidence Range
Trades
Win Rate
Total P&L
ROI
P&L by Confluence Band · graph-grounded
›
Sliced by the number of distinct intel source-types touching the trade's primary entities at trade time. Tests whether multi-source corroboration actually predicts outcomes.
Confluence Band
Trades
Win Rate
Total P&L
ROI
P&L by Primary Entity · graph-grounded
›
Replaces theater string-matching with the ontology-tagged primary entity for each trade.
Primary Entity
Trades
Win Rate
Total P&L
ROI
Probability Calibration Curve
›
Predicted probability vs actual resolution rate. A perfectly calibrated model sits on the diagonal.
Prob Bucket
N Evaluated
Predicted Mid
Actual Win Rate
Calibration Error
Live Configuration
Current bot settings
Strategy: Global conflict — Middle East · Ukraine · Taiwan · N.Korea
Exchange: Kalshi (US-regulated)
Signal model: claude-sonnet-4-6
Review model: claude-haiku-4-5
Poll interval: 10 minutes
Position management: Continuous (every 30s)
Trades per cycle: One — highest-edge candidate only
Min confidence: 0.72
Take-profit: +20¢ price move OR 50% return on entry cost (dynamic — Claude can extend)
Stop-loss: −10¢ price move (dynamic — Claude can override)
Min entry price: 5¢ per contract
Escalation decay: Half-life ≈ 6.4 h, mean-reverts to 25% between signals
Market Regime
Auto-classified from rolling escalation probability, urgency distribution, and vol regime across recent signals
—
Avg Escalation
—
Esc Trend
—
High Urgency
—
Vol Regime
—
Edge Distribution
Claude probability − market price across all scored markets in signal history
Mean Edge
—
Median Edge
—
+Edge Markets
—
Samples
—
Equity Curve & Drawdown
Cumulative P&L across closed trades — red shading = drawdown from peak
Total P&L
—
Max Drawdown
—
Win Rate
—
Closed Trades
—
Kelly Position Sizing
Optimal bet per market based on edge vs market price. Showing ½ Kelly (conservative). $500 estimated bankroll.
Market
Side
Claude / Mkt
Edge
Full Kelly
½ Kelly $
vs $3 base
No markets with meaningful positive edge in current signal
YES: Kelly = (claude − market) ÷ (1 − market) · NO: Kelly = (market − claude) ÷ market · Capped at 25% · ½ Kelly applied for safety
Knowledge graph
The graph now has its own home.
We promoted the interactive ontology explorer to a standalone page so it can use the full viewport and be linked / shared on its own URL. Same Cytoscape view, same data, same click-to-explore, just bigger.
We promoted the platform write-up to its own URL so it has proper marketing-page typography,
social-share metadata, and a more readable layout for prose. Same content, more depth.