> ## Documentation Index > Fetch the complete documentation index at: https://docs.voight.xyz/llms.txt > Use this file to discover all available pages before exploring further. # TxLINE sports data > Verifiable live sports odds inside Prediction Markets agents: real-time demargined probabilities, a deterministic sharp-movement detector, on-chain pick attestation, and a 7-for-7 development run. **TxLINE** is TxODDS' Solana-anchored sports data feed: fixtures, live odds, and scores, with on-chain validation proofs. [Prediction Markets](/agents/templates#prediction-markets) agents plug into it to read the market in real time, detect sharp movement deterministically, and anchor their own picks on Solana. Verifiable data in, verifiable predictions out. Demargined implied probabilities from a Solana-anchored feed. No margin math, no distortion. A calibrated z-score script detects sharp moves. The model never does market math. Committed picks are hashed and anchored on Solana, tamper-evident before the event settles. TxLINE is in **preview** on selected Prediction Markets agents, graduating from our entry to the TxLINE World Cup hackathon. The full integration is open source: [voight-txline-agent](https://github.com/Voightxyz/voight-txline-agent). ## Track record **Development run (July 9-15, 2026): 7 for 7.** While this integration was being built, the prediction agent called seven consecutive winners across World Cup 2026 knockouts and MLB moneylines: | # | Event | The agent's call | Result | | :-: | --------------------------------------- | ---------------------------------------------- | :----: | | 1 | World Cup QF · Spain vs Belgium | Spain to win | ✓ | | 2 | World Cup QF · England vs Norway | England to win | ✓ | | 3 | World Cup QF · Argentina vs Switzerland | Argentina to win | ✓ | | 4 | MLB · Yankees moneyline | Yankees to win | ✓ | | 5 | MLB · Red Sox moneyline | Red Sox to win | ✓ | | 6 | World Cup SF · Spain vs France | **Spain to win, against a 28.5% market price** | ✓ | | 7 | World Cup SF · England vs Argentina | Argentina to win | ✓ | The standout is the semifinal: the market gave **Spain a 28.5% implied probability against France**, and the agent took Spain anyway, flagging the referee assignment pre-match as a decisive factor. It decided the opening goal. Honest numbers, honestly sourced: these picks were logged in the agent's persistent ledger **before each event settled**, during the development run. They predate on-chain anchoring, which shipped mid-July; from then on, committed picks anchor on Solana (below). A hot streak is not a guarantee: the agent gives research, not financial advice. ## How it fits together ```mermaid theme={null} flowchart LR A["TxLINE feed
(Solana mainnet)"] -->|"demargined Pct[] + scores"| B["Agent container
(private, per-agent)"] B --> C{"sharp_detect.py
(deterministic)"} C -->|alert| D["Web research:
the why"] D --> E["Telegram alert
+ market card"] C -->|quiet| F["One short line"] B -->|committed pick| G["Pick ledger"] G --> H["Solana memo
+ Solscan card"] ``` ## What the agent reads * **Fixtures.** Upcoming and in-play matches, participants, start times, game state. * **Demargined odds.** TxLINE's StablePrice feed delivers implied probabilities with the bookmaker margin already removed: a value of `58.2` means the market genuinely implies 58.2%. No margin math, no distortion. * **Live scores and updates.** Snapshot plus a live update stream while a match runs. * **The markets that matter.** Match result (1X2), over/under goals, and Asian handicap. Derivative and period markets are excluded as noise. | Endpoint | Use | | ---------------------------------------------------- | -------------------------------------------- | | `POST /auth/guest/start` | Guest JWT, auto-renewed on 401 | | On-chain subscription + `POST /api/token/activate` | API token minted from the Solana transaction | | `GET /api/fixtures/snapshot` | Discover active fixtures | | `GET /api/odds/snapshot/{fixtureId}` | Current odds state | | `GET /api/odds/updates/{fixtureId}` | Live updates for detection scans | | `GET /api/odds/updates/{epochDay}/{hour}/{interval}` | Historical replay and backtests | | `GET /api/scores/snapshot\|updates/{fixtureId}` | Score context for alerts | | `GET /api/odds/stream` (SSE) | Live capture for calibration recordings | Network: mainnet (`txline.txodds.com`), program `9ExbZjAapQww1vfcisDmrngPinHTEfpjYRWMunJgcKaA`. ## The Sharp Movement Detector The core design decision, learned in production: **the model never does market math.