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 A past verified announcement can now be outdated.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.
SN34 BitMind reports a benchmark reset and fixes reward-result truncation
Today’s SN34 BitMind channel brief reports gasbench 0.9.8 with new holdouts, removal of the rendering category and a full benchmark reset. It describes a temporary upload closure during holdout preparation and a six-hour window for selecting new leaders; the exact reset schedule remains flexible. Separately, the public validator 5.0.8 change removes a 1,000-result cap and fetches the full two-hour reward lookback, preventing eligible older results from being displaced simply by newer activity. The reward fix is verified in code; reset timing is attributed to the channel brief.
Both the evaluation reset and result-window fix affect how miners qualify and receive rewards.
SN55 NIOME reports another scoring interruption after a validator crash
SN55 NIOME’s latest brief says a validator crash left only five miners scored on one task while others received zero. The team reports restarting the validator and investigating continuing anomalies. This is a new interruption after yesterday’s task-generator repair, so the earlier fix should not be read as resolution of every scoring issue.
Repeated scoring disruption is material to participants whose work may not be evaluated.
SN74 Gittensor separates access and health checks across leased compute
SN74 Gittensor merges one SSH port per machine with workload access over private addresses, alongside revised lease accounting. Related changes distinguish unexecuted checks from benchmark failures and allow a clean operating period to reduce accumulated strikes. These changes target access reliability and misleading penalties in its compute pool. They establish code changes, not measured fleet-wide uptime gains.
Access reliability and failure classification directly affect compute suppliers and customers.
SN76 Ormas jumps 73.6% in the rolling 24-hour snapshot
At 11:26 AM ET, SN76 Ormas alpha is up 73.6% against TAO over 24 hours. Its current metadata describes an outcomes API for coding work that is still in development, with miners paid for independently accepted delivery. The pool holds about 609 TAO in reserves, so the move should be read in the context of a small pool. No new official announcement was established as the cause. This is a timestamped market move, not proof of a product launch or a price forecast.
A large move deserves visible market coverage, while limited liquidity and no verified catalyst keep it below a five-star product story.
SN78 UMI targets its first evaluation batch while bridge weights remain
SN78 UMI’s channel brief targets closing its first evaluation batch and starting scoring today, subject to code cleanup. It says temporary bridge weights remain until evaluated results replace them and submissions will stay open for later batches. Exact cutoff instructions are still to be published. This is a planned transition, not confirmation that the first results have already set on-chain rewards.
The handoff from temporary weights to evaluated results is material to miners awaiting scoring.
SN81 Reliquary avoids duplicate checkpoint staging for remote proofs
SN81 Reliquary’s new remote-proof path avoids staging model checkpoint weights unnecessarily when proof work is handled remotely. Companion environment changes add a per-turn token budget and make telecom task state serializable. The main operational benefit is avoiding redundant model-weight handling; a realized speed or cost reduction was not independently measured.
Reducing repeated checkpoint handling is a substantive training and proof-infrastructure improvement.
SN90 KubeTEE reports two image-generation backends live behind its gateway
SN90 KubeTEE’s deployment documentation now reports end-to-end checks for two Michigan image backends: FLUX.2 Klein 4B and Cosmos 3 Super. It documents protected infrastructure and mutual TLS between services. Klein is exposed through an OpenAI-style image-generation route, while Cosmos uses its native inference schema through a separate passthrough. This follows yesterday’s attestation news with specific model-serving infrastructure. The live tests are maintainer-reported; this scan did not submit image requests or independently audit the security design.
Named model backends and documented gateway paths are a concrete product step beyond a general infrastructure roadmap.
SN91 Cascade tightens GPU comparability for challenger evaluations
SN91 Cascade now keeps a manifest’s evaluations on the same GPU type and requeues challengers when that condition cannot be met. Related changes let available lanes continue work past a deadline without unnecessarily holding the incumbent. These are evaluation-scheduling changes intended to keep comparisons consistent while avoiding blocked capacity, not a new model-quality result.
Comparable hardware and clearer scheduling matter when competition results determine the winning model.
SN100 Cortex merges a new Bounty reward profile; activation is still a separate step
SN100 Cortex’s new algorithm assigns up to 30% of emissions to Bounty and 70% to Proof. Each valid Bounty report earns one point, authors share proportionally and ten valid reports unlock the full Bounty share; unused allocation burns. The code explicitly requires an owner-signed profile and coordinated activation before these rules apply. At this scan, the public network page still says automatic payouts are not wired and weights are not sealed, and its older scoring description differs from the new repository rules. Report intake and a merged reward design therefore should not be confused with confirmed live 30% payouts.
The reward design is material, and the activation requirements prevent a premature claim that miners are already receiving the new split.
SN105 Beam tightens transfer-integrity enforcement and warns of stale dashboard data
SN105 Beam’s channel brief attributes permanent UID and worker bans for false transfer chunks to the team. It also says dashboard worker-linkage data may be stale and advises operators to consult local data. This report does not identify or accuse individual miners; the operational news is the enforcement policy and the dashboard limitation.
Transfer integrity and unreliable operator visibility are practical issues for a file-routing network.
StonkFun reports $1.25M revenue, $750K buybacks and $60M cumulative holder rewards
StonkFun’s September 19 operating report lists $1,247,924 in revenue, $749,636 in buybacks and 2.93 million STONK burned. A separate post reports cumulative holder rewards reaching $60 million. These are fresh, project-reported platform totals. They are not independently audited accounts, TAO-only revenue or TAO token burns, and today’s report is not a new announcement of the previously covered TAO pairing.
Fresh operating figures update the ongoing StonkFun watch without conflating platform-wide activity with TAO-specific economics.
