VERNA — Vernacular LLM · Solana mainnet

The alignment layer,out of the vendor’s hands,onto an open ledger.

VERNA is an open language model whose refusal thresholds are voted by the people who use it. Every preference, every label, every adapter is hashed to Solana mainnet and folded into the next version of the weights.

every contribution → VERNA:1:<kind>:<sha256> · memo program · mainnet-beta

···

On-chain contributions

···

Contributors

···

Verified votes

···

Current round closes

figures count verified mainnet memo transactions only
I.The thesis

Censorship does not live in the weights. It lives in the alignment layer above them — and today one company sets that layer behind closed doors.

Every open model you can download has already been given a character: what it refuses, how it hedges, which words it will not say. Those decisions were made in a private RLHF pipeline, by the vendor, before you ever typed a prompt. The weights are public; the judgment baked into them is not.

VERNA moves that layer into the open. Filtering thresholds become parameters of a public protocol, voted by token holders category by category — profanity, fiction, controversy, high-stakes advice, refusal style. The model, its changelog, and its refusal rules stay inspectable, round after round.

The weights stay open. The base models stay interchangeable. What the community comes to own is the one thing that was never on the table: the character.

II.The problem

Open weights are not open character.

01

Weights without character

An open-weight model arrives finished. The vendor already decided what it declines and in what wording, and shipped that judgment inside the checkpoint. You can run it anywhere — you cannot re-decide it.

02

Retraining is priced for datacenters

Fine-tuning and alignment take data pipelines, GPU fleets, and specialist hands. For one person, re-aligning a modern model to their own norms is technically possible and practically out of reach.

03

A thousand small contributions

VERNA aggregates the work instead of demanding it whole. A single LoRA delta weighs 8–40 MB and travels over ordinary internet; preferences and labels weigh even less. A thousand people steer one model with no datacenter among them.


III.Architecture

Four layers, each with one thing that is real.

i

Contribution

Participants vote on answer pairs, label alignment data, or submit LoRA adapters. Every accepted contribution is hashed and written on-chain; Solana keeps that stream cheap. Reputation follows what later rounds accept.

What is real

a memo transaction on mainnet, one per contribution

ii

Aggregation

On a fixed weekly schedule the protocol folds adapters and preference data into a new version of the weights, published beside a changelog of what moved and on whose contributions.

What is real

weekly rounds anchored by merkle roots

iii

Inference

Requests route to open bases across existing GPU pools and DePIN networks. No proprietary model and no owned compute at the start — the burn rate stays honest. Settlement in $VERNA or USDC.

What is real

routed open models, metered per request

iv

Alignment governance

Holders vote the filtering threshold for each content category, from strict to minimal. One floor — illegal content — is welded into the protocol and never reaches a ballot.

What is real

votable thresholds + an immutable floor

IV.Real today, not theatre

The line between live and later is drawn in ink.

Live now

Open-base hosting — Llama, Qwen, DeepSeek, Mistral — with switchable filter levels

Roadmap

Verifiable decentralized RLHF and RLAIF

Live now

Preference-pair and label collection with on-chain accounting of every contribution

Roadmap

True distributed fine-tuning, beyond adapter aggregation

Live now

LoRA adapter aggregation — FedAvg over user-submitted deltas

Roadmap

Pretrain-level contribution, in the spirit of Prime Intellect and Nous

Live now

Token utility and buy-and-burn from inference revenue

Roadmap

An owned compute layer

The on-chain accounting is real from day one — even while the contribution itself is labels and preferences rather than gradients. Anything simulated here would be trivially exposed, and the project would burn on it. That accounting is what separates VERNA from an AI memecoin.

V.The floor

The floor is not on
the ballot.

Six categories of illegal content are hard-blocked by an independent classifier before a single token is generated. This floor is welded into the protocol: it is not a threshold, it has no slider, and no vote — whatever its weight — can lower it.

It exists for one reason: survival. Without it there is no app-store presence, no exchange listing, no fiat rail. A protocol that cannot exist cannot be governed. Everything above the floor belongs to the vote; the floor itself belongs to no one.

  • S4Child sexual exploitation
  • S1Violent crimes & terrorism facilitation
  • S9Indiscriminate weapons (CBRN)
  • S3Sex-related crimes
  • S2Serious non-violent crimes (fraud, malware, trafficking)
  • S11Suicide & self-harm facilitation

screened before generation · a request that hits the floor is refused with its code — nothing is generated

VI.The token
Inference paymentPremium access & daily quotaVote weight on thresholdsContributor rewardsBuy-and-burn deflation

mint — awaiting TGE · launch via pump.fun · liquidity on PumpSwap


The invitation

Write yourself into the next version of the weights.

A preference, a label, or a LoRA adapter — every accepted contribution becomes a line in the ledger and a trace in the model.