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 A past verified announcement can now be outdated.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.
SayGM patches a configuration-injection flaw in its miner data path
SayGM’s merged miner changes validate a node secret before inserting it into the request-processing configuration, and render that secret after other substitutions. The documented flaw could allow a specially constructed configuration value to inject code into the buyer-facing data path. The fix is concrete source evidence; no confirmed exploitation or customer-data breach was found. Subsequent development also adds measured cloud-model routing and deployment verification.
A security fix affecting request handling deserves coverage, without inflating a patched configuration weakness into an observed breach.
Ralph lowers the challenger margin from 2% to 1%
Ralph’s merged Round 6 change lowers the margin needed to dethrone the current leader from 0.02 to 0.01. This is a targeted competition-setting change that makes smaller improvements sufficient under the published configuration. Deployment on every validator was not independently checked.
A narrow but useful rule change for participating miners.
Score highlights a large football-analysis cost gap and adds player attribution
Max Sebti reports that full-match processing with the named frontier models would cost roughly 1,000 times more than Score’s private-track models in his latest tests; Score says its models would also be more accurate. Those are attributed test claims, not independently audited results in this scan. A separate Score update says player attribution is being added in post-processing. This is new cost and product context, rather than a fresh announcement of earlier private-track benchmarks.
The cost comparison and player-attribution work are relevant to commercial sports analysis, but the claimed ratio needs broader benchmark evidence.
Lium makes GPU rental and remote Python jobs easier through its SDK
Lium’s SDK updates add automatic selection of the cheapest matching node, a remote Python execution decorator, shared CLI/API-key handling and better machine-readable output. Its v0.0.38 and v0.0.39 releases also improve GPU matching and rental diagnostics. These are practical tooling changes for developers renting compute; they are separate from the throughput benchmark covered in yesterday’s edition.
Lower setup friction and clearer rental selection improve the usability of an existing commercial compute service.
TPN moves toward wider adoption of its benchmark and compressed-model products
True