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 A past verified announcement can now be outdated.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.
Verathos links alpha conviction to daily inference credit
Verathos introduced Conviction Allowance: users who lock SN96 alpha receive daily model-usage credits while the alpha stays locked. The team says credits start after the lock is confirmed on chain and their amount changes with the alpha price. This connects holding and using the subnet’s token with access to its model lineup. The announcement does not establish a fixed dollar return or a permanent inference allowance independent of the lock.
This creates a direct product benefit from alpha conviction rather than relying only on speculative demand.
Cortex opens mining; iOS remains in Apple review and payouts have limits
Cortex’s September 11 update says mining is open and its iOS app has entered Apple’s review queue. Its public network site lists Proof and Bounty as accepting submissions, but also says weights are not yet sealed, judging remains partial and automatic payout is not wired. Unmatched emissions burn in the meantime. The team targets next week for the full application’s production rollout, a DeepSeek V4.1 upgrade and a launch campaign with a 14-day trial. These are distinct stages: submissions are open; App Store approval and the planned production rollout are still ahead.
Opening participation and moving the mobile app into review are material milestones. The live site’s payout limitations are essential context for miners.
NOVA reports a 7.02× inference gain and opens Boltz-2 testing
Metanova reports that NOVA’s latest Blueprint winner increased PSICHIC throughput from 214 to 1,502 protein–ligand pairs per second on an RTX 3090. That is a 7.02× result for the stated benchmark, not a demonstrated improvement in clinical outcomes. The team also opened the Boltz-2 integration for testing: scoring moves to a dedicated GPU oracle that competitors can query during their search. The September 10 announcement says the competition goes live this week, so test access alone is not proof of completed activation.
A measured improvement in a scientific workload and a more capable scoring model are substantial research and competition developments.
SayGM adds DeepSeek V4.1 Flash—and the discount is now 50%
DeepSeek V4.1 Flash is available through SayGM. The initial announcement advertised 5% off, but Mog posted a correction to 50%, and the live price table now agrees. At this morning’s check, SayGM listed $0.15 per million input tokens, $0.003 cached input and $0.60 output. The gateway describes this model as routed to an external provider, whose API terms still apply. It should not be confused with a model running entirely inside SayGM’s confidential-compute tier.
A new model plus a verified pricing correction directly affects users’ costs. Checking the live table avoids publishing the already-outdated launch discount.
Late-found September 9 news: Almanac crosses $10M in trading volume
Almanac announced on September 9 that it had processed $10 million in cumulative trading volume and recorded its first $200,000 day. TAO Daily highlighted the milestone today, but the original event is two days old and is included here as a clearly dated catch-up. The figures are reported trading volume, not platform revenue, profit or a new enterprise contract.
A concrete adoption milestone deserved coverage. Its original date and the meaning of volume are retained so a recap is not mistaken for a new announcement.
Epago adopts a 95% burn rule while its observed allocation stays at zero
Epago’s new competition configuration assigns 95% of miner emissions to a burn destination and the remaining 5% to the king, replacing its earlier variable schedule when that configuration is used. Another change requires private checkpoint submissions so competitors cannot copy unpublished work. At 10:07 AM ET, SN36’s observed subnet allocation is still zero and its alpha is down 22.58% against TAO over 24 hours. The repository changes do not themselves reverse the administrative disable documented yesterday, and a proposed reward split is not evidence that miners are receiving those rewards.
The emissions and reward rules matter immediately after yesterday’s surge and confirmed shutoff. The new configuration is distinguished from an actual resumption.
Witness publishes its mainnet validator and scorer v1.1.0
Witness’s public release now provides a CPU-only validator preset for Finney SN20, with no model-provider API key required for validator setup. The follow-up scorer v1.1.0 adds signed round feedback and the current documentation specifies a 70% burn / 30% winner allocation. When nobody has a positive qualifying reward, the allocation burns in full. This is a concrete software release beyond yesterday’s research-preview description; independent benchmark quality and the competitiveness of submitted miners remain separate questions.
A working mainnet participation path and an explicit reward contract materially change the subnet’s readiness.
Apex sets September 14 for its next humanoid challenge
Apex announced Box Scramble for Monday, September 14. A simulated Unitree G1 must reach a raised exit by moving and climbing loose crates, with their sizes, weights and positions changing each round. The task expands control from the earlier 12-joint legs-only setup to 22 joints with working arms. It tests manipulation and planning as well as locomotion; it is a scheduled competition, not an already completed robot deployment.
A more demanding, clearly dated robotics challenge gives miners a concrete next target.
Allways improves miner setup and rewards only active swap pairs
Allways added a guided funding and setup flow, shared EVM-key handling and checks for whether the miner container is actually running. Its reward change assigns direction-pool shares only to pairs that cleared qualifying volume during the relevant window. Registering inactive pairs therefore does not dilute the active-pair floor. The changes are in the public repository; they are not evidence of a measured increase in bridge volume.
Participation becomes easier and the reward mechanism is more closely tied to swaps that actually occur.
