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 A past verified announcement can now be outdated.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.
Swarm confirms private submissions while winning models still become public
Swarm’s September 7 announcement confirms that submissions now remain private while competing. When a model becomes champion, its full source and weights are published. This provides primary confirmation for yesterday’s developing rollout report. The change protects challengers during evaluation while preserving public access to the winning work.
A confirmed visibility change affects copying risk, miner participation and the availability of winning models.
NOVA plans to move Blueprint scoring to Boltz-2 this week
Metanova says NOVA Blueprint is transitioning from PSICHIC to Boltz-2-based scoring this week. The project expects the structural-biology model to provide a richer evaluation layer, at higher inference cost. That makes efficient search more valuable: algorithms should narrow the candidate space before the expensive evaluation. The announced transition is planned; a completed activation was not established.
Changing the core scientific scoring model affects both research quality and the economics of the search competition.
AdTAO expands its reported reliable predictions to seven change types
AdTAO’s updated dashboard results rate seven of nine Google Ads change types as RELIABLE over seven days, up from six four days earlier. Four change types also meet its RELIABLE rating over fourteen days. These are the project’s own evaluation labels and reported results, not an independent guarantee of advertising performance. The update adds longer-window evidence to the previously announced prediction product.
Measured improvements and a longer evaluation horizon provide more useful product evidence than a generic growth claim.
Follow-up: Endure opens a testnet for alpha-collateral risk recommendations
Endure’s September 7 announcement, posted before yesterday’s cutoff and added here as a late follow-up, says testnet 504 is live. Miners submit structured alpha-collateral risk recommendations, and validators evaluate them against realized market behavior using the risk.v1.subnet_alpha schema. Miner and validator documentation is linked in the original announcement. This is the SN30 project’s testnet, not a claim of mainnet activation.
A working testnet gives the risk-intelligence project an actual participation path. The original date and testnet status are explicit.
Score’s Astra Ultra replay shows why F1 and subnet rewards can diverge
Score tested Astra Ultra on one 30-second football clip and reported F1 of 76.92%, but zero under SN44’s published scoring formula after timing adjustments and false-positive penalties. Its breakdown lists 22 predictions, seven extras and two misses. Score explicitly calls this one offline replay, not a live-latency test or a season-long benchmark. It is a useful example of the cost of invented rare actions under the subnet’s scoring rules.
A detailed, reproducible scoring breakdown explains what the specialized evaluation rewards. The narrow sample does not establish broad model superiority.
Lium benchmark follow-up: 256 streams explain the 14,499 tokens-per-second figure
The full September 6 thread supplies the conditions missing from our previous edition: Qwen3.6 35B FP8, vLLM 0.28.0, one B200 with 192GB, and 256 concurrent streams producing 512 output tokens each. Lium links its LocalMaxxing run and a reproduction repository. Its cost calculation is $5.60 per GPU-hour divided by about 52.2 million output tokens per hour, or roughly $0.107 per million. This is aggregate throughput under the stated workload, not one user’s generation speed.
Adding the original workload and cost methodology corrects a material gap in yesterday’s coverage of a widely shared result.
Vanta publishes scoring explanations generated from its live code
Vanta announced a transparency page that renders scoring rules and their changes in plain English directly from its open-source repository. The page provides a public way to inspect the competition’s logic instead of relying only on informal explanations. It is a transparency feature, not a change to the network’s emission status.
Making scoring rules easier to inspect improves the ability of participants to audit the competition.
Iota channel reports a 950 TAO burn of excess emissions
Today’s team-channel brief attributes a manual 950 TAO burn to macrocrux, describing it as excess miner emissions. The specific transaction was not independently reconciled during this scan. Keep that reported TAO amount separate from circulating alpha-burn alerts, which use different units. Claims that reward concentration proves misconduct are not established by the available evidence.
A reported burn of this size warrants coverage, with the transaction-verification gap and token units explicit.
Pareton adds score transparency and a two-day delay before patch publication
Pareton merged per-prompt score reports, including explanations for zeroed results, and a two-day visibility delay for newly ingested patches after their first finalized scored or disqualified evaluation. Hashes, status, scores and logs remain public; older submissions keep their existing links. Its September 7 benchmark announcement also said the team was holding while maintaining and recalibrating, a detail omitted from yesterday’s summary. A competition restart was not confirmed.
The new audit detail and delayed source disclosure affect how miners understand results and protect their work.
Mantle confirms Blockmachine’s RPC support from the receiving side
Mantle’s official account shared Blockmachine’s Mantle RPC offering on September 7, highlighting request-based pricing and low-latency reads. This adds partner-side acknowledgment to yesterday’s provider announcement. It confirms recognition of the infrastructure offering, without disclosing a commercial contract or independently validating the provider’s savings claims.