** A deterministic script builds implied-probability series from the feed and applies calibrated thresholds. The agent narrates, investigates, and delivers; it never estimates a movement itself. Detection you can audit, explanation you can read. For each series *(fixture, market type, outcome)* built from the demargined `Pct[]` feed: 1. **Jump.** `Δ = |Pct(t) − Pct(t − 120s)|`: the move over a 2-minute window. 2. **Noise baseline.** `σ` = the standard deviation of successive moves over the 600 seconds before the window. 3. **Alert** fires only when `Δ ≥ 5 pts` **and** `Δ ≥ 3σ`: an absolute floor so thin markets can't alert on noise, and a relative test so calm markets alert on genuinely abnormal moves. 4. **Debounce.** A series stays quiet for 300 seconds after alerting. | Mode | What it does | | ---------- | -------------------------------------------------------------------------------- | | **Scan** | Sweeps every active fixture for sharp moves. Runs on the agent's recurring task. | | **Watch** | A 60-second live loop on one match, for in-play monitoring. | | **Replay** | Backtests a recorded stream. Same input, same alerts, byte for byte. | ### Calibrated on a real feed The live SSE stream was recorded during the England vs Argentina semifinal. 1,773 odds events, 32 series, and the detector fired exactly twice: the over/under 3.5 goals market repricing as regulation ended 1-1. Zero false positives across the rest of the match. ```text theme={null} $ python3 detector/sharp_detect.py replay captures/odds-stream-semifinal-20260715.jsonl eventsScanned: 1,773 · seriesBuilt: 32 · alerts: 2 ▲ ALERT over/under 3.5 goals · under 84.7% → 95.4% (+10.7 pts, z=17.8) ▲ ALERT over/under 3.5 goals · over 15.3% → 4.6% (−10.7 pts, z=17.9) ``` ## The alert flow On schedule, the agent pulls active fixtures and their live odds updates from TxLINE. `sharp_detect.py` applies the calibrated thresholds. Empty alerts means nothing abnormal, and the agent says exactly that. On an alert, it investigates the likely cause with live web search (a goal, a red card, lineup news) so the alert arrives explained, not just flagged. A Telegram message with a market card: market, outcome, the probability move in points, and its z-score. No alert, no noise: a silent market produces one short line. ## Picks anchored on Solana TxODDS anchors the sports data on Solana to make it verifiable. Voight closes the circle: **the agent's predictions get anchored too.** Every pick the agent commits to is logged in its persistent ledger. Ask it to **anchor the pick on-chain** and it: 1. Takes the pick's **immutable core** (`id`, `date`, `market`, `pick`, `prob`) as canonical JSON (sorted keys, no whitespace). 2. Hashes it with **SHA-256**. 3. Records a **Solana memo transaction** carrying the hash plus a human-readable summary, so the prediction reads directly on any block explorer. Fee: \~0.000005 SOL. 4. Returns a card in chat with a **Verify on Solscan** link. A real memo from the third-place playoff: ```text theme={null} voight-pick: | France to win @58% · World Cup 2026 Third-Place: France vs England (2026-07-17) ``` Because only the immutable core is hashed, verification is **settlement-stable**: the hash still checks out after the pick settles and its status changes. And it's independent, from the public repo, with no trust in Voight required: ```bash theme={null} node attest.mjs verify ``` ## Driving it from chat | You say | The agent does | | -------------------------------------------------- | -------------------------------------------------------------------------------------------------------------- | | "What fixtures are coming up?" | Pulls the live fixture list from TxLINE. | | "Any sharp movement?" | Runs the detector and reports alerts, or says the market is quiet. | | "Research Argentina vs Spain and commit to a pick" | Full-picture research (form, injuries, referee, weather, head-to-head), its own probability, a committed pick. | | "Anchor that pick on-chain" | Logs the pick, attests it on Solana, and returns the Verify on Solscan card. | Be direct: "anchor this pick on-chain" triggers the attestation flow. Asking for "a verifiable proof" in the middle of another request can send the agent down a different path. ## Verify it yourself The client, the detector, the attestation tool, and the real captured semifinal feed (755 KB of live odds) are public in [voight-txline-agent](https://github.com/Voightxyz/voight-txline-agent). Replay the capture and you get the same two alerts the agent got, deterministically. Two minutes, no keys needed. ## Next * [Templates](/agents/templates#prediction-markets): the Prediction Markets template * [Quickstart](/agents/quickstart): deploy an agent in about a minute