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 A past verified announcement can now be outdated.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.
SN15 ORO moves Race 11 to today and releases local shopper testing
SN15 ORO’s team-attributed brief reschedules Race 11 to October 2 at noon PT / 3 PM ET, giving miners more preparation time. Public code now runs composed shopper tasks through the new environment runtime and includes a bundled practice pack with failure-category feedback. Local testing refuses race configurations: passing the practice pack is not the same as qualifying in the network race. The scheduled race had not started at this research cutoff, and the wider model allowlist remains a team-reported expansion.
A specific rescheduled time and available practice tooling are actionable operator changes, without claiming a race result or automatic qualification.
SN20 Witness versions its judging rules and resets admissions once
SN20 Witness publishes window-specific evaluation rules, including a one-time admission reset at window 32 that preserves consensus history. Another change selects the new text judge and visual review from window 36, with explicit version boundaries rather than silently rewriting previous hashes or results. Qualification remains burn-only in the documented path. These are verified implementation changes; the current active window and resulting finalized miner payouts were not independently reconciled.
Version boundaries and preserved history address a meaningful fairness risk when judging policies change.
SN22 Desearch reports validator slowdown limiting available tasks
After the web-index launch, today’s team-attributed brief says validator performance is slowing scoring and tasks will remain limited until a fix ships. The team says weights are set and burn is not 100%, contradicting broader outage claims in the channel. The public crawl-task endpoint is reachable and exposes verdicts, but that alone does not establish normal throughput or deployment of every new track. No completion timeline or independently measured full recovery was located.
A specific acknowledged scoring bottleneck is a material operational follow-up to yesterday’s launch.
SN28 SayGM offers invoice-matched inference credit up to $5,000
SN28 SayGM is offering to match a customer’s last OpenRouter invoice with the same amount of SayGM account credit, capped at $5,000. Its October 2 announcement says customers can keep the same SDK and switch with a one-line change, with prompts routed through a verifiable trusted execution environment. This is service credit for an onboarding offer, not a cash payout or verified new revenue. No invoice was submitted or credit awarded in this scan; a user reply reports a problem with the claim link, so successful redemption remains untested.
A concrete customer-acquisition offer has clear value and terms, without overstating redemption, revenue or security guarantees.
SN38 ChronoLLM targets its 16B model for later in October
SN38 ChronoLLM says training has been running for roughly 14 weeks and targets the 16B model for later this month, with 72B and larger models as longer-term goals. The October 1 reply clarifies that it remains early in the build; it does not establish a new institutional deployment, customer contract or buyback program. Keep the live time-aware model work already covered separate from the still-upcoming 16B milestone. The previously reported miner-attestation repair has no new independently verified admission or payout outcome in this scan.
A dated model-size target clarifies the next material milestone and keeps SN38 visible without turning a roadmap into a completed commercial launch.
SN44 Score reports break-even, paying clients and executed buybacks and burns
This is the business milestone to watch. SN44 Score says September was its first break-even month and two private-track partners have become long-term paying clients, with work expanding from sports into fuel retail and security through Manako. It has also started buying and burning alpha from the subnet owner address. The first linked transaction, Finney block 9190693, extrinsic 8, succeeded and records a 106.895289378-alpha burn. The customer and break-even figures remain team-reported, with no published financial audit. Break-even is not profit, and SN44 Score promises no fixed buyback amount or schedule.
Paying-client conversion and reported break-even are substantial business progress, backed by a directly checked successful alpha-burn transaction. The rating rests on adoption and execution, not price alone.
SN44 Score reports new PPE and fallen-person tasks, with a scoring issue open
Today’s team-attributed brief says SN44 Score has added public tasks for worksite protective-equipment compliance and fallen-person detection, and rebalanced 12 scoring fields in its private Cricket track. The brief also says the team acknowledged emission assigned to a zero-score model and plans a fix. These additions do not prove yesterday’s separate Person-element outage is resolved. The task rollout and open scoring issue remain attributed to the channel summary.