Emission watch: tiny allocations return to SN71 Leadpoet and SN76 Ormas
At 11:26 AM ET, SN71 Leadpoet has 0.000239% observed network allocation and SN76 Ormas has 0.000001%, both above zero after zero in yesterday’s saved snapshot. SN92 and SN58 remain at zero. These are small snapshot changes, not confirmation of meaningful miner payouts. Network allocation, voluntary burns and actual distributions are different measurements; the data does not establish that Const, Arbos or another named person manually switched these subnets on.
The new zero/nonzero changes continue the emissions watch, with scale and unknown causes made explicit.
BaseStonk reports more than 1,700 launches
BaseStonk’s latest post reports more than 1,700 launches, updating the platform’s earlier count. The number is project-reported and platform-wide; it does not establish that those launches use TAO or that each has meaningful trading activity.
A new dated platform count belongs in the ecosystem watch with its scope stated clearly.
SN28 SayGM and NEAR AI confirm live confidential inference integration
NEAR AI Cloud confirms that its confidential inference is available through SN28 SayGM’s confidential tier. The announced design protects both steps: SN28 SayGM routes inside an Intel TDX enclave, while NEAR AI runs models in protected CPU and GPU environments. For developers handling private code or customer data, this adds a named external provider and a public verification path to the subnet’s product. NEAR publishes instructions for obtaining attestation without an account. Its documentation also states the limit: an environment report alone does not prove that a particular gateway request reached that environment. The confidential tier matters; the same protection does not automatically apply to every frontier or open model in the catalog.
A partner-confirmed live integration expands the subnet’s commercial product and provides inspectable security evidence, without implying a completed independent security audit.
SN66 Conjectures publishes a new proof claim for Erdős Problem 416(i)
SN66 Conjectures reports a solution to the first part of Erdős Problem 416, a 52-year-old question about the number of distinct values taken by Euler’s totient function. The claimed result establishes the doubling limit V(2x)/V(x) → 2. The team links a public Lean proof and says it passed verification. This is a new result after the Problems 108 and 653 claims covered yesterday. It is specifically part (i), not a claim to settle every question on the Problem 416 page. The independent Erdős Problems directory still marks the problem open at this scan; the proof was not independently rerun for this report.
A new longstanding mathematics result with public formal-proof material deserves lead coverage; importance is separate from both market direction and independent mathematical acceptance.
SN114 SOMA confirms live inference routing through SN28 SayGM
SN114 SOMA says it now compresses context and routes agent requests through SN28 SayGM, with real traffic already flowing. SN28 SayGM independently confirms that it is providing inference for the integration. This is another concrete example of one subnet using another’s output in a product. Neither announcement quantifies traffic, customer revenue or realized cost savings, so those are not implied by the launch. Separately, today’s channel brief schedules SN114 SOMA’s next competition for September 22, with DeepSeek V4.1 Flash and IssueScout tasks; those competition changes are still upcoming.
A live integration confirmed by both participating teams is materially stronger than a proposed partnership or roadmap teaser.
SN11 TrajectoryRL shifts competition to routing policies for SN53 Engy
SN11 TrajectoryRL’s Season 2 release changes what miners build: routing policies that choose which SN53 Engy models answer each request. The release retains the existing scenario harness and verification process while metering model calls and enforcing a spending cap. Updated project documentation says the policy holding the winning seat is served to SN53 Engy customers as engy/auto, connecting competition output to an inference product. The code and documentation are public; customer requests were not independently exercised. The broader recursive self-improvement platform remains a roadmap, not an achievement established by this release.
A new competition mechanism with a documented customer-facing use for the winning artifact is a significant product and incentive change.
SN105 Beam introduces private Rooms; new coverage details a 100 GB transfer in 70 seconds
SN105 Beam introduces Beam Rooms for private transfers across peers, cloud buckets and connected services. Fresh TAO Daily coverage of its demonstration reports 100 GB moved in roughly 70 seconds using parallel routing, while the Opentensor account points viewers to the full demonstration and a Beam Studio preview. The key distinction is the measured example versus the projection: the cited demonstration is 100 GB in 70 seconds; a 1 TB transfer in under two minutes is the team’s extrapolation, not the demonstrated result. These performance claims were not independently rerun.
A new product surface and a concrete large-file demonstration make the infrastructure easier to assess beyond a general speed claim.
SN88 Investing says RootReborn weight setting is disabled ahead of a redesign
SN88 Investing’s account says RootReborn weight setting has been disabled on-chain and is to be replaced with active rebalancing. Its description says root validators fall back to passive flat baskets for now. The immediate news is the reported disabled mechanism and interim behavior; active rebalancing is the proposed replacement, not confirmed as already operational. This concerns the named root-basket mechanism, not a claim that Bittensor staking or all subnet emissions have stopped. The operational claim is attributed to the project and was not independently checked against the relevant on-chain setting.
A reported interruption and redesign of an allocation mechanism is important for participants; current behavior must be distinguished from the replacement plan.
SN90 KubeTEE adds customer-facing attestation and publishes its key-broker endpoint
SN90 KubeTEE has merged an attestation endpoint and a verification guide for API clients. Requests can carry a fresh nonce, with evidence exposed separately or alongside a chat response. A companion change publishes the key-broker endpoint and reports testing the attestation exchange with a real TDX guest. These are concrete steps toward allowing users to inspect the infrastructure serving them. They are verified repository changes and project-reported testing, not an independent security audit or proof that every generated answer is correct.
Customer-accessible attestation moves confidential infrastructure toward a verifiable product, with a public implementation and guide.
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.
