Skip to content

PROTOCOL

A shared process for improving intelligence.

Intelligence improvements are often developed in isolated systems. Evaluating changes, establishing reproducibility, tracking dependencies and rewarding downstream contributions can require disconnected tools and institutions. The protocol proposes one process for proposing, evaluating, recording, attributing and reusing improvements.

at a glance

Settlement
Robinhood Chain · 4663
Evaluation
controlled · replicated
Records
hashed · lineage-preserving
Token
none

PROBLEM AND PROPOSAL

Isolated improvements do not compound.

A result produced in one lab, one company or one agent framework rarely becomes a component another team can verify and build on. Reproduction is manual, attribution is informal, and the knowledge of why something worked is lost with the environment it worked in.

Kuxer treats the improvement itself as the unit of coordination. A proposal names its baseline and parents, an experiment runs it under fixed conditions, evaluators reproduce the result, and the accepted record becomes a parent for later work. The economic layer settles commitments and, later, contribution accounting on Robinhood Chain · 4663.

CORE OBJECTS

Nine logical objects.

These are the proposed logical objects of the protocol. Final schemas are agreed before contracts are extended; the current draft is schema version 0.1.

Capability
A defined capability or research target, with its primary metric and direction.
Baseline
Reference implementation and evaluation conditions a candidate is compared against.
ImprovementProposal
A proposed modification: source, parents, contributor and a content hash.
Experiment
A reproducible evaluation run of candidate versus baseline inside an isolated sandbox.
EvaluationRecord
Results, metrics, evaluator attestations, reproduction status and limitations.
LineageRecord
Parent–child relationships between contributions: derives, combines, supersedes.
Contribution
Attribution and reuse relationships between addresses and records.
RewardClaim
A claim on a defined reward.
planned · not part of version 1
Challenge
A dispute regarding an evaluation or protocol claim, resolved by appending an outcome.

MECHANISM

Propose, experiment, evaluate, validate, inherit, evolve.

Each stage produces a record. Together they make an improvement traceable from the first proposal to the later work that builds on it.

  1. 01

    Propose

    ImprovementProposal
    Submit a candidate improvement to a defined capability, algorithm, reasoning method or research procedure, naming its baseline and parents.
  2. 02

    Experiment

    Experiment
    Run the candidate against the baseline in an isolated environment with identical instances, seeds and limits; capture versions, resources and outputs.
  3. 03

    Evaluate

    EvaluationRecord
    Assess performance, generalization to held-out tasks, efficiency, regressions and reproducibility under preregistered criteria.
  4. 04

    Validate

    EvaluationRecord · status
    Replication runs must reproduce the result before the record moves from pending replication to validated.
  5. 05

    Inherit

    LineageRecord
    Validated improvements become parents: later proposals derive from them or combine several of them.
  6. 06

    Evolve

    research objective
    Test whether the research methods themselves (suites, baselines, evaluation procedures) can be improved, and record those changes the same way.
See the lifecycle in detail

IMPROVEMENT CRITERIA

What counts, and what does not.

Evaluation may consider task performance, generalization, compute and execution cost, reliability, regression risk, reproducibility and contribution to future discovery. Research domains are never compressed into a universal intelligence score.

not an improvement
  • Unsupported claims: a self-reported score without a controlled comparison.
  • Evaluation leakage: a candidate tuned on the held-out set.
  • Irreproducible results: a gain that disappears when the run is repeated.
  • Unverified downstream attribution: credit inferred from proximity instead of recorded contribution.
  • Benchmark gains that hide material regressions in validity, generalization or cost.
limitations
  • Cryptographic commitments establish record integrity, not scientific truth. A hash proves a retrieved record matches what was anchored.
  • Independent evaluation can fail through collusion, test leakage or poorly designed metrics; signatures alone do not prove independence.
  • The first release covers one bounded domain. Claims about other domains require their own baselines and evidence.

next

Read how the layers fit together, then the methodology.

ArchitectureEvaluation methodology