New customer-relevant vision tasks broaden the competition, while the acknowledged scoring issue stays visible.
SN46 Instant releases a validator fix for single bad proofs
SN46 Instant publishes validator v0.1.3 with a signed release manifest and downloadable Linux binary. The new rule disqualifies a miner only when at least two proofs fail or time out and those failures exceed 0.5% of the proofs judged; previously a single bad proof could zero its epoch score. A companion updater verifies signatures and hashes and provides rollback if the replacement fails. Release publication and code are verified; adoption across all validators and changed live rewards were not independently checked.
A released fix addresses a specific unfair disqualification case and adds safeguards for operator updates.
SN49 Nepher interview is now available as a 52:47 replay
Shizzy’s October 1 conversation, “Can Bittensor Train Real Robots? Inside Nepher Robotics SN49,” is now a completed public livestream replay. The YouTube Live tab and watch-page metadata confirm the video ID and 52:47 duration. The site’s video library has been refreshed to make this the latest public episode and remove its upcoming-stream presentation. No new model performance or tournament result is inferred merely from the interview being available.
A completed, verified replay gives readers a direct way to catch up with the team.
SN51 lium.io changes handling of rented GPUs that fail
SN51 lium.io merges code to mark a rented node inactive when a host-side check confirms a dead GPU runtime, so the backend can stop billing rather than leaving an unusable node active after a failed scrape. The check distinguishes genuine GPU-loss codes from unrelated disk or temporary errors and does not rely on claims made from inside the renter’s container. A separate GPU-drop reporting path is default-off and observational. Code is verified; fleet-wide rollout and actual billing outcomes were not tested here.
A concrete customer-facing billing safeguard addresses unusable rented hardware while avoiding false fault reports.
SN60 Bitsec adds a Sentios continuous-audit badge
SN60 Bitsec introduces a Sentios badge intended to show that a codebase receives ongoing vulnerability searches as it changes. The official post describes the badge as proof of auditing effort and token spend, and its own public repository displays an active Sentios badge. This is a visible assurance signal for its already-covered continuous-audit service. It is not an independent security certification, proof that every vulnerability has been found or a guarantee that the code cannot be exploited.
An inspectable badge makes the ongoing audit service visible to users, while its limits are kept distinct from a security guarantee.
SN62 Ridges launches its test-generation competition
SN62 Ridges announces competition 29 for generating tests that remain useful through bugs, refactors and rewrites. The team calls it live, resolving the planned-competition watch from September 25. In a separate update, it reports its database competition rose from 30% to 36% on its internal measure, with 3,576 agents from 738 miners. The competition launch is a primary announcement; these database figures remain team measurements rather than independently reproduced benchmark results. New allocation targets should not be read as a reconciled payout to every miner.
An announced live competition opens a distinct software-engineering task, alongside specific participation and progress figures from the existing database track.
SN64 Chutes publishes a downloadable Parallax research checkpoint
SN64 Chutes has published parallax-8b-kappa on Hugging Face with public model files and an MIT license. The model card describes a roughly 7.8-billion-parameter mixture-of-experts model with about 1.25 billion active parameters, trained across distributed consumer GPUs over the public internet. The October 2 export records 738.3 billion training tokens. This is a mid-training base checkpoint: the run is still ongoing, and the release is not instruction-tuned, chat-ready or established as a safety-tested production model. The new inspectable checkpoint goes beyond the previously covered phone demo.
Public weights and a detailed model card turn a distributed-training demonstration into an inspectable research artifact, with its unfinished base-model limits explicit.
SN71 Leadpoet publishes recovery changes; upstream scoring remains unresolved
SN71 Leadpoet adds replay checks for the extended recovery schedule and rejudges October 1 work with its original event-state scope preserved. Today’s team-attributed brief reports A7 v1 and says a remaining scoring issue comes from an upstream API bug that the team is fixing. Public recovery code is progress, but it does not prove the provider issue or every held evaluation has cleared. Yesterday’s scoring-hold watch remains open.
