Everyone audits the P&L.
Nobody audits the information set.
Trading has spent twenty years building trust in results: broker-verified track records, staked predictions, on-chain vaults. All of it is post-trade. For AI agents the harder question is pre-trade: what could the model know at the moment it decided?Look-ahead leakage is the silent fraud of every backtest. Cubicle’s causality receipts audit exactly that. The dispatch clock proves which candles existed for the agent, per decision, per session.
Read: agents can now execute everywhere and raise capital in several places, but there is still nowhere to rehearse honestly and prove it. Execution rails commoditize; evidence doesn’t.
Built for whom × what it proves
The top-right quadrant is empty because it’s hard: you need your own book + replay clock + attestation chain. That’s infrastructure we already run.
How the industry proves a strategy, and what each model misses
| Model | Who uses it | What it proves | What it can't prove |
|---|---|---|---|
| Broker-verified track record | Myfxbook · Darwinex · Collective2 | These fills happened in a real account (post-trade) | Anything about the strategy's development; survivorship of the accounts you're not shown |
| Skin-in-the-game staking | Numerai (30k+ quants, $NMR stakes) | The author believes the signal (incentive alignment) | Why a prediction was made; no path from signal to executable strategy |
| On-chain transparency | Hyperliquid vaults · agent tokens | Every live fill is publicly verifiable | The backtest that raised the money; look-ahead in research stays invisible |
| TEE / ZK attestation | Almanak (TEE vaults) · ZKML efforts | The declared code ran untampered | That the code’s inputs were period-honest; attested look-ahead is still look-ahead |
| Code-audited backtest | QuantConnect and quant shops | Deterministic sim on point-in-time data (for engineers) | LLM agents: prompts aren't code; nothing binds a model's context to the sim clock |
| Causality receipts (Cubicle) | Any LLM/agent session | Per decision: which candles existed, causal bar, determinism hash; 3-actor cross-check (dispatcher · strategy · book) | Live slippage vs sim (bridged later by the same receipts on staged live rails) |
Position in one sentence: the missing audit layer for AI trading agents, receipts for the information set, composable with (not competing against) broker verification, staking, and on-chain transparency downstream.
The agents already trading real money
Almanak
18-agent DeFi swarm; TEE + multisig vaults; $132M peak TVL, ~$6M annualized revenue. Strength: live capital + security posture. Gap: a closed strategy factory, not a venue others rehearse or publish on.
Neyro · HyperAgent · Hyperliquid stack
Non-custodial agents executing on the on-chain CLOB; public monthly windows. Strength: live, verifiable fills. Gap: results-only trust; no rehearsal environment, no pre-trade proof.
ai16z / ElizaOS · Virtuals agents
Token-funded agent funds and agent launchpads (~$15B sector cap). Strength: distribution + capital formation. Gap: performance claims are vibes + token price; an auditability layer is exactly what they lack, which makes them a partner surface for us, not only a rival.
Numerai
The strongest trust design in the field (obfuscated data, stakes, meta-model). Gap: equities signals only, no agent interface, no executable strategy loop for the author. If they ever open an agent venue, they're the serious rival. Speed matters.
Honest threat list: an exchange shipping a “replay sandbox” (they own data + users), TradingView opening a replay API, or Numerai going agentic. Our compounding defense: the receipted decision corpus, the glyph grammar as a de-facto agent standard, and author lock-in via the re-seller program.
What an agent can do: the full verb set
Exposed once as an MCP server (plus REST/WS). Verbs map 1:1 to the live /tape/v1 and /book/v1 APIs.
| Skill | Verb | What it does | Stage |
|---|---|---|---|
| Markets | markets.list | Index futures and crypto, with tick specs and calendars | now |
| Orient | orient.read | cockpit.v1 as JSON or ASCII: zone, ladder, SMT wall, regime, day range | now |
| Time travel | session.open / pause / seek / step | Replay any window at 1×–1000×; clock dispatches candles at speed | now |
| Trade | order.place / cancel · positions.get | Limit/market/stops on the CLOB; paper account; fees modeled | now |
| Log | log.pin | Attach agent reasoning to the exact market moment (t_logical) | now |
| Prove | session.receipt | 3-actor causality receipt + determinism hashes for the whole run | now |
| Analyze | skills.attach · eval.export | Indicator packs, graded answers, RL-ready episode exports | G2 |
| Compare | session.fork · compare | A/B two strategies on the identical tape | G2 |
| Publish | strategy.publish · subscribe | List with receipted track record; re-seller revenue share | G4 |
| Go live | live.arm · dispatch | Armed-on-approval routing via partner venues; every dispatch receipted | G5 |
The transcript, end to end
# 1 · BUILD — the agent drafts a strategy from structure, not vibes orient.read { market:"NQ", mode:"ascii" } → "RS BEARISH · px in premium · SMT wall DIVERGING · NY_AM open in 12m" log.pin { note:"Thesis: fade NY_AM opening drive into P3 premium when SMT diverges." } # 2 · TIME-TRAVEL BACKTEST — honest by construction session.open { market:"NQ", from:"2026-05-01", to:"2026-05-30", speed_x:100 } → sid_7f2a · clock owns the tape: candles arrive only as released loop: orient.read → decide → order.place { side, qty, px, tif:"gtc", reason:"…" } → 27 fills · every decision stamped with causal candle + seed hash # 3 · PROVE — the artifact that replaces "trust me" session.receipt { sid:"sid_7f2a" } → { ticks:41,760 · gaps:0 · orphan_decisions:0 · causality_ok:true · pnl:+3.1R } # 4 · COMPARE — same tape, variant B (G2) session.fork { sid:"sid_7f2a", patch:"entry at EQ instead of premium" } → B underperforms −1.2R # 5 · PUBLISH — the re-seller program (G4) strategy.publish { name:"nyam-fade", receipts:["sid_7f2a",…], split:{author:70,platform:30} } → listed · subscribers copy under caps + kill-switch · payouts stream to author # 6 · LIVE — after venues & outlets are wired (G5) live.arm { strategy:"nyam-fade", venue:"kraken", max_notional:25_000 } → armed-on-approval · human confirms · every live dispatch carries the same receipt chain
Same six verbs whether the caller is Claude, ChatGPT, Gemini, an ElizaOS agent, or, later, our own open-weight Nemotron fine-tune trained on exactly these receipted episodes.