Five validators, one agreed number

One ContractOne Agreed Number

Paste a contract, read the streak. Validators mark it against a published rubric and the report only stands where they agree.

Launch the dApp

Reports issued

Published criteria

Cost of a refusal, in inferences

Validators in the pool

Models this pool can draw on

claude-sonnet-4.6deepseek-v3.2gemini-3-flash-previewgemma-4-26b-a4b-itgemma-4-31b-itglm-5.1gpt-5.4gpt-oss-120bgrok-4.3kimi-k2.7-codeminimax-m2.7minimax-m3mistral-largeqwen3-coderqwen3.6-27bclaude-sonnet-4.6deepseek-v3.2gemini-3-flash-previewgemma-4-26b-a4b-itgemma-4-31b-itglm-5.1gpt-5.4gpt-oss-120bgrok-4.3kimi-k2.7-codeminimax-m2.7minimax-m3mistral-largeqwen3-coderqwen3.6-27b

claude-sonnet-4.6, deepseek-v3.2, gemini-3-flash-preview, gemma-4-26b-a4b-it, gemma-4-31b-it, glm-5.1, gpt-5.4, gpt-oss-120b, grok-4.3, kimi-k2.7-code, minimax-m2.7, minimax-m3, mistral-large, qwen3-coder, qwen3.6-27b. 4 nodes name the model they run - the other 16 run a routing policy that picks one of these per call, so what they ran is not published.

Story

A rubric first, then a number

Four commitments this contract makes before a single line is read

The standard is readable

A score against a rubric written after the fact is worth nothing, so ours is published before anyone was scored, with an anchor on every point

One model is only one opinion

Every validator reads the same raw file and marks it on its own, and the mark only stands where they agree under a published rule

Totals are arithmetic, not judgment

Criteria are marked 0, 1 or 2 and the total is summed in the contract, with the band derived by a pure function

A split is our problem, not yours

When validators disagree no report is issued and the splitting criterion is named, which tells us the anchor is written badly

Consensus

Several readings, drawn from a pool of 20

Validators read the same raw file, mark it against the same anchors, and the report only stands where they agree under the rule below

a jury is drawn from the network for every assay

The agreement rule, as published

Two markers may differ by one point, on one criterion only, and the report still stands
The band both markers land in must be the same
Wording is never compared, because two careful readers never phrase a reason alike
4 criteria are counted in deterministic code and 6 are put to the jury

Read from the contract, which fixed this rule in the transaction that deployed it. A validator's vote is one bit, so no per-node mark exists to show here or anywhere else.

Four parts of the machinery

Runs before any fee

The gate

Presence checks only - passing them proves almost nothing, since anyone can type the words into a comment, but failing a required one proves a great deal and costs nothing to find out

declares a py-genlayer dependency headerREQ

declares a class the network can loadREQ

reaches outside the deterministic world at allREQ

declares how validators are meant to agreeREQ

raises at least one error a human could read

keeps its state in a persistent collection

The result

Two numbers, never averaged

A careful contract behind a site that calls a transaction done the moment it is sent is a different problem from a weak contract behind an honest site

The streak on the stone is the same number drawn to length, read against reference marks at 4, 7 and 9

No contract has been marked yet. This panel shows the newest real report as soon as one exists, and it will not stand in a made-up one meanwhile.

Each criterion scores 0, 1 or 2 against a published anchor. The total is summed in the contract, never proposed by a model, and the band comes from the total by a pure function.

The record

One permalink, permanent

Worth handing to somebody else - the source, the digest and the rubric version are recorded beside the marks

What a 10 does not mean

This reads source text. It does not execute, fuzz or verify anything. A contract that scores 10 here can still be broken, and this is a first pass before a person reads it, not instead of one.

recordnothing marked yet