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 A past verified announcement can now be outdated.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.
SN56 Gradients plans a check for exposure to Intercode test material
SN56 Gradients’ team-attributed brief says it is addressing training setups that expose Intercode evaluation material and plans an automated check, with manual submission review in the meantime. The team says tournament rules will not change before further discussion with miners. The operational issue is evaluation integrity; no individual miner allegation or completed automated fix is asserted here.
Protecting hidden evaluation data is material to the fairness of model rankings.
SN71 Leadpoet prepares longer Arena runs and previews integrated outreach
SN71 Leadpoet’s repository adds a 60-minute execution profile for new Arena rounds and a guarded migration for the unstarted September 22 round, preserving historical 45-minute results. Separately, founder Gavin Zaentz says LinkedIn and email outreach will soon connect to the same system that finds buying signals and contacts. The longer profile is implemented in code; the outreach workflow is an upcoming product feature, not a confirmed launch.
More execution time changes miner preparation, while the announced outreach direction explains the intended customer workflow.
SN80 OpenRoboto prioritizes new benchmark and data tracks
SN80 OpenRoboto’s latest channel brief says new benchmark and data tracks are expected next week, with emission shares to be adjusted. It also says a physical-robot track is not the immediate priority and no common robot standard has been finalized. This is a roadmap update following yesterday’s simulation comparison; it does not announce live new tracks, a confirmed allocation split or a physical deployment.
The next evaluation priorities matter to miners choosing where to build, with timing and allocation still pending.
SN104 TaoStatus moves reward calculation to the previous completed epoch
SN104 TaoStatus’s merged reward code pays each epoch using verified usage reports from the preceding completed epoch, replacing the carried-forward moving average. A miner with no qualifying report in that window gets no weight from this calculation, and an accumulator that spans a multi-epoch validator outage is discarded. The documented burn fallback still applies when the usage pipeline is unconfigured. This confirms new reward logic in the repository, not universal live adoption or a new network emission allocation.
Reward freshness and removal of stale carried-forward scores materially change miner incentives.
SN106 Nodexo reports all 94 executors scoring normally
SN106 Nodexo’s team-attributed channel brief says all 94 executors were scoring and rejects an externally observed “seasonal reset” message as coming from its validator or documentation. The count is an operator report at that time, not an independently measured uptime guarantee.
A direct operational clarification helps miners distinguish an actual reset from an unrelated message.
SN120 Affine’s team reports Arbos leading with an automated training script
SN120 Affine’s channel brief attributes to Const an account of team-controlled Arbos leading the competition using an automated GRPO training script costing about $500 per day in compute. It says logs and scripts were shared and calls for more competing miner teams. The figure and ranking are team-reported, not independently reproduced here. This describes competition participation; it is not evidence that Const or Arbos manually switched network emissions on or off.
Team participation and the claimed training cost are relevant competitive context, while the emissions interpretation must remain evidence-based.
Emission watch: SN10 Pareton and SN71 Leadpoet return to zero allocation
At 10:28 AM ET, SN10 Pareton and SN71 Leadpoet show zero network allocation, versus approximately 0.000812% and 0.000239% respectively in yesterday’s saved snapshot. SN70 moves from zero to 0.008036%; its current metadata has no project name. These compare two observed snapshots, not an uninterrupted history or confirmed miner payouts. Network allocation and a subnet’s voluntary burn policy are separate. The data does not establish manual intervention by Const, Arbos or any other named person.
The zero/nonzero changes continue the requested emissions watch without overstating tiny allocations or inventing causes.
StonkFun adds AVAX pairs and reports $1.10M daily revenue
StonkFun says AVAX-paired launches are now available through Sunrise. Its September 20 report lists $1,100,719 revenue, $660,554 buybacks and 2.35 million STONK burned, while another announcement puts cumulative rewards to holders of reward coins at $65 million. These are platform-reported figures, not audited accounts or TAO-only revenue and burns. The new pairing is AVAX; previously covered TAO launch pairs are not being announced again.
A new supported pairing and fresh operating totals update the broader StonkFun ecosystem watch.
Exploit Summit confirms livestream coverage and a Bittensor 101 opener
Subnet Summer announces livestream coverage from Exploit Summit at New City Gas in Montréal. The summit’s official account also confirms a September 28 Bittensor 101 opening session with TAO Templar and MCJ of Learn Bittensor, ahead of two days of talks and demonstrations. This adds a way to follow the event remotely. Future product reveals remain future announcements; this is Subnet Summer coverage, not a newly scheduled Shizzy livestream.
A confirmed public viewing route and introductory session are useful upcoming ecosystem information.
SN80 OpenRoboto publishes a robot-control demo with sharply lower Jev API costs
SN80 OpenRoboto has released code, recorded trajectories and an offline verifier for a simulated xArm7 moving an apple onto a plate. In its published trial, Typesafe’s Jev completed the task for $0.018825 in about 182 seconds, versus $5.933624 and 707 seconds for GPT-6 Astra with low reasoning. That is roughly 1/315 of the API cost and 26% of the wall time. GPT-6 used seven fewer control cycles; GPT-4.1 mini reached the 160-cycle limit. The useful advance is an inspectable comparison with a shared controller and physics environment. It is one trial per model using structured simulator feedback, not camera perception, a success-rate study or a physical robot deployment. The recordings were not independently rerun.
