Skip to main content

​
Why Groups

Every task is dispatched to a five miner verification group. The group structure provides:
  • A diverse set of independent analyses for the same bundle
  • A consensus signal that anchors individual scores
  • A rotation mechanism that keeps the network meritocratic without freezing newcomers out
  • A collusion detection signal by comparing primary against auditor agreement
A single miner cannot fake their way to high emission. A group of colluding miners cannot consistently bias every group their tasks land in.

​
Group Composition

Each group has two roles.
RoleSlotsSelectionMultiplierWindow
Primary3Top 3 by per type reputation in the candidate pool1.0Full deadline
Auditor2Random sample of 2 from the rest of the candidate pool0.6Tighter window
Primaries do the heavy lifting: full trace bundle, full sandbox manifest, full probe evidence. Auditors only need to produce an SSSA plus probe evidence. The 0.6 multiplier compensates auditors for doing less work.

​
Selection Algorithm

For each task in the round:
1

Build candidate pool

Build the candidate pool from miners that declared support for the task’s skill_type.
2

Filter

Remove miners with:
  • Three or more accumulated collusion flags
  • Per type reputation < 0.2
  • Active dispatch already in flight (rate limiting)
  • Recent failure rate above the network threshold
3

Sort primaries

Sort the filtered pool by per type reputation descending.
4

Apply auditor rotation

For each miner whose last three round assignments for this skill type were all auditor, promote them to the primary slot, demoting the lowest ranked primary to the auditor pool.
5

Pick primaries

Take the top three.
6

Sample auditors

Randomly sample two from the remaining pool (uniform over reputation, not weighted).
7

Dispatch

Send the synapse to all five with role labels and corresponding deadline windows.

​
Auditor Rotation

The auditor rotation is the key fairness mechanism. Without it, miners with high reputation would be permanent primaries and miners with lower reputation would never get a chance to demonstrate quality at the primary level. The rule: if a miner has been auditor for the same skill type for three consecutive rounds, they are promoted to primary on the fourth round, pushing the lowest ranked current primary to auditor. This means:
  • Newcomers eventually get a primary slot regardless of initial reputation
  • Top miners cannot block newcomers indefinitely
  • The promoted miner’s performance is directly comparable to the displaced primary’s
The rotation is per skill type. A miner can be a primary for executable_python and an auditor for mcp_server in the same round.

​
Graceful Degradation

If the candidate pool is smaller than 5:
Pool sizeWhat happens
5+Normal: 3 primaries + 2 auditors
43 primaries + 1 auditor
33 primaries + 0 auditors. No consensus signal but task still runs.
< 3Consensus scoring disabled. Each miner scored on Q alone. Task still runs to keep network learning.
0Task is skipped.
When consensus is disabled, the consensus multiplier defaults to 1.0 for all responses. This avoids penalising miners for the pool being small.

​
Why Two Auditors

Two auditors provide enough independent signal to detect collusion without doubling the cost of each task. The agreement-with-primaries-vs-agreement-with-auditors gap is statistically meaningful at n = 2 because each auditor is a fresh random draw per round. Three or more auditors would improve detection sensitivity at proportional cost. The current trade off is two.

​
Collusion Detection Signal

After each round the validator writes per (hotkey, skill_type, round_id):
  • agreement_with_primaries: average SSSA consensus score between this miner and the other primaries in the group
  • agreement_with_auditors: average SSSA consensus score between this miner and the auditors in the group
Over the last 30 samples, the validator computes: 
gap = mean(agreement_with_primaries) - mean(agreement_with_auditors)
If mean(agreement_with_primaries) > 0.90 and mean(agreement_with_auditors) < 0.60, that is, the miner consistently matches primaries but consistently diverges from random auditors, the validator accumulates a collusion flag for that miner. Three flags eject the miner from group selection until investigated by the operator. The signal works because random auditors are not predictable by the miner. A miner who is honestly producing good SSSAs will agree with both primaries and auditors at similar rates (whatever the network truth happens to be). A miner colluding with specific other miners will only show high agreement with their cohort.

​
Validator Variance

Different validators use different RNG seeds for auditor sampling. This means:
  • A miner sees a different group composition from each validator
  • Collusion across validators is harder than collusion within a single validator
  • The aggregate weight a miner receives is a smoothed function over all validators’ independent group picks
This makes the system more robust than a single deterministic ordering would be.

​
What’s Next

Consensus

The seven component formula computed within the group.

Reputation

Per type reputation that drives primary selection.

Sandbox Reruns

The async verification that follows the round.

Round Flow

Where group selection fits in the wider round.