Macrocosmos will present Project Orion’s summer results at Exploit
Macrocosmos says it will show the results of its summer work on Project Orion at Exploit Summit, addressing whether distributed AI training can compete commercially. The September 10 post is a preview of the presentation. It does not disclose a new paying customer or establish the project’s commercial performance in advance.
A relevant upcoming research presentation is worth tracking, with a lower rating until the results are public.
ORO brief reports submissions reopening after the benchmark migration
The September 11 channel digest reports that submissions reopened for the updated benchmark after backend fixes, with an initial 30% qualifying threshold and later dynamic cap. Public changes also address model-output and local testing failures. The original Discord discussion was not independently accessed, so the reopening and race details remain developing. The zero allocation in this morning’s separate emissions snapshot does not settle whether the submission service is functioning or identify an administrative intervention.
A reported recovery after a submission pause matters to miners, with service status kept separate from token allocation.
Blockmachine fixes a startup failure that could cost validator rewards
Blockmachine validator v0.2.13 fixes a path where a failed startup request could leave a validator using unusable storage defaults and missing rewards. Startup configuration now retries, while refreshed storage settings can recover instead of remaining stuck. The release notes identify an affected September 10 epoch. The patch addresses the failure mode; it does not establish compensation for any missed dividends.
A documented failure affecting validator earnings warrants a practical release notice.
Endure explains the eight values scored in its alpha-risk schema
Endure’s new explainer specifies four risk outputs—maximum drawdown, realized volatility, time-weighted average price and liquidity depth—over both five-day and thirty-day horizons. Missing any of the eight combinations causes a submission to be rejected before scoring. It is a clarification of the current Alpha Risk V1 contract rather than a newly announced customer launch.
A precise explanation helps participants understand what constitutes a valid submission.
Ralph replaces the random judge draw with multi-judge scoring
Ralph’s update has every available judge evaluate the same generated answers in each round, then averages their protected slice scores. Previously, the drawn judge could change a close contest’s outcome. The update also reduces the dethroning margins after a long reign. These are competition-rule changes in the public implementation, not a claim that a new champion has already been selected.
A change to judging and incumbent protection materially affects how miners compete.
Score outlines the specialist models coming after Studio
Score says Studio is about to go live and will give users access to SN44’s vision models. Its new roadmap lists intrusion, fire and smoke, license plates, vehicles, crowd density, falls, abandoned objects, violent behavior, weapons and thermal/night imagery. This expands the planned product range beyond the access announcement covered yesterday. The post does not confirm that Studio or all those specialist models are already available.
A broader, concrete product roadmap shows the intended customers and workloads, while its launch status remains explicit.
Lium simplifies miner onboarding and adds workspace-aware tools
Lium’s CLI now documents a single registration command that takes a bare host through setup to a listed node using a portal token. Workspace-aware commands and API keys also give teams an explicit billing and access context. A separate rental-security change keeps the encrypted-volume passphrase out of the provider’s disk and process arguments. These are new operational improvements alongside the billing and lending announcements already covered yesterday.
Simpler supply onboarding and clearer team access help the compute service operate at greater scale.
A coldkey swap is scheduled for the wallet owning SN58
SubnetRadar records a coldkey-swap scheduling event at block 9,039,596 on September 10 at 20:42 UTC, from 5CLUz…9gtJ to 0x572…682e. Its follow-up identifies the wallet as owning SN58 greevils and notes a mandatory delay of about five days. The event is scheduled, not completed; it does not by itself establish a subnet sale or the identity of a buyer.
An upcoming owner-wallet change is material, but the chain event must be described at its actual stage.
Conjectures merges a new 15-day bounty curve
A September 11 merge changes the default mathematical-bounty policy to start at 10% of the uncommitted SN66 alpha treasury and increase to 12.5% over 15 elapsed days. The quoted alpha amount still changes with available treasury funds and outstanding commitments. Existing submission locks and ages are preserved. New formalization-defect awards are capped at the lesser of $750 converted to alpha or the submission’s locked bounty. The repository requires deployment and a policy-version check; that live activation was not independently confirmed.
A material change to reward sizing and review terms affects future submissions, while preserving the distinction between merged code and deployed policy.
Harnyx widens the main contest for challengers that beat the incumbent
Harnyx announced broader access to its main round for artifacts that beat the champion. The corresponding implementation compares qualifying challengers with the incumbent and caps the main field at 30. That cap matters: the announcement should not be read as an unlimited-size final. The change broadens the route to the title while retaining a bounded evaluation field.
The rule directly changes how qualifying work reaches the championship round and should be visible to competitors.
Gittensor reports faster Qwen3.8 generation on the same RTX 5090
Gittensor reports 25% faster generation and 43% more context capacity than Nvidia’s NVFP4 build of Qwen3.8-27B on the same RTX 5090, with comparable accuracy. Its repository also records a separate packed-kernel improvement at eight concurrent requests. These are workload-specific project benchmarks; the separate speedups should not be multiplied or presented as a universal performance advantage.
Same-hardware measurements make an inference-efficiency claim more useful, while its scope remains tied to the tested model and workload.
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.