A public robotics comparison with code and replayable evidence is material; a single simulated trial does not establish general model superiority.
SN92 is renamed Available as MicroTensor seeks a new subnet slot
SN92’s current metadata now reads “Available,” matching SubnetRadar’s new rename report. Today’s channel brief says MicroTensor validators abstained on round 1242 after the coordinator’s signing hotkey was pruned, removing the anchor used to verify rules. It attributes a new burn hotkey to the slot owner and says the MicroTensor team is seeking another slot while continuing its platform and revenue work. At 11:26 AM ET, SN92 alpha is down 16.4% against TAO over 24 hours and its observed network allocation is zero. The metadata and market snapshot are checked; the operational account is digest-attributed. This is a slot transition, not evidence that the MicroTensor company or product has shut down.
A confirmed slot-name change, reported loss of the validation anchor and a substantial price decline make this a priority participant-risk story.
SN118 Ditto announces live Telegram group memory
SN118 Ditto says its Telegram integration is live, allowing a connected group’s conversation history to be read and summarized. Its demonstration reports reading 322,400 tokens for $0.01 and turning three days of chat into a six-paragraph summary. That gives the subnet’s memory product a concrete everyday use: catching up on a busy community without scrolling through every message. The figures describe the team’s example, not a guaranteed price or accuracy level for every group; an authenticated integration was not independently exercised.
A live product integration brings subnet-backed memory into a widely used community workflow.
SN118 Ditto adds a reproducible V13 enterprise benchmark candidate
SN118 Ditto’s repository adds deterministic V13 enterprise datasets and checks that scorers support the intended benchmark version. Seeded generation makes evaluation inputs reproducible, and the release separates that path from deferred private dataset production. The code comments explicitly say V13 has not activated publicly. This is a benchmark implementation and readiness update, not proof of a completed public scoring migration or a new independently confirmed leading agent.
Reproducible evaluation and verified scorer compatibility improve the ability to inspect competition results.
SN124 Swarm targets April 2027 field trials with a European security partner
SN124 Swarm announces a planned field evaluation of Interceptor and SwarmSAR with an unnamed European special-operations security partner in April 2027. The team says preparatory discussions have been underway for months and describes the evaluation as a step from simulation and lab work into field conditions. This is a significant prospective test of the subnet’s output, but it is a future evaluation reported by Swarm. The partner has not been publicly identified or independently confirmed, and the announcement does not establish a third signed commercial contract, new revenue or an already completed deployment.
A specific field-trial plan is commercially relevant, while future timing and an unnamed partner keep the story developing.
SN11 TrajectoryRL sets Season 2’s opening; Engy auto mode remains a commitment
SN11 TrajectoryRL announced a September 20, 00:00 UTC opening for Season 2, in which miners build fusion policies across eight open models. Its examples report that seven of eleven tested combinations beat the best individual model and four also cost less; these are project-run benchmarks. A material clarification to yesterday’s coverage: updated project documentation now explicitly describes serving the winning policy as SN53 Engy auto mode as a commitment, not something confirmed live. New code also restores the alpha submission-fee burn, while the documented copycat audit currently covers SKILL.md and leaves policy-file checking as future work. The opening time has passed, but a live submission was not independently tested.
The competition changes model-routing incentives, and the corrected distinction between a promised customer route and an operating one is important.
SN88 Investing targets next week for its locked-alpha dividend experiment
SN88 Investing says its incentivized-conviction experiment will go live next week, redirecting a portion of emissions to increase dividends for locked alpha. The announcement advances a previously discussed SN88 experiment with a new launch target. It does not establish a network-wide change to Bittensor rewards, an activated protocol-native feature or a guaranteed yield. The final allocation and live implementation should be checked when the rollout occurs.
A planned change to rewards for locked alpha can materially affect participant incentives, but remains an upcoming subnet experiment.
SN3 Teutonic lowers its configured evaluation delta threshold
SN3 Teutonic’s public chain configuration changes delta_threshold from 0.0125 to 0.0095 while retaining the listed evaluation dataset and shard count. This is a concrete change to the configured improvement threshold used in evaluation. The repository change does not by itself confirm adoption by every live operator or a new winning model.
A small but specific competition-parameter update is useful to active miners.
SN15 ORO prepares new qualifying and race suites
SN15 ORO’s team is preparing new qualifying and race problems to reduce overlap between suites and the need for subjective review. The channel brief also describes consideration of a single task family, with current races continuing under sanity checks. The new suite and single-family design should not both be treated as already active.
Evaluation design changes can alter what miners build and how fairly submissions are compared.
SN21 AdTAO makes gate verdicts and model transitions easier to inspect
SN21 AdTAO adds public gate verdicts with per-horizon measurements and exposes the transition from a hotkey’s previous model to its new model, including remaining allocation and ramp factor. These changes help participants see why a model passes or fails and how its rewards transition. They are verified transparency changes, not evidence of a new advertising-performance record.
Visible eligibility and allocation transitions make reward decisions more understandable to miners.
SN23 Trishool acknowledges a mismatch between jailbreak detection and scoring
The SN23 Trishool channel brief records a case in which content was flagged as jailbroken but received a zero score. A team member acknowledges the mismatch and says the judge model will be tightened. This is an acknowledged scoring issue and a planned improvement, not a verified completed repair.
A mismatch between detection and scoring affects trust in the security evaluation.
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.
