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 A past verified announcement can now be outdated.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.
Gittensor’s SparkInfer 0.5.4 adds image input and reproducible benchmarks
SparkInfer 0.5.4 adds image-input support and longer-context DSpark benchmark coverage. The team also committed the workload corpus and measurement harness after acknowledging that previous published benchmark numbers could not be reproduced from the repository alone. This is a new verification and capability update following yesterday’s compute-serving launch.
Reproducible measurements and broader input support make performance claims easier to assess.
Hippius bridge update awaits runtime rollout after a merged fix
The team points to Subtensor PR 3140 for a forthcoming bridge-related update. The fix, merged September 4, restores contract-proxied stake transfers affected by a runtime filter. Merge status is confirmed, but runtime activation and restored Hippius bridge operation were not confirmed by this edition’s cutoff.
Bridge availability matters for fund movement and miner onboarding; merged code is only part of the recovery.
UMI follow-up: reference implementation is available, translation rewards are inactive
Following SN78’s pivot from Vocence, UMI has published its reference implementation, baseline model and inspectable miner pilot, including Apple Silicon support. The repository explicitly says public calibration has not started and translation weights remain inactive. Component review and offline rehearsal are open; this is not confirmation that the new translation competition is paying rewards.
The implementation gives miners something concrete to inspect, while inactive weights materially limit launch status.
MVTRX updates its trading interface; new scoring remains under review
The team reports a redesigned swap page, better chart controls and an expanded Hangar for local testing ahead of miner grants. Version 0.6.1’s earnings-sensitive scoring change is still being tested and has not been confirmed live. The α/TAO market snapshot showed a 27.65% 24-hour decline, making the rollout and pending scoring decision material to watch.
The sharp market decline and a pending change to miner earnings make rollout status especially consequential.
OpenRoboto clarifies the season’s grasp-and-lift baseline
A September 5 documentation change specifies a green-block grasp-and-lift baseline, including object dimensions and the requirement to lift clear of the table. Placement is not required for this baseline. The locked season task specification remains authoritative, giving miners a more precise target for local testing and submissions.
A clearer physical task definition reduces ambiguity in robotics training and evaluation.
Compelle’s King of the Hill streak reset highlights the Elo-gap rule
The team discussion describes an 18-block streak resetting when the miner’s Elo lead fell below the required 200 points at the validator tempo. Operators also questioned eligibility and strategy classification. Those disputes remain unresolved; this edition does not treat copying or manipulation allegations as established facts.
The reset helps explain reward eligibility, while disputed claims need stronger evidence.
Vidaio calls for miner v0.5.4 during infrastructure migration
The team’s September 5 brief says infrastructure migration is around 70% complete and miners must move to v0.5.4 using the new CPU-based onboarding guide. Validators are directed to a thin-validator setup without a GPU requirement. Temporary disruption is possible during migration; the first compression competition is planned for next week, with details still pending.
Required miner upgrades and an incomplete infrastructure migration directly affect service continuity and participation.
MicroTensor brings the classification arena into its live competition setup
New commits add label-accuracy scoring for the classification arena and configure classification as the live track, alongside coordinator and round-window fixes. The team’s September 5 brief describes Arena 4, intent classification and resource limits, while acknowledging deployment-configuration issues. SN92 was up 24.38% over 24 hours in the 9:32 AM ET α/TAO snapshot.
A substantive new evaluation track and a large market move warrant attention; deployment issues remain part of the story.
Albedo reports a new benchmark winner while reviewing training variance
The latest team discussion reports that King CXVII beat Genesis on both cited benchmarks. It also highlights substantial variation between training runs and duels. The comparison is team-reported and has not been independently reproduced in this edition.
A benchmark result is worth tracking, but its broader significance depends on reproducibility and variance.
Cortex adds its own inference judging backend and narrows live challenges
Public commits add an InferenceOffer backend for RLM judging and remove Prism, Relearn and Design from the active product challenge set, retaining Bounty and Proof. The changes focus the network on code bounties and research proofs. They verify a change in the software and evaluation structure, without establishing new customers or revenue.
