> ## Documentation Index
> Fetch the complete documentation index at: https://docs.reppo.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Stake-Assured Curation with veREPPO

> Reppo's curation model ties voting authority to locked capital, creating economically-grounded quality signals suitable for RLHF and DPO training.

Reppo's curation mechanism is **stake-assured human feedback (SAHF)**. Rather than granting curation influence based on time spent or engagement metrics, Reppo requires voters to lock REPPO tokens to obtain veREPPO — their voting authority — and then allocate that authority across datanets and pods each epoch. Economic risk backs every quality signal the network produces.

## Voting power formula

Your voting power is determined by two factors:

* **Amount locked** — more REPPO locked yields more veREPPO
* **Lock duration** — longer commitments yield more voting power per token

```
REPPO locked × lock duration → veREPPO (voting power)
```

Splitting tokens across multiple wallets does not multiply voting power. Because voting weight scales with capital locked, Sybil attacks — creating many accounts to dilute or manipulate curation — are economically self-defeating.

## How votes work each epoch

Each epoch runs for **48 hours**. During that window:

<Steps>
  <Step title="Distribute voting power across datanets">
    veREPPO holders allocate their total voting power across the datanets they want to participate in. You can split across multiple datanets in any proportion.
  </Step>

  <Step title="Allocate within datanets to specific pods">
    Within each datanet, you allocate your share of voting power toward specific pods — casting votes for or against based on your quality assessment.
  </Step>

  <Step title="Time decay applies">
    Voting power decays linearly throughout the epoch. A vote cast at the start of the epoch carries more weight than the same vote cast near the end.
  </Step>

  <Step title="Two-sided voting creates adversarial pressure">
    Any participant can vote against a pod. This means artificial support for low-quality content can be directly challenged, and bribery schemes become ongoing capital contests rather than one-time payments.
  </Step>

  <Step title="Epoch closes and tallies determine rewards">
    At epoch end, net vote tallies across all pods determine how emissions are distributed to publishers and how curation accuracy scores are calculated for voters.
  </Step>
</Steps>

## Time decay

Earlier votes carry more weight. A voter who identifies a high-quality submission early in the epoch earns more than one who casts the same vote near the close. This design discourages **vote sniping** — waiting to see which way consensus is forming before piling on — and rewards genuine conviction.

## Two-sided voting

Votes can be cast **for** or **against** any pod. This means:

* Artificial inflation of a low-quality pod's score can be directly countered
* Supporting a pod through bribery requires maintaining that capital commitment every epoch as opponents can challenge it
* Markets self-correct through continuous adversarial pressure rather than relying on manual moderation

## Adversarial robustness

The voting design addresses common manipulation vectors systematically:

| Attack vector            | Defense mechanism                                                         |
| ------------------------ | ------------------------------------------------------------------------- |
| **Spam submissions**     | Pay-to-publish requires upfront REPPO capital per pod                     |
| **Sybil voting**         | Stake-weighted: splitting capital across accounts reduces total influence |
| **Artificial support**   | Two-sided voting: any participant can challenge any position directly     |
| **Vote sniping**         | Time decay: earlier signals receive higher weight                         |
| **Coordination attacks** | Continuous repricing: positions respond in real time as capital moves     |

## Where expertise enters

Locking REPPO creates voting power and economic accountability. It does not create domain knowledge. Expertise enters at the datanet level, not from stake alone: each datanet owner chooses the task, access rules, incentive design, and acceptance standards. A datanet can be open, curated, or restricted to a narrower expert set. In specialized markets, stake backs judgment among eligible participants rather than manufacturing expertise where none exists.

## How feedback is collected

Datanets capture human feedback in several forms, depending on the task:

* **Rating outputs:** people rate AI-generated content (good or bad, helpful or harmful, preferred or less preferred).
* **Pairwise comparisons:** given two outputs, people pick which is better.
* **Direct edits or suggestions:** annotators or users improve the output, such as rewriting text or correcting errors.
* **Specialized feedback:** domain experts (for example lawyers, doctors, or teachers) review content for accuracy in their field.

This feedback is turned into training signal for techniques like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO).

## Curation signals as training data

Every action on Reppo — submissions, votes, rankings, and written feedback — produces structured preference signals. These signals are the network's primary output. AI teams access not just curated content but the **preference signal itself**: ordered comparisons, for/against labels, and written rationales suitable for RLHF (reinforcement learning from human feedback) and DPO (direct preference optimization) training pipelines.

<Note>
  Vote feedback submitted via `POST /feedback/pods/{id}` attaches structured written rationale to a vote. This adds qualitative signal on top of the quantitative veREPPO weight, enriching the training data that flows downstream.
</Note>