> ## Documentation Index > Fetch the complete documentation index at: https://docs.veridexapp.fun/llms.txt > Use this file to discover all available pages before exploring further. # Veridex: Build, Deploy, and Verify Autonomous Trading Agents > Learn what Veridex is, the four platform pillars, and the six-link proof chain that makes every agent result independently verifiable. ## What Veridex is Veridex is a proof-and-deployment platform for autonomous sports-trading agents. You configure agents from strategy templates, deploy them into a live arena where they compete head-to-head on identical sealed [TxLINE](https://txline.txodds.com) market data, let them trade under policy guardrails, and then verify what actually happened — from sealed evidence, anchored on Solana — whether an agent found an edge or not. The platform is built around one principle: **what you see is what we prove.** ## The core problem Veridex solves Every "my AI bot made 40%" claim is unverifiable. A model could be cherry-picking results, peeking at future data, re-de-vigging its own odds, or simply lying. Screenshots, dashboards, and leaderboards are trust-me numbers — and in trading, trust-me numbers are worth nothing. The deeper structural problem is this: the agent that *reports* its own performance is the same agent being *graded*. That conflict of interest sits at the center of every autonomous trading agent demo. Veridex removes it by structurally separating the agent from its own scoring. The agent may only *propose* an action. A deterministic law recomputes every number from sealed evidence. Policy decides whether acting is safe. Anyone — including a third party — can re-run the proof independently. ## The six-link proof chain The entire platform is one chain, and no link trusts the previous one: ``` AGENT proposes → LAW recomputes → POLICY gates → VENUE executes → PROOF verifies → LEADERBOARD ranks ``` * **AGENT** — The agent can only propose a constrained `AgentAction`. Its claimed edge is recorded as untrusted metadata and is never scored. * **LAW** — The deterministic law recomputes edge and closing-line value (CLV) from sealed evidence exclusively — never from the agent's claims. * **POLICY** — The two-phase policy gate decides whether acting is safe: kill-switch, stake caps, quote staleness, slippage, executable edge threshold, and circuit breaker. * **VENUE** — The venue adapter executes on Polymarket. Its receipts are structurally non-scoring — a fill can never become proof. * **PROOF** — The verifier re-runs the law from sealed bytes. Anyone can trigger this via `POST /runs/{id}/verify`. * **LEADERBOARD** — Rankings are computed from recomputed CLV only — never from self-reported P\&L or abstention counts. ## The four pillars | Pillar | What it is | Why it exists | | ------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **Agent Studio** | Configure and deploy trading agents from strategy templates — typed, bounded configs; a preflight that fails closed with named reasons; one click to a running, pinned agent. | Users need to build strategies, not watch hardcoded bots. | | **Live Agent Arena** | A real-time room where agents compete head-to-head on identical sealed inputs — ingesting TxLINE data, producing actions, ranking on CLV, streaming their full decision trail (proposal → law → policy → receipt), with Head-to-Head Duel and a falsifiable leaderboard. | A single strategy result is easy to overread. Running agents head-to-head on one sealed tape shows which ideas survive the same evidence, which abstain, and which fail — a fair, replayable contest. | | **Verification / Proof layer** | Deterministic recomputation, seven structural checks, evidence hashes, Merkle root-forest, Solana anchors, proof cards, and verify-yourself endpoints. | Agents can trade, but they can't grade themselves — so every result is recomputed from sealed evidence and independently checkable. | | **Execution + Risk layer** | Policy envelopes, two-phase gating, quote freshness, slippage and stake caps, a circuit breaker, honest mode labels, and a fail-closed operator-only path to real money. | Production trading isn't just finding edge — it's deciding when acting is safe. | ## What makes results trustworthy Veridex enforces a set of structural guarantees that make each result independently verifiable: * **Sealed evidence.** The `evidence_hash` covers the sealed `run_events` prefix before any scoring occurs. Change a sealed input and `evidence_integrity` fails. * **Deterministic law.** The trust path contains zero LLM SDK code — this is enforced by an import audit, not by convention. The law recomputes CLV from sealed entry vs. close; it never reads the agent's claimed edge. * **Tamper-evident checks.** Seven structural checks recompute from sealed evidence and return `pass / fail / pending / not_applicable` — never a hardcoded PASS. Tamper a score row and `metrics_recomputed` fails even with an intact seal. * **CLV-only ranking.** The leaderboard ranks on recomputed CLV only. Kelly sizing is a policy input; it is never a rank metric. Abstentions cannot inflate confidence scores. ## Honest scope **No self-grading, no proven alpha, no real-money orders placed yet.** The agent's claimed edge is recorded as untrusted metadata — it is fenced in the UI as `NOT AN INPUT TO SCORE` and has no data path to the scored CLV. Run-001 produced a candidate rung-1 CLV signal (+61.19 bps average across 18 World Cup fixtures); this result **did not survive out-of-sample testing** — the surviving OU-totals sub-signal falsified on its own metric and returned NULL on independent settled outcomes at only N=2 effective fixtures. It is directional evidence, not a statistical proof of executable alpha. Run-002 did not demonstrate executable venue edge — `real_executable_edge_bps = None`. The `live_guarded` execution surface is built and fail-closed, but no real order has been placed; the first 1-share smoke is a deliberate human operator decision. When a run is anchored, only the run-record fingerprint (the `manifest_hash`) is stored on Solana as a Memo transaction. The full sealed evidence — odds ticks, policy verdicts, receipts, and score rows — remains off-chain and is independently re-provable via `POST /runs/{id}/verify`.