A change in the live challenge set and core judging dependency materially reshapes the subnet’s work.
ConnitoAI reports restored validator status and repaired cohort grouping
The team says validator status is back online and the cohort-grouping problem has been resolved. This is a reported operational recovery; the scan did not independently replay resulting scores or reward allocations.
Restored evaluation infrastructure is useful for participants, with a contained operational scope.
Capcomp reports its first cycle of sealed on-chain mining commitments
The team reports 27 recipes sealed before results existed, using timelock encryption and drand randomness, with run bundles on Hugging Face and digests committed on-chain. It plans to expand the adapter pool from 30 to more than 50. The repository documents the commitment workflow, but this edition did not independently replay the 27 recipes or audit their on-chain records.
Commit-before-result evidence could make mining outcomes independently auditable; the completed-cycle counts remain team-reported.
Beam reports petabyte-scale transfers and hardens encrypted room handoffs
Beam’s September 4 official update reports petabyte-scale data flowing among services including Hugging Face, Hippius and major object stores. That volume is a project claim without an independently verified measurement window. A September 5 commit separately requires end-to-end-encrypted room-transfer capability during handoffs.
Reported usage at this scale and concrete transfer-security work are meaningful infrastructure signals, with the metric kept attributed.
Nodexo explains executor-related 503 errors and stale canary flags
The team attributes affected validator API 503 responses to missing executor registry entries. It also says flags left by old IP and port configurations clear after a successful canary check, rather than simply with time. The discussion does not support describing the issue as a system-wide outage.
Correctly identifying registration and endpoint failures helps operators restore working nodes.
Minos reports reproducible scores; an emissions restart is still unconfirmed
The team says shared ground-truth files now make scores reproducible. The same discussion includes a request to re-enable emissions, but no confirmation that the restart occurred. Although the secondary brief’s headline says emissions resumed, its underlying summary supports only a request. Validator ranking and reward mechanics remain under discussion.
Score transparency and paused rewards are central to miner confidence; a restart must be confirmed before being reported as complete.
Finsight reports a working testnet prototype and targets integration within two weeks
Finsight’s developers say a testnet prototype is working and Bittensor integration testing is underway. They target completion within two weeks and plan to share the repository. This remains a team-reported development milestone; production launch and the timetable have not been independently confirmed.
A working prototype is a step beyond planning, while integration and public verification remain unfinished.
Green Compute plans on-chain miner payouts for greater transparency
Subnet lead Josh says miner payouts will move on-chain so the community can inspect them publicly. The team is also working on additional transparency measures. The reviewed discussion contains the commitment, but no completed migration, activation date or reconciled payout history.
Publicly inspectable payouts would improve accountability; completion is the key follow-up.
Ditto ships coding practice tasks and encrypted benchmark delivery
Ditto added ten public coding practice tasks and encrypted private benchmark payload delivery through Hippius. The latest v0.217.0 release also follows a hosted control admission API that is off by default. These are inspectable preparation and security changes for the coding track; they do not establish a generally available hosted coding launch.
Practice material and protected benchmark delivery are meaningful steps toward a usable, harder-to-game coding competition.
Affine makes submissions private; private-duel issues are under investigation
Catch-up from September 4: Affine’s official account says submissions are now private, crowned models remain public, and losing checkpoints will be released later for verdict review. The September 5 team discussion also acknowledges a private-model evaluation problem and reviews overfitting penalties. The privacy change should not be mistaken for confirmation that private dueling is currently working reliably.
Submission privacy changes competitive incentives and audit timing, while evaluation failures can affect rewards.
Sundae Bar’s SBC25 submission window closes Sunday at 2 PM EDT
The team says SBC25 is open from September 4 at 2 PM EDT until September 6 at 2 PM EDT. It closed the previous challenge early after reporting that its production-performance criteria had been met. The new deadline is actionable; the claimed enterprise performance has not been independently benchmarked.
A live submission deadline is directly relevant to miners, with the team’s performance claim clearly attributed.
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.
