Bittensor Weekly — October 30, 2026

Since the chain handed owners their own dial this month, 18 of 128 subnets have moved off the standard 18% owner, 41% validators, 41% miners. Nine raised the owner cut. Three lowered it. Six kept the owner at 18% and moved the rest around.

The seven at the cap are SN20, 35, 36, 72, 73, 98 and 117. What each one did to the other two columns is the part worth reading. SN20, SN73 and SN98 left miners on 60% to 61% and cut validators to 6% or 7%. SN35, SN36 and SN72 split the cost, taking validators to 30% and miners to 37%. SN117 took both down to around a third.

Then the subnets that went the other way. Vanta (SN8) and SN74 take nothing for the owner and split the rest 50/50 between validators and miners. Almanac (SN41) takes 10%. Ten subnets pay miners above the old 41% ceiling, led by SN47 at 69.4% with its validators on 12.6%, SN122 at 64% and SN16 at 61.5%. ReadyAI (SN33), Yanez (SN54) and SN76 went the other direction, paying miners 33% and validators 49%.

The big earners have not touched it. Chutes (SN64), Lium (SN51), Targon (SN4) and Templar (SN3) all still run 18/41/41, which is where the whole network started eleven days ago.

For a staker, the split is now a number to look up before picking a subnet. Two subnets with identical emission can pay you very differently, and the owner's cut has to come from somewhere.

ReadyAI (SN33) had its conversation task taken apart in public, conceded the flaw, and is retiring the task type that carried it.

A miner laid the mechanism out on Thursday. A validator cuts each conversation into five random ten-line windows, sends each window to six miners, and all five of those tasks carry the same enrichment cards, so the cards fingerprint the parent conversation. Hold enough slots and the windows can be stitched back together, and scoring is taken from the centroid of the whole conversation rather than the window that miner was sent.

"If you have 96 slots you will have ~100% chance of being able to rebuild the conversation," one of them wrote. Another asked whether the post had just leaked an exploit on the way past.

The owner did not dispute the mechanics. Conversation tasks are already less than half of what miners get, papiercirer said, and the Skill Store launch takes the subnet to almost all skill tasks. Asked whether conversations go to zero, the answer was yes, gradually, over the next few weeks.

A task type that pays for reconstructing a conversation across UIDs was always going to reward the miner holding the most slots. ReadyAI is choosing to stop paying for it instead of patching it.

Image: /home/wiggly/Documents/bittensor-subnet-sentiment/daily_reports/20261030_shots/sn33_1565417541155688614.png

Swarm (SN124) opened its first client-facing task and had 39 submissions in the queue before evaluation could keep up.

Sentinel launched Thursday. A drone leaves its dock, patrols a solar park through the night, reports every intruder and lands. The three sim tasks miners had been grinding, Autopilot, Swarm Autopilot and Interceptor Office, closed with v6.0. The drone and dock are modelled in detail because the winning policies are meant to fly on hardware.

Thirty-nine models went into the Sentinel lane at 0.5 TAO each, and each one runs the full 2,000-seed benchmark in sequence. That is days of queue for a competition that opened the same day. "The 39 submissions paid 0.5TAO to register at least, so its not like we can ban them for no reason," miguelik wrote, turning down a request to halt the run and start over.

The scoring got more honest in the same pass. The website had not been applying the 5x penalty, so a model that never reports, or reports everything, looked better on the board than it was. Every missed intruder and every false alarm turns five of a model's best patrols to zero once the penalty lands.

Hot. 1,415 messages across 84 channels in the last two days of the week, from 414 different accounts. ORO (SN15) and Swarm (SN124) tied at 130, Score (SN44) and Affine (SN120) at 110 each, Yugos (SN34) at 77, Ditto (SN118) at 59, RedTeam (SN61) at 51. The top five carried 39% of everything posted.

Building.

Friction.

Affine (SN120) is the loudest room in the network. More than 650 UIDs are registered, one submission gets trained per epoch, and the round sat stuck for over 12 hours on Thursday. "The 600 deep queue is not the solution," consτ wrote, floating a switch to batch competition with registration as the queue. Miners asked to go back to model submission. One of them found a queue loophole by swapping hotkeys and handed it over rather than using it.

ORO (SN15) had the opposite problem: a wall of rejections. Miner after miner asked why their code came back as an anti-cheating failure, and the answer did not move. "Folks I have looked at all of the hotkeys that have been sent this morning and all are correctly being rejected," Iron Seth wrote, then linked the code requirements page. Race 8 has the qualifying bar under protest from a dozen angles, and the bar stays: "we aren't changing the qualifying score today."

Niome (SN55) miners published arithmetic against the winner. Using the published contract's weights and 250 rows, they put the ceiling at Q 265.55 for a winner reporting D 0.932, while UID 243 reports Q 280.499. The owner says the team is investigating. Nothing has been published since.

ChronoLLM (SN38) watched its alpha slide and lost a miner over it. "I have no energy to continue my training. Bye Chronos!" he wrote on the way out. The team says the mechanism is unchanged and incentives still cover the compute for a 2B model.

Instant (SN46) has miners asking when the first buybacks land, with the API portal live and listings still arriving. The answer is Friday to Sunday, on the dashboard.

Cold. 44 of the 128 subnets posted nothing in the last two days, and another 33 managed five messages or fewer.

Thursday in Astrid (SN127), a subnet owner had to say out loud that his own scoring formula was the problem.

The setup was ugly. Payouts for two competitions had been held while integrity checks ran, several miners got flagged for sharing infrastructure, and a group of people who share an office and an internet line got flagged alongside everyone else. Léo Mercier could have defended the check. He wrote out the business case instead: Arena pays alpha for trading intelligence, and buyers will not pay for what they can get free. "In MS6 and MS7, many top results were one or two trades timed around a single BTC move, then weeks flat. One trade can't tell skill from luck."

Then the line that cost him something. "Our Astrid Score ranked that profile at the top. You learned the formula and tuned to it, which is exactly what we asked. So the formula has a gap, and the fix is ours to make."

The miner who had been flagged answered inside the hour, and the answer was not a demand. He said he appreciated the honesty, that miners had tuned to the exact formula they were asked to optimize for, and then: "Not many subnet owners take accountability like that and we genuinely respect it."

Nobody got paid that afternoon. What they got was an admission, and a miner said thank you for it. That is the exchange this network runs on, and it costs nothing but nerve.

Next week's issue goes out Friday. Subscribe and it lands in your inbox: https://newsletter.bittensor.quest

Shaping the next issue: tell me what to keep, what to cut, and what you want more of — mailto:[email protected]