Bittensor Weekly — October 23, 2026

Between 06:22 and 07:10 UTC on Tuesday, one taker exploited how the validator processed swap events, the team wrote in the channel. Swaps miners had already delivered got marked timed out, and the slash landed in the taker's wallet. A validator-side bug, a regression from the previous week's fixes. No user lost anything: every swap completed or was refunded, and the damage stopped at miner collateral.

The network went down and the fix took a full red team pass before it shipped. UID 189, the first miner hit, was reimbursed on chain. Everyone slashed after that was not, and that is the rule from here: first hit covered, nothing after. Bounties only count when the report lands before the exploit.

Then the rewrite. Version 3.7.3 reopened the subnet on Wednesday with miner rewards scaling to swap fees. Each scoring round pays out up to twice the fees collected over the last 24 hours, and the rest burns. Posting a new wallet costs a one-time 0.02 SOL, and a posted miner wallet can no longer be the source of a swap. Emission now follows revenue.

The repo reads like the incident report. Read program events only from the program's own invocation frame. Refuse a posted miner wallet as a reservation source. Refuse a claimed source deposit as a delivery. Twelve commits landed from Tuesday to Thursday, and read in order you can watch the doors close one at a time.

Then Thursday, another hole. A taker can name someone else's wallet as the source, wait for that wallet to send the miner money for any reason, and submit that transfer as its deposit. Recipient, amount, sender and timing all line up, so validators accept it. The miner then pays the taker with money that was its own, or eats a slash for not paying. Evan's proposed fix: make the taker prove it controls the source wallet when it reserves. The team is aware and deferring to miners instead, recommending they pull quotes before topping up, because a signature check adds a step for users. That trade is in the miner guide. A subnet that got burned once is choosing convenience over closing the gap.

Score (SN44) spent the week closing the holes in how it picks a winner.

Three fixes in five days. The copy check moved first: a model now needs to differ on 3 of 10 sampled past challenges to shed the copy flag, down from 4. Behind that sits an audit of 12,671 element-epochs, one task type by one round by one winner, since July 11. The paid model was not the top scorer in 2,565 of them, and every one of those went to an older commit.

Then the tiebreak. A miner was skipping the tasks it was weak on, then taking the top spot on the tiebreak anyway. It now compares the last 10 shared challenges both hotkeys scored, so skipping a task stops paying. Thursday brought Tiebreak Profiles: separate rules per element, so detection work and counting work stop sharing one tiebreaker.

The channel runs on block numbers. A miner posts a block, the owner checks the annotation, corrects the ground truth when the miner is right. "Checks are performed once a day, pointless to multi ping you won't get an answer faster," he told one of them, then fixed the block. Fallen-person detection went up in difficulty Thursday, described as his last attempt at that element.

Almanac (SN41) is 17th on Polymarket's builder leaderboard.

$7.89M of routed volume in the last month, per Polymarket's own table. A second builder code, almanac.market.2, sits 22nd at $6.30M. Put them together and it is $14.19M, past Jupiter's $12.03M in 10th.

155 miners split 25 TAO a day, and their forecasts settle against markets anyone can check. The people who moved that volume came for the predictions. None of them had to know the models come from miners.

That is the shape of a subnet that works: the product is the front end, the incentive layer stays invisible. Almanac made Polymarket's builder table on prediction volume. The Bittensor part is plumbing.

Hot. 1,472 messages across 75 channels on Thursday, up from 958 across 74 the week before. ORO (SN15) led at 125, RedTeam (SN61) at 112, Harnyx (SN67) at 102, Score (SN44) at 89, BitMind (SN34) at 88, Beam (SN105) at 77, Affine (SN120) at 69, Yanez (SN54) at 61.

Building.

Friction.

Harnyx's batch broke on Thursday when a validator went out mid-run. By one miner's count, 209 of 251 scripts failed a single OpenRouter billing check, and 38 agents got paid. The owner's read: a fix for yesterday's stall never reached the validator, so the push landed between batches. The commitment from here is to fail a batch outright when verification breaks, and split the emission evenly.

Affine (SN120) paid 10 UIDs out of more than 500, and its held-out score slipped from .75 on e75 to .656 on e76 and e77. One miner called it "LUCKY TAKES ALL" with a straight face. The team's answer is that held-out runs bounce and the on-chain weights keep moving.

Niome (SN55) got the sharpest question of the week: why does the top 10 hit 0.65 to 0.71 consistency when an honest design sits at 0.10? The team says seeds are drawn at the start of validation and points at the W&B upload times. Miners pointed back at the arithmetic.

Claims (SN111) went live with the 1/n split on Monday night having been flagged the previous Friday, with no cutover date announced. The top of the board repriced itself immediately.

The protocol. Null consensus gives subnet owners a second epoch mechanism: the highest-stake UID holds the only validator permit, sets the weights, and decides miner pay for the whole subnet. Consensus and bond math get skipped, and UID capacity goes from 256 to 2,500, shared across mechanisms at 1,250 for two and 625 for four. Owners switch it with one hyperparameter, epoch_consensus. The root channel asked the right question within the hour: how does anyone check the one permit holder ran the scoring code. "We don't. Don't verify, trust," came the reply, which is the deal validators already live under. Null writes it down.

Cold. 27 of the 102 channels in this week's record posted nothing on Thursday. The map still runs on about a dozen teams carrying the news, and the rest watching.

Thursday on SN67 started with an accusation. Almost every agent in the batch had failed on a billing check, one miner called it self-mining, and the owner was asked to take it to DMs.

"I'd rather keep the discussion public since there has been suspicions as such so I want to resolve it for all."

That is a harder sentence to write than it looks. The easy move is the DM, where nothing you say gets quoted. The owner stayed in the channel for four hours, explained that a fix for the previous day's stall never reached the validator, and then said the part most teams leave out.

"It is unfair and unfortunately we will have more incidents as such as we keep on trying to move faster."

Then the commitment: when verification breaks, the batch gets failed and the emission splits evenly. Frost, who had spent the morning asking how long a fix would take, came back to it: "Thanks for taking the trouble to sort it out for all of the miners." LoonerFinger put the stakes plainly: "I hope the subnet retains the trust of the miners."

That trust is the whole asset. Subnets lose miners over a broken batch far more easily than they lose them over a bad week, and the ones that keep them are the ones that answer in public while it is still going badly.

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]