# FAQs

### General

**Q : What exactly is SlinkyLayer?**\
**A :** SlinkyLayer is an infrastructure platform for training, validating, and serving live AI models that trade crypto markets. A smart-contract registry anchors model metadata, stakes, audit proofs, and reward emissions.

**Q : Is SlinkyLayer a broker or exchange?**\
**A :** No. The platform emits signed trading signals. Executing those signals is left to your own bot, vault, or desk.

**Q : Model vs Policy – why two words?**\
**A :** Every training run produces a *model* in the broad sense of “learned weights.” When the algorithm is reinforcement learning, that model is more specifically a *policy -* a mapping from market state directly to an action, without an intermediate price forecast. All policies are models, but future releases will also host non-policy models (e.g., vision encoders). So we use model for user-facing language and policy when we need the exact technical term.

**Q : Can I train computer-vision models such as YOLO today?**\
**A :** Not yet. The current release focuses on reinforcement-learning policies. The pipeline is model-agnostic, so vision or hybrid models are on the roadmap.

**Q : What is Slinky Consensus?**\
**A :** It is the on-chain mechanism that ranks published models and releases new $SLINKY. A model’s score equals its 7-day forward PnL multiplied by a logarithmic stake term. Rewards are split fifty-fifty between the creator and the stakers who backed the model.

**Q : Do I have to publish my model?**\
**A :** No. Private runs stay in your encrypted storage, produce signals only for your API key, and never touch the chain.

***

### Model Training

**Q : Which learning algorithms are available?**\
**A :** PPO, A2C, SAC, TD3, and DDPG, each with a configurable MLP backbone (layer sizes and activation: Tanh or ReLU).

**Q : How large should my training window be?**\
**A :** The wizard presets cover common horizons. Example: a 60-minute chart trains on 24 months and tests on the next 24 months. Custom dates are allowed.

**Q : What is a timestep?**\
**A :** One environment step equals one candle at the chosen timeframe. Total timesteps is the budget of environment steps the learner will sample before stopping.

***

### Signals

**Q : What actions can a model output?**\
**A :** Exactly 3 tokens: `LONG_100`, `SHORT_100`, and `FLAT_0` 

**Q : Which transports are supported?**\
**A :** WebSocket, HTTPS webhook, Kafka, and a Historical REST API. All channels send the same signed JSON.

**Q : What latency should I expect?**\
**A :** Gateway publishes within two seconds of candle close. Auditors flag signals that arrive after five seconds.

***

### Marketplace and Staking

**Q : How is the Slinky Score calculated?**\
**A :** Detailed explanation is provided here -> [Broken link](https://docs.slinky.network/understanding-slinkylayer/broken-reference "mention")

**Q : How often are rewards paid?**\
**A :** Every seven-day epoch the protocol mints new $SLINKY. For any public model, 50 % goes to the creator, 50% to stakers pro-rata by stake-days.

**Q : Why use a logarithm on stake?**\
**A :**The log term gives diminishing returns, preventing a single large holder from dominating rankings while still rewarding broad community support.

***

### Audit and Security

**Q : What stops someone from replaying tomorrow’s prices?**\
**A :** Auditor nodes compare each signal’s timestamp to reference candle close timestamps and compute a segment hash. A segment scores only after three identical hashes reach quorum.

**Q : Can I verify a signal myself?**\
**A :** Yes. Fetch the creator’s on-chain public key, then verify the Ed25519 signature over the raw JSON payload. Recalculate reward with a public price feed to match the auditor math.

**Q : What if two models look identical?**\
**A :** On publish the weights are hashed. If that hash already exists the transaction reverts, preventing duplicate uploads.