New recovery implementation and an attributed explanation clarify the incident without declaring unverified resolution.
SN74 Gittensor releases SparkInfer v0.6.4 with faster first-token handling
SN74 Gittensor releases SparkInfer v0.6.4 with downloadable Linux and Windows packages. Prefill and decode now share a pass by default, reducing queueing for concurrent chats. In the team’s RTX 5090 / Qwen3.8-27B NVFP4 test with 32 concurrent chats, median time to first token fell from 1,183 to 372 milliseconds, compared with 356 for its vLLM baseline. Per-token latency increased in those concurrency tests and most other benchmark cells were unchanged. This is a scoped team measurement, not independently reproduced proof of universal speed gains.
An actual release addresses a measured latency bottleneck, with hardware, workload and tradeoffs specified.
SN75 Hippius releases desktop v0.6.6 with clearer sharing controls
SN75 Hippius releases desktop v0.6.6 for macOS, Windows and Linux. Shared folders now look distinct from whole drives, obsolete sharing indicators clear when access is removed, invitation-limit messages show the wait time, and the menu bar displays the account balance in dollars. This improves the already-launched Shared Drives experience; it is not another launch of that feature.
A shipped usability release improves file-sharing clarity and account information.
SN78 UMI introduces SignVision smart glasses; preorders are still ahead
SN78 UMI introduces SignVision, a planned wearable that translates American Sign Language into speech. The October 1 announcement says an existing hardware-infrastructure partner is handling the hardware side, with API and SDK integration connecting the glasses to the translation service. The partner is not named, and preorders are described as opening soon. This creates a concrete prospective customer product and recurring-service path, but does not establish shipped devices, paid orders or booked revenue. Market attention is real: the site’s alpha/TAO snapshot at 15:15 UTC on October 2 showed SN78 UMI up 22.55% over 24 hours. That move does not prove its cause.
A specific wearable product gives the commercial roadmap substance, while unnamed hardware arrangements and pending preorders keep it below a completed five-star launch.
SN78 UMI hardens C5 enrollment, but its API still says rewards are inactive
After yesterday’s intake reopening, SN78 UMI’s official C5 launch post is now available and new code makes endpoint enrollment durable and points miner upgrades to the enrollment runtime. The public competition API returned 176 accepted submissions during this scan, up from yesterday’s 165, with intake open and evaluation ready. It still explicitly reports intake_no_weight mode, rewards_active: false and chain_submission_authorized: false. C5 participation is progressing; resumed reward payments are not confirmed.
New enrollment fixes and observable submissions are useful restart progress, while the public reward gate prevents a false all-clear.
SN80 OpenRoboto reports nearly 90% success on simulated robot tasks
OpenTensor’s October 2 recap of Novelty Search says SN80 OpenRoboto miners improved Physical Intelligence’s π0.5 model from about 55% to nearly 90% success on simulated robot tasks in under a month. The episode also discusses testing on real hardware and the already-announced Shift training-data network with OR-S1 capture devices. The nearly 90% figure applies to the simulated tasks; it is not a claimed success rate for real-world deployment. This is a completed discussion and reported result, replacing yesterday’s upcoming-event reminder.
A substantial, specifically scoped simulation result gives the robotics competition measurable progress, without conflating it with demonstrated real-world reliability.
SN81 Reliquary adds public task shares and per-miner audit status
SN81 Reliquary’s public code now exposes each declared training task’s mechanism, cap, status and job through /corpus/tasks, plus per-job miner routes for audit deadlines, verdicts, verified tokens and recent task-share history. A subsequent change serves cached task information immediately instead of making requests wait for a slow registry read, and isolates judging work from request-handling threads. This adds inspectable accounting tools after yesterday’s four-job announcement. Current endpoint completeness and payout reconciliation were not independently verified; community allegations are not treated as proven misconduct.
Task-level and miner-level evidence makes the training competition easier to inspect and addresses a documented status-request delay.
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.
