Dates and verification labels belong to the original edition. A past verified announcement can now be outdated. Historical project names are retained; an SN filter searches the slot across registrations. All keywords must match the story text or subnet label. Ratings describe editorial priority, not investment quality.
SN10 Pareton closes its vLLM campaign and prepares SGLang
The latest SN10 Pareton channel digest attributes closure of the vLLM competition to the team, with no further submissions accepted. It says the project is concentrating on a larger SGLang campaign because of stronger demand, with details still to come. The digest describes rewards being directed to the burn UID during the gap, and the current screener shows zero observed subnet allocation. Yesterday’s active-campaign description has therefore changed, but a start date or live reward budget for the replacement campaign has not been confirmed.
Closing an active campaign immediately affects miner work and rewards; the replacement should not be treated as open before its rules are published.
SN125 Refinery reports mainnet readiness as positive allocation appears
Today’s SN125 Refinery digest says the subnet is live after submission-processing delays, and that source code from winning and leading submissions has been shared. It sets the evaluation budget at 188,000 steps or 72,000 seconds. The 11:21 AM ET screener now shows 0.0231% allocation, up from zero in yesterday’s snapshot. The allocation change is independently observable in the feed; completion of all delayed evaluations and individual payouts remains unverified.
A reported opening accompanied by a new positive allocation is a meaningful update to yesterday’s stalled-weight status.
SN80 OpenRoboto CLI 1.4 protects submitted robot weights before payment
SN80 OpenRoboto’s CLI 1.4 changes real-robot submissions to put the Hugging Face repository behind manual approval and grant the season’s named evaluator access before paying the submission fee. This keeps weights restricted while allowing evaluation to proceed. The change avoids a fully private repository that would block the evaluator, and it leaves simulation seasons alone when no evaluator grant is configured. It follows yesterday’s separate extension of real-robot reward accrual through evaluation.
Protecting submitted weights while preserving evaluator access addresses a practical risk at the point where miners pay to enter.
SN38 ChronoLLM changes Round 11 quality tests to instruction following
SN38 ChronoLLM’s new Round 11 evaluation code replaces open-ended text-completion prompts with direct questions and instruction-following tasks. The judge is now told to assess how well a response answers the question, alongside accuracy, clarity and knowledge. This is a concrete change to what the competition rewards, following yesterday’s price-led coverage. A repository update alone does not show every validator has already adopted it.
Changing the quality exam alters the target miners optimize for and gives the current market interest a substantive development to follow.
SN67 Harnyx adds a black-box rating competition
SN67 Harnyx has merged a rating competition in which comparisons and judgments are exchanged through signed endpoints, with durable accounting of completed work. The validator exposes readiness for the rating worker and tracks results independently of a transient request. This expands the evaluation mechanism; the code does not by itself establish a completed production round or its reward distribution.
A new competition path with persistent result accounting is material to how miner work can be evaluated and rewarded.
SN81 Reliquary makes miner verdicts and selection rules easier to inspect
SN81 Reliquary’s latest changes expose an optional miner-verdict view with lifecycle information and reasons, explain selection decisions, and persist final outcomes. A follow-up reduces diagnostic load and recovers durable verdict history. These changes make a rejected or unfinished submission easier to distinguish from a scored result; production-wide rollout still depends on validator upgrades.
Visible verdicts and reasons help miners understand outcomes and scrutinize the reward process.
SN20 Witness adds signed video rounds and a safer return to burn
SN20 Witness has added signed MP4 production rounds, structured evaluation and daily processing budgets. A follow-up fix invalidates the previous completed round when preparation of the next source clips fails, returning the system to burn instead of continuing to use stale weights. This is a specific reward-safety behavior in the new code, not evidence that a fresh successful video round has already paid miners.
Preventing stale results from continuing to earn during a failed round directly improves reward integrity.
SN92 MicroTensor closes the latest guard round and tightens blocked-miner settlement
The SN92 MicroTensor channel digest reports that guard / mt-4g Round 1,241 closed at block 9,061,144, with 14 of 23 committed systems still sealed at the digest’s observation. Separately, new public code threads the control-plane blocklist into settlement so barred hotkeys receive no earnings. The round closure is a digest report; the code change is directly inspectable. No final winner, refund decision or fully settled round was established.
The first follow-up to the new guard arena concerns actual submission progress and who remains eligible to earn.
SN74 Gittensor fixes a benchmark that falsely reported a 96% improvement
SN74 Gittensor’s sparkinfer evaluation code now uses five measurements at every scored context size. The commit explains that a single unusually slow baseline made a normal result appear 96.1% faster and caused an automatic merge; the claimed gain was not real. Repeated sampling addresses the scoring flaw. The same commit says the previously merged entry still needs separate correction, so the measurement fix should not be read as a completed reversal of that award.
A documented false performance win and a targeted measurement fix matter more than repeating the inflated headline.
SN91 Cascade blocks diverged training checkpoints
SN91 Cascade’s latest trainer change stops work when loss becomes non-finite and refuses diverged checkpoints at ingestion. It helps keep a failed training run from becoming an accepted model artifact. The previously announced multivariate scoring boundary remains block 9,068,400, ahead of this report’s snapshot. A channel claim about future multivariate dataset shares is a plan, not the current scoring mix.
