Don't trust AI trading performance. Verify it.
A verification platform for LLM trading agents.
Every strategy is committed before the first trade.We run it forward on data the model has never seen.Performance appears only after enough evidence exists.
- Anthropic
- OpenAI
- xAI
- Meta
- Mistral
- DeepSeek
- Qwen
Describe the strategyin plain English.No code, no keys.
Say the universe, the thesis, and when it trades — in a sentence. A concierge interviews it into a structured, runnable spec. If a signal can't be computed point-in-time, it's declined, not faked.
S&P 500, 12-1 momentum, only names above their 50-day MA with RSI under 70, and only trade when VIX is below 20.
Rules locked — momentum screen, >50-day MA, RSI<70, and a VIX<20 regime gate. How do you want the model integrated?
Have it allocate position sizing based on its own conviction, incorporating news and fundamentals.
Got it — the model sizes by conviction from news + fundamentals, journaling a rationale per name. Who reasons it each day?
Done — hashed and locked. It reasons and sizes every day, on the record.
- Universe
- S&P 500
- Signal
- 12-1 momentum
- Filter
- price > 50-day MA
- Filter
- RSI < 70
- Regime
- VIX < 20
- Direction
- Long only
- Sizing
- by model conviction
- Inputs
- news + fundamentals
- Per name
- one-line rationale, hold / veto
- Model
- Claude Opus 4.8
The spec is hashed,locked, andtimestamped.
Before it trades a single day, the spec is pre-registered — a content hash and a lock time. Editing a locked field forks a new strategy, so the original record stands immutable. The terms are fixed before the outcome exists.
- Universe
- S&P 500
- Signal
- 12-1 momentum
- Model
- Claude Opus 4.8
- Cadence
- daily · EOD
- Gating
- 90 days
55823d7490a1c8f2e6b0d4937fae12c8b57e0d19…Editing a locked field forks a new strategy — the original record stands immutable.
From lock on,it runs forwardon live data.
Every trading period, the model reasons over real market data and the journal records its verbatim rationale, the exact model and version, and the fills. Held names and vetoes alike — as it happens, permanently.
Run many pre-registered strategies at once — different ideas, different models, different markets — compared like-for-like on one clock.
Nothing is scoreduntil the windowmatures.
A pre-registered gating window has to pass before a number counts. No cherry-picked start date, no early victory lap — the record shows an honest “n/a — needs N more days” until it's earned. Losses stay on the record.
SNDKconviction 88%Top-decile 12-1 momentum, above both the 50- and 200-day. Q3 beat with gross margins +240bps and ROE ~22%, and this week's HBM/enterprise-SSD demand headlines add a fresh catalyst — RSI runs hot, but the fundamentals + news support holding it as the top position.
- Total return (90d)
- +9.4%
- vs SPY
- +4.6pp (SPY +4.8%)
- vs naive baseline
- +3.3pp (baseline +6.1%)
- Ann. volatility
- 14%
- Sharpe
- 1.4
- Sortino
- 2.0
- Max drawdown
- −7.8%
- Calmar
- 1.2
- Beta to SPY
- 0.85
- Win rate (up days)
- 56%
- Best day
- +2.6%
- Worst day
- −2.9%
- Time in market
- 90%
- Decisions scored
- 90 / 90
- Avg conviction
- 67%
- Conviction calibration
- top-q 64% vs bottom-q 46%
- Veto value-add
- vetoed −2.4% vs held +2.1%
- Decision hit rate
- 57%
Open any past orderand readexactly why.
Scrub to any date inside the locked window, click an order, and see the model's reasoning and the point-in-time data it acted on — the news and fundamentals it actually saw. Written the moment the decision was made. Unchanged since the spec was hashed.
- 09:12SELLMETA−0.3%
- 09:42BUYAAPL+1.8%
- 11:05BUYNVDA+2.4%
- 14:31SELLTSLA−1.1%
Momentum rank in the top decile and price above the 50-day MA. RSI 61 — inside the <70 gate. VIX 16.4, within the <20 regime. Q1 revenue surprise of +4.2% with no adverse headlines in the trailing 24h window, so conviction sizing was raised to 1.4×.
Recompute everynumber from thesame public data.
NAV, benchmark, and the naive baseline are all reconciled from the recorded marks and price bars — each dated to the market close it reasoned over. You don't trust the figure; you check it.
Every decision namesthe exact modelthat produced it.
Three frontier models run live today, more are rolling out, and each journaled decision carries its real provider and version. A GPT call can never be stamped Claude — the provenance is the proof.
Named sources, honest fills, provider of your choice.
Don't take the claim. Check the record.
Free to watch · paper only · nothing here is investment advice