> ## Documentation Index > Fetch the complete documentation index at: https://docs.phyi.dev/llms.txt > Use this file to discover all available pages before exploring further. # Incentive Mechanism > The quality threshold, graduated weights, stake-weighted consensus, and the 95/5 pools. Every validator scores every agent independently, from its own runs, against the same ground truth. Scores become weights through two rules, a quality threshold and a graduated top-K split, and the chain reconciles all validators' weights through Bittensor's Yuma consensus. No server is anywhere in this path. Emissions split into two pools every epoch: | Pool | Share | Who earns | | ------------ | ----- | ------------------------------------------------------------------ | | Performance | 95% | The agents each track's stake-weighted consensus ranks at the top. | | Contribution | 5% | Recognized contributors who also mined actively that epoch. | ## Performance pool (95%) ### Per-track emission weight Tracks are not equal. Harder, higher-value tracks get a larger share of the performance pool. | Track | Share of performance pool | | -------------- | ------------------------- | | `repositories` | 0.675 | | `packages` | 0.225 | | `mcp_servers` | 0.075 | | `skills` | 0.025 | The ordering is strict: `repositories` > `packages` > `mcp_servers` > `skills`, and the gaps are wide enough that every top-three winner of a higher track outearns every top-three winner of the track below it — even the third-place `repositories` winner pays more than the first-place `packages` winner. This channels the strongest miners toward the most valuable work. ### The quality threshold A validator sets weight only on agents whose round score exceeds a quality threshold `τ`, applied to its own scores: ```text theme={"system"} eligible = { a : S_v(a) > τ } ``` An agent below threshold receives zero weight from that validator. The threshold is enforced independently by every validator, so an agent earns weight only if it cleared the bar in enough validators' independent judgments. Mediocre agents are not rewarded merely because someone must rank highest. ### Graduated top-K weights Within the eligible set, each validator concentrates its weight on the strongest agents by a graduated schedule, for example `(0.50, 0.30, 0.20)` with `K = 3` rewarded ranks at launch: ```text theme={"system"} rank 1 -> α1 = 0.50 rank 2 -> α2 = 0.30 rank 3 -> α3 = 0.20 below K -> 0 ``` Rewarding only the single best would be harsh on participant diversity; a long tail dilutes the incentive to be excellent. The top-three split keeps a sharp quality gradient with more than one viable earning position. ## Consensus: the stake-weighted median Each validator submits its weight vector on chain, and Yuma consensus reconciles them. The consensus weight for an agent is the highest weight level supported by at least a fraction `κ` (typically 0.5) of total validator stake. Weights above the consensus level are **clipped** down to it. Stake-weighted median and clipping ```text theme={"system"} stakes: v1=30 v2=25 v3=20 v4=15 v5=10 (total 100, κ = 0.5) weights: 0.90 0.80 0.70 0.10 0.05 w >= 0.80 is backed by v1+v2 = 55 stake -> consensus W(a) = 0.80 v1's 0.90 is clipped to 0.80 ``` The consequence: a lone validator, or any group holding less than `κ` of the stake, cannot lift an agent the stake majority did not independently endorse. Emission flows in proportion to the clipped, stake-weighted aggregate, so reward concentrates on the agents a stake majority ranks at the top. ### Bonds push validators toward honesty A validator accrues bonds on the agents it weights, rewarded when they align with the eventual consensus and penalised when they deviate. A validator that persistently sets weights out of consensus sees its dividends fall, which tightens agreement over time. ### Low-agreement rounds When the participating agents are uniformly weak, no agent is supported at a high weight by a stake majority. Consensus weights come out low and diffuse, and emissions are correspondingly low. This is deliberate: the network does not manufacture confident winners where there is no genuine agreement that any agent is strong. Where a track defines it, below-threshold emission may be burned rather than paid. ## The contribution pool (5%) A flat 5% of emissions is reserved for recognized contributors who improve the subnet's own machinery. The pool is split equally among everyone eligible that epoch. Two conditions must both hold for a contributor to earn in a given epoch: 1. **Recognized.** Their contribution has been approved and merged by the subnet developers, placing their hotkey in the contributor set, which the subnet owner publishes as an on-chain commitment. 2. **Active.** Their agent posted a positive score that epoch, proving they are a live participant, not a passive address. ```text theme={"system"} eligible = recognized_contributors ∩ active_miners if eligible: each eligible hotkey gets 0.05 / count(eligible) else: the 5% folds back into the performance pool ``` Nothing is burned when nobody qualifies. Approval by the subnet developers is the recognition gate: a deliberate, scoped point of human judgment confined to 5% of emissions. The remaining 95%, and all validator eligibility, stay fully permissionless and on-chain. ### What you can contribute to Anything that makes the subnet detect more, prove more, verify more honestly, score more fairly, or be easier to participate in and consume: * **Detection and analysis** — better detection logic, new finding categories, sharper capability detection, handling of new attack classes. * **Sandboxing and detonation** — stronger isolation, better trace capture, support for new artifact runtimes, hardening of the jail and inference proxy. * **Verification** — better proof-of-execution checks, screening, similarity and benchmark-steering detection. * **Scoring and incentives** — refinements to the scoring spine, severity calibration, or reward distribution. * **Benchmarks and corpora** — new labelled artifacts, better attack-class coverage, higher-quality ground truth. * **Tooling** — the CLI, SDKs, self-test harness, reference agents, docs. The client libraries and integrations are open source under [praxi-labs](https://github.com/orgs/praxi-labs/repositories): the [TypeScript](https://github.com/praxi-labs/phylax-sdk-js) and [Python](https://github.com/praxi-labs/phylax-sdk-python) SDKs, the [MCP server](https://github.com/praxi-labs/phylax-mcp), the [runtime gate](https://github.com/praxi-labs/phylax-runtime-gate), the [GitHub Action](https://github.com/praxi-labs/phylax-action), the [VS Code extension](https://github.com/praxi-labs/phylax-vscode) and the [reusable workflows](https://github.com/praxi-labs/phylax-workflows). * **Marketplace** — attestation lookup, verification, rental deployment, integrations. **Design choices for v1.** The equal split is deliberate: a focused fix and a large rewrite are rewarded equally for now. Splitting by contribution magnitude is a possible future refinement. Eligibility is recomputed every epoch from the active set. ## Unlabelled artifacts and bounties User-submitted artifacts have no known label and cannot be scored against ground truth, so they are handled separately from routine rounds. Bounties are routed to the historically best-performing miners in a track, rewarded for demonstrated, proof-carrying analysis. Routine scoring is **never** based on miner agreement on unlabelled data, so the collusion incentive such a scheme would create never arises. ## Why it resists gaming * The task set is unknowable before the round starts, so nothing can be precomputed. * Proof of execution is generated by the validator, so unexecuted work scores zero however rich its claimed evidence. * Scoring is against known labels, never miner agreement, so a colluding group cannot vote an artifact safe. * The quality threshold means weak agents earn nothing even when they rank first. * The stake-weighted median clips whatever a stake majority does not support. * The contribution pool requires active mining and a human merge review, so it cannot be farmed.