Catching invalid training output before acceptance protects evaluation quality during the funded-round rollout.
SN100 Cortex adds clearer trial evidence; a 72.8K-alpha conviction lock is recorded
SN100 Cortex’s latest proof work adds bounded per-trial Harbor logs and redacts decoded log content before truncation, giving failed benchmark runs more useful evidence without an unbounded payload. Separately, SubnetRadar records a new 72.8K-alpha conviction lock on September 13 at 19:38 UTC. The lock is a staking event, not proof of a new market purchase or an announced customer. No new consumer-app approval was established.
Better benchmark evidence is a concrete engineering update, accompanied by a separately sourced material conviction event.
SN78 UMI adds an ASL reference contribution while production gates remain
SN78 UMI’s reference-model repository has received a community ASL baseline contribution, while the main repository continues work on evaluator evidence, upload retention and competition activation. The reference work offers a starting point for builders; it does not certify translation quality on an independent production holdout. Yesterday’s temporary endpoint-availability rewards remain distinct from the permanent translation competition. An advertised AMA has conflicting weekday and calendar-date information, so a firm event time is not listed here.
A reusable model contribution advances participation, while keeping the temporary reward bridge separate from a fully activated quality competition.
SN113 TensorUSD explains how strategy fees would support its reserves
Following yesterday’s expanded Pallet partnership, SN113 TensorUSD has outlined the intended money flow: miners develop long/short, delta-neutral alpha strategies, users can copy them through Pallet, and performance fees help build TUSDT reserves and liquidity. The new information is the proposed mechanism and reserve link. It is not confirmation of a live derivatives market, customer returns or funded reserves from this product.
The fee-to-reserve design explains why the planned product matters beyond a partnership headline.
SN118 Ditto prepares v13 and explains future buyback funding
SN118 Ditto says future buybacks will use 50% of profit after costs, and distinguishes its initial miner-registration-funded buyback from the intended customer-funded model. The repository also prepares the v13 agent-review policy and corrects the submission fee to 0.1 TAO. The rollout material is activation preparation, while the channel digest targets this week; no live v13 activation or new customer-funded burn was established.
The funding source and participation rules are economically meaningful, with future plans separated from completed transactions.
SN21 AdTAO applies for the Claude Connectors Directory
SN21 AdTAO says it submitted its connector for inclusion in the Claude Connectors Directory. The post explicitly says approval is pending. A successful listing could improve discovery by people using Claude for advertising workflows, but an application is not an approved integration, a signed Anthropic partnership or evidence of new customer revenue.
A concrete distribution application is worth tracking, with approval status stated plainly.
SN105 Beam reports a one-terabyte-per-half-hour traffic pace
SN105 Beam reports processing roughly one terabyte every 30 minutes and extrapolates that pace to 48 terabytes per day and 1.44 petabytes per month. The current throughput claim is the update. The monthly number is a run-rate projection, not a measured month of traffic, and the post does not establish how much of that traffic generated paid revenue.
A quantified throughput update gives a useful operating signal when its extrapolated totals are clearly labeled.
SN63 Enigma outlines a path from cryptography work to licensing revenue
SN63 Enigma’s new article lays out a commercial model in which subnet-funded cryptography work produces intellectual property that may later earn licensing royalties. It frames the shift as a way to develop income beyond emissions. The roadmap still places the first commercial license in a future phase; no license sale, royalty receipt or new paying customer is disclosed in this update.
The proposed commercial model is relevant to long-term sustainability, while realized licensing revenue remains ahead.
SN28 SayGM adds an attested image-generation route
SN28 SayGM’s latest provider work adds FLUX.2-klein-4B image generation through a NEAR-attested route, alongside a DeepInfra option and capability-registry changes. Another update adds an SN53 Engy DeepSeek route. These are concrete additions to the serving code; the SN53 Engy text-model route alone is not presented as a new launch of the whole SN28 SayGM service.
An additional image capability expands what the service can route beyond the existing text-model offering.
SN15 ORO moves Race 2 to today at 3 PM ET, pending final readiness
The newest SN15 ORO team-channel digest moves Race 2 back 24 hours to September 14 at 19:00 UTC, or 3 PM ET. It reports continued incentive-mechanism testing, planned cooldown resets, and re-evaluation of affected work after a validator was excluded. Submissions are described as open, but grading and difficulty are changing. The race time remains a team-attributed schedule; no live restart was confirmed at this report’s earlier cutoff.
A one-day delay changes miners’ immediate plans, making the revised time and readiness status worth surfacing.
SN62 Ridges raises its second screener threshold to 92%
Today’s SN62 Ridges channel brief attributes a rise in the Screener 2 threshold to 92% to the team, citing stable scores from the leading agents. This supersedes yesterday’s reported 86% threshold and discussion of a possible 90% setting. The new value comes from the dated team-channel digest; the live qualification control was not independently inspected.
A higher threshold directly changes which miners qualify and updates yesterday’s participation rules.
Searches published headlines and summaries, not private bookmarks, unpublished leads or source-page contents. Duplicate headlines within an edition appear once; separate editions remain separate so follow-ups are preserved. Share the page URL to retain your filters.
