> For the complete documentation index, see [llms.txt](https://docs.konnex.world/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.konnex.world/understand-konnex/validator-metascore.md).

# Validator metascore (design)

On Konnex, **validators score miners**: they rank physical work against the task instruction and subnet policy using replays, sensors, benchmarks, and—where the workload requires it—**AI-assisted scoring** (e.g. kinematic smoothness, grasp success, VLA behaviour). Choosing a good scoring function is therefore security-critical.

Plain smart-contract checks alone do not close the loop: **AI scoring can hallucinate**, and **validators can collude**. The protocol treats “who validates the validators” as a first-class problem and implements a **hybrid, three-layer metascore** on the Substrate-based chain.

## Layer 1 — Weight alignment (peer consensus on scores)

The primary cryptoeconomic signal for validator quality is **agreement with the rest of the validator set** on the same miner outcomes.

Each validator submits a **weight vector** (a list of scores) over a batch of miners or tasks. The chain builds a **global consensus vector**—the same dimensions, aggregated across validators—using a governance-tunable rule such as **median per dimension** or a **stake-weighted** blend. Call validator *i*’s vector `W_i` and the network aggregate `W-bar`. When `W_i` tracks `W-bar`, that validator’s standing improves; **systematic drift** away from the aggregate cuts **validator trust (vTrust)** and **lowers effective APY**. The consensus term in the metascore (below) is the formal similarity between `W_i` and `W-bar`.

**Lazy copy-paste consensus** (everyone mirroring the same scores without doing work) is mitigated by **pseudorandom hidden spot checks**: validators cannot assume which submissions are audited only by alignment, so blind copying remains risky.

## Layer 2 — Honeypots (deterministic reference tasks)

Robotics workloads have a structural advantage: **simulators** (e.g. ManiSkill, Isaac Sim, subnet-specific benches) can produce **deterministic ground-truth metrics**.

The network (or a governed oracle module) **injects honeypots** into the task mix—runs whose correct grades are known **before** validators vote. Validators do not know which jobs are honeypots.

If a validator’s AI stack **awards high scores to an objectively failed reference** (e.g. dropped object marked as success), the protocol applies a **hard penalty**, up to **slashing** on severity tiers defined by governance.

## Layer 3 — Two-tier scoring inside the validator node

Each validator’s **local scoring pipeline** is designed as **two independent channels** that must agree before a vote is safe to publish:

1. **VLA / LLM layer** — High-level task understanding (did the miner follow the instruction semantically?).
2. **Deterministic layer** — Lightweight heuristics or closed-form checks: torque limits, timing windows, joint limits, replay deltas against sim logs, etc.

If the two layers **conflict**, the validator **abstains** or **requests an extended check** (subnet-defined escalation path). Publishing a confident vote despite internal disagreement is scored against the validator under Layer 1 and Layer 2.

## Onchain metascore (L1)

At the base layer the protocol combines alignment, honeypots, and penalties into a single **validator metascore** for each validator (written `S(V_i)` in the figure):

![Validator metascore: S of V sub i equals alpha times C of W sub i and W-bar, plus beta times H of V sub i, minus gamma times P sub i](/files/ZDxw999n82lgBZRVxC3h)

Read the formula left to right:

* `S(V_i)` — The scalar metascore for validator *i*; higher is better after governance normalization.
* `C(W_i, W-bar)` — **Consensus term**: how close validator *i*’s score vector `W_i` is to the network’s consensus vector `W-bar` (for example cosine similarity, or another distance mapped to a score on the closed unit interval from 0 to 1).
* `H(V_i)` — **Honeypot accuracy**: share (or count) of hidden reference tasks where *i*’s grades match the known-good labels.
* `P_i` — **Penalties** for operational faults: high latency, missed voting epochs, broken commitments, and similar slashing-adjacent behaviour.
* **α, β, γ** (alpha, beta, gamma) — Positive **governance weights** that set how much consensus, honeypots, and penalties move `S(V_i)`. They are tuned per subnet phase and mainnet rollout.

## Relation to mining and PoPW

Miner rewards and PoPW acceptance still flow from **subnet verifiers and task semantics**. The metascore **does not replace** workload-specific scoring; it **governs who is trusted to emit those scores** and at what economic weight.

## See also

* [Validating on Konnex](/participate/validating.md)
* [Proof-of-Physical-Work (PoPW), TEE, and contracts](/understand-konnex/contracts-and-popw.md)
* [Protocol architecture](/understand-konnex/protocol-architecture.md)