Confirmation from the supported chain strengthens the evidence behind yesterday’s infrastructure announcement.
Trishool targets a small, faster streaming-output guard
Trishool says its current adoption work prioritizes latency and that its streaming-output guard is being built at the same 0.8-billion-parameter scale as HaloGuard 1.0. The aim is safety filtering that can run on laptops, devices and edge servers with lower compute requirements. This is a development update; the post does not announce a completed streaming-guard release or a new OpenAI agreement.
The update explains a concrete product direction and its deployment target without repeating the earlier program-acceptance story as new news.
NIOME replaces one-shot benchmark scores with averages across multiple seeds
NIOME says every task now broadcasts multiple seeds and validators benchmark each submission several times, using the average as its final score. The change is intended to reduce lucky one-off results and make miner comparisons more reliable. The project also says it has moved to Bittensor v11.
A confirmed change to benchmark design directly affects the reliability of miner rankings.
Conjectures expands its published task pool and implements account-based submissions
The validator repository now documents 259 reviewed theorem targets and 518 proof/refutation bundles, alongside a 10MiB proof policy. A separate implementation lets an authenticated website account spend a submission credit without a Bittensor signing key. The repository contains explicit release and deployment stages; these source changes are not proof that the expanded pool or every submission route is active in production.
More reviewed targets and a simpler submission route broaden the potential research and contributor pool.
SubnetRadar flags a pending SN73 owner-key rotation
SubnetRadar reported a pending owner coldkey swap for SN73 Parked on September 7, with an approximately five-day delay. This is an operational watch item: the alert does not establish that ownership has already transferred, identify the people behind the keys, or announce a new product. The underlying chain transaction was not independently retrieved.
A pending control-key change deserves a watch entry while confirmation and its effective date remain incomplete.
Gittensor updates its serving release to Qwen3.8-27B
Gittensor merged the Qwen3.8-27B serving release, updating the miner checker and published deployment configuration to the new model. The setup supports pinned model-directory artifacts as well as the earlier single-file path. This is new serving infrastructure, separate from previously covered model-download milestones.
A concrete serving configuration changes what operators can deploy and verify.
Catch-up: Phylax deregistration wipes a 56,076-alpha conviction lock
A finalized-chain event at block 9,008,717 shows that the former SN76 Phylax slot was deregistered on September 6 at 1:37 PM ET, wiping 56,076.232 locked alpha. SubnetRadar says the locked stake was converted to TAO pro rata through the AMM and returned to free balance. This was a deregistration exit, not a scheduled cliff, and it is not evidence that the returned TAO was sold. The event fell inside the September 7 edition's coverage window but was missed there, so it is labeled as a catch-up.
The forced release exceeded the report's 50,000-alpha threshold and materially affected holders in a deregistered subnet slot.
UMI opens its public endpoint pilot to all eligible registered miners
UMI’s updated pilot rules allow every verified SN78 miner hotkey without a validator permit to enroll; there is no exclusive pilot slot. Each miner can complete one case after declaring readiness. The test checks a public endpoint against a known clip and publishes replayable evidence. It remains a component test with no translation-weight activation, ranking significance or ongoing traffic implied by enrollment.
Wider pilot access is a real participation change, with clear limits on what the test demonstrates.
Vidaio discussion reports a one-slot rule for identical decoded content
The September 8 channel brief says schema v17 went live at epoch 24978 and gives identical decoded content one scoring slot regardless of host or coldkey, with copies receiving duplicate-content zeros. It also reports fixes for clips with non-standard frame rates. The brief does not establish which statements came from authorized operators, and a matching primary release was not retrieved, so the rollout remains developing.
A duplicate-content rule could materially change miner scoring, but the available attribution is insufficient for a verified activation claim.
Cascade changes duel seating to include all screened entrants
Cascade’s new default seats the full screened field in duel-only rounds, queuing extra legs when the available GPU fleet is short. An explicit field cap can still override this behavior. The change follows a documented round where only two of five entrants were seated and extends yesterday’s dashboard transparency work into actual scheduling rules.
The rule changes who gets evaluated in the current round when compute capacity is constrained.
MicroTensor measures full-request cost and rejects empty-output profiles
MicroTensor’s mechanism 0.4.0 code adds total-task latency alongside time to first token. Requests that produce no output are failures, and an entirely empty profile faults the artifact instead of recording zero cost. Coordinator changes also defer the assignment plan until the freeze. These address the scoring and scheduling issues discussed in yesterday’s edition; publication of the fixes is confirmed, while adoption by every validator was not measured.
The changes close an important measurement loophole and make cost comparisons better reflect completed work.
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.
