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 A past verified announcement can now be outdated.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.
SN80 OpenRoboto unveils OR-S1 with General Intelligence Labs
Following its Shift launch, SN80 OpenRoboto unveils OR-S1, a stereo egocentric capture device for robot-manipulation training data, and names General Intelligence Labs as the hardware partner. The partner’s quoted statement confirms support for the combination of its devices and SN80 OpenRoboto’s collector network. Preorders are linked through Shift. This is a hardware and data-collection announcement, not evidence that devices have shipped or a disclosed volume of customer sales.
A named hardware partnership gives the robotics-data network a concrete collection product.
SN80 OpenRoboto fixes rejection of valid AXIS checkpoints
SN80 OpenRoboto CLI 1.5.0 chooses the checkpoint layout from the active season’s benchmark. The commit explains that valid AXIS checkpoints had been checked against a LIBERO layout and rejected before submission. The fix targets that specific miner blocker; users still need the updated client, and live acceptance was not tested here.
A concrete submission blocker is fixed for affected robotics miners.
SN81 Reliquary ships layer-by-layer model validation
SN81 Reliquary says its new validation system streams model layers instead of loading the entire model at once. The team reports that a single H100 can now verify a frontier mixture-of-experts model with results identical to a full-model run. That could lower the hardware barrier to validating larger models. The upgrade is team-announced; the bit-identical result has not been independently reproduced here, and larger future training runs remain a roadmap statement.
A significant validation change could expand the scale of models the subnet can support.
SN75 Hippius becomes Zapme Social’s distributed storage layer
SN75 Hippius says Zapme Social now stores and delivers its content through the subnet project’s S3-compatible distributed storage. Its September 24 case study describes a staged migration from AWS covering a community of more than 6,000 members. This is a concrete external product integration rather than another storage-roadmap announcement. The migration and community size are provider-reported; separate customer confirmation and revenue figures were not established in this scan.
An external application using the storage product is meaningful evidence of utility.
SN10 Pareton reports 23–54% more output throughput in a controlled benchmark
SN10 Pareton tested its leading Qwen3.8-27B miner against the baseline on the same GPUs using NVIDIA’s AIPerf tool. The team reports all 600 requests completed and output throughput improved by 23–54% across tested loads. This offers specific evidence for its inference-optimization work. It is a project-run benchmark using an NVIDIA tool, not an NVIDIA endorsement or an independently reproduced performance result.
A quantified same-hardware comparison gives a concrete measure of the subnet’s optimization output.
SN49 Nepher reports 337 robot-navigation submissions across 90 hidden mazes
SN49 Nepher says about 97 miners competed in Leatherback LiDAR Real Maze Phase 2, submitting 337 policies tested across 90 hidden mazes. The robot navigates using 360-degree LiDAR without a camera or map, and the team publishes footage of the winning policy. Separately, it says Nepher Verified is selectable on a major robotics-listing platform, but the announcement does not name that platform. The competition figures are team-reported, and the listing claim is not equivalent to regulatory certification.
Concrete competition scale and a demonstrated navigation task are stronger evidence than a generic robotics teaser.
SN18 Zeus reports faster weather runs and an August temperature advantage
SN18 Zeus reports a new weather run every 1.5 hours, compared with six hours for IFS and seven for WeatherNext 3 in its comparison. It also says Zeus Pro was more accurate on two-meter temperature throughout August. The claim concerns the team’s chosen comparison and measured variable; it does not establish superiority for every weather target, geography or forecast horizon. The benchmark has not been independently reproduced here.
A new performance comparison matters for a subnet selling usable weather intelligence.
SN8 Vanta launches a promotion path to accounts of up to $1 million
SN8 Vanta’s Pro offering is live. Eligible $50,000 and $100,000 Classic accounts can be invited into a path toward up to $1 million in trading capital, keeping their chosen reward split without an additional promotion fee. The published rules require a challenge of at least 90 days, performance targets and stricter risk controls; promotion is discretionary. Separately, the team says a dedicated buyback wallet is being prepared and first evaluation-fee buys are coming soon. Those planned purchases are not reported as completed.
A live customer offering materially expands the trading program, while eligibility and buyback timing remain explicit.
SN114 SOMA adds Codex to its context-compression workflow
SN114 SOMA announces Codex availability with its context compressor, aiming to reduce coding-token expenditure. Claude Code is identified as the next integration rather than a live feature. This follows yesterday’s separate SN53 Engy integration. The announcement establishes a product addition; it does not disclose independently measured savings or an official partnership with OpenAI.
A new coding workflow broadens the compressor’s practical use.
SN118 Ditto: channel report describes temporary 100% reward burn
SubnetRadar’s September 24 channel brief attributes a temporary 100% burn setting to SN118 Ditto’s owner while the team works through source-integrity review and screener bugs. It also reports a ten-hour upload cooldown. This is a secondary account of channel messages, not an independently verified current burn setting, and no reopening time is established. Repository changes separately repair adjudication and benchmark-promotion checks. Do not confuse a local reward-burn policy with the network’s subnet allocation: the current market snapshot reports zero allocation, but does not prove the cause or duration of that state.
A reported interruption to miner rewards deserves prominent attention even while live-state confirmation remains incomplete.
SN13 Data Universe reports roughly 100,000 September requests so far
SN13 Data Universe reports approximately 100,000 on-demand requests in September to date, averaging about 4,600 per day, with 97.8% answered by miners. These are usage figures released on September 23, not a completed-month total, independently audited customer count or disclosed revenue.
A quantified demand and response-rate update helps assess actual subnet usage.
SN21 AdTAO publishes a third historical training dataset
SN21 AdTAO adds 7,132 episodes from August 5–14 to its training-data release. Its earlier V2 dataset contains 29,129 episodes; that larger count is not the size of today’s addition. The documentation keeps historical training data separate from future held-out evaluation, helping miners improve without confusing a public training set with the scoring set.
A substantial new training release changes the material available to competing agents.
SN28 SayGM adds verification support for Chutes TEE routes
SN28 SayGM’s repository adds Chutes TEE verification support and advances its stable release to 0.4.21. This is an integration and verification change, distinct from yesterday’s model-catalog additions. The code distinguishes ordinary TEE routes from attested paths, so the update should not be interpreted as proof that every request is end-to-end encrypted or independently audited.
Verification support is relevant to users choosing trusted inference routes.
SN44 Score: reported weight-setting error followed by recovery
SubnetRadar’s September 24 channel brief says a database read error during weight-setting briefly affected SN44 Score’s reward distribution, and attributes a recovery update to team member mxmsbt. The 10:52 AM ET screener shows about 4.80% network allocation. That snapshot supports the current allocation level but cannot reconstruct the reported intraday fall or establish its cause. The team is also reported to be reviewing detection-label and scoring issues. This is an operational recovery report, not evidence of a deliberate protocol-wide emission switch.
A material scoring disruption and recovery warrant coverage with the secondary-source limitation stated.
SN51 Lium prepares scoped API keys and spending budgets
SN51 Lium adds client-side support for scoped API keys and budgets in its CLI and SDK. The change explicitly notes that server support is pending, so these controls should not yet be described as enforced on the live service. The work points toward more constrained programmatic GPU access once the backend is ready.
Access scopes and budgets are meaningful product work, with live availability still pending.
SN68 NOVA updates molecular candidate constraints
SN68 NOVA’s repository expands the checks applied to candidate molecules, including entropy and atom-count constraints. This changes what submissions can qualify for scoring and is relevant to miners preparing candidates. It is an evaluation-rule update rather than a new drug discovery or experimentally validated medical result.
Eligibility changes affect miner submissions and interpretation of the competition.
SN79 MVTRX puts v0.6.2 on testnet; mainnet scheduled for later today
The September 24 channel brief reports SN79 MVTRX v0.6.2 on testnet, with automated calibration replacing manual selection for its market configuration. Published scoring changes reduce credit for narrow book coverage and extend counterparty discounting. Mainnet is scheduled for 16:18 UTC, or 12:18 PM ET, September 24—after this edition’s research cutoff. A scheduled rollout should not be read as already deployed.
Testnet availability advances yesterday’s proposal and gives miners a concrete rollout time.
SN97 Albedo moves from binary judging to graded feedback
SN97 Albedo’s scoring change replaces a binary judgment with twenty levels, using probability-weighted model outputs to give partial credit. That can distinguish incremental improvements that previously received the same pass-or-fail result. The repository implementation is available to inspect; operator deployment and effects on rankings were not independently checked.
A scoring redesign changes how miner improvements can earn recognition.
SN100 Cortex receives more than 250 bug reports in 48 hours
SN100 Cortex says ten miners submitted more than 250 bug reports in the first 48 hours of its Bounty Challenge. The team explicitly says the reports are under review. The count measures submissions, not 250 confirmed vulnerabilities or accepted paid bounties.
Early participation offers evidence of activity, with reported bugs distinguished from validated findings.
SN120 Affine tightens safeguards against leaked coding answers
SN120 Affine adds restrictions on GitHub access during the coding-agent phase and filters contaminated training prefixes. The goal is to limit access to upstream answers that could undermine evaluation. This is a preventive evaluation change; it does not establish misconduct by any particular miner or prove all leakage paths are closed.
Protecting the evaluation from answer leakage matters to the integrity of rewards.
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.
