See every subnet, not just the headlines.

Reposted community content highlighted a mobile storage app going live on Android, a research paper analyzing a year of production LLM traffic from an inference subnet, and a preview of an upcoming context-compression marketplace subnet event.

Miners raised concerns that an incomplete re-evaluation process left some submissions scored unfairly compared to others, and are requesting a full re-evaluation using all validator scores.

The subnet owner clarified that scores carry over between rounds after some confusion, and reactions remain disabled to curb harassment; miners also asked routine questions about submission timing and local testing errors.

Published an analysis of jailbreak techniques used against its top-ranked prompt -- its output guard's F1 score improved from 73.9% to 78.53% since release two weeks earlier.

Released an update adding support for both adversarial scanning and adversarial training, adjusted reward splits for top scanning miners, and said a second round of token buybacks is being prepared.

Models run on confidential computing infrastructure so prompts structurally cannot be read by the operator, and run about 20% cheaper than a comparable commercial alternative.

The subnet's min_burn hyperparameter was changed twice in quick succession (down then back up), while community members debated whether certain activity constituted self-mining.

A bug causing miner responses to be truncated at 50 tokens during evaluation was acknowledged and is being fixed, following debate over whether it gave certain miners an unfair advantage in recent rounds.

GPU rental listings on the subnet's marketplace continued to update in real time, with pricing and availability for high-end GPUs fluctuating throughout the day as capacity was rented out and added.

The team clarified that a specific UID represents the subnet's dual incentive mechanism and said the current cycle is projected to end around September 24, pointing miners to documentation for details on the upcoming sandbox phase.

A change requiring first- and second-place agents to use distinct hotkeys took effect this round to spread rewards across more teams, while miners questioned dashboard timing data and pushed back on keeping runner-up agent code private.

Highlighted a published dataset and accompanying research covering one year of production LLM inference traffic, comprising 6.12 billion requests, intended to support research on real-world serving workloads.

Miners found a solution to a previously open mathematical problem, Erdős Problem 1062(ii), with a formally verified proof published alongside a paper, and the team also released a new set of harder problems.

Cohort 3's evaluation completed but missed its final settlement certification deadline and did not activate current weights, while Cohort 4 continues with plans for certified weights to keep renewing until superseded; the team acknowledged reliability problems with recent cohorts.

The subnet's whitepaper is now complete and the project is on track ahead of an upcoming industry conference, with a further funding round planned to coincide with the whitepaper release.

A new release introduced rolling intake, which activates automatically once validators representing 51% of stake upgrade, and multivariate scoring for related time series went live on mainnet.

A bug-bounty and evidence-review program continued processing a queue of reported dashboard bugs, with a portion confirmed as real issues and a public log of validated reports published; miners also asked about the pace of updates.

A review of top competition submissions found several miners disqualified for using prompt injection, with winners announced separately; the team also outlined two planned business models, subscription and API access, for its context-compression product.

Benchmark v13 scoring is underway but has not yet reached the quorum needed to replace benchmark v12 for emissions, and miners pressed for a timeline on the switch while raising concerns about transparency of the top-ranked code.

Regular corpus-refresh updates continued to expand the production training corpus by hundreds to thousands of turns per epoch, while some community members questioned whether certain high-performing activity constitutes self-mining or unfair use of remote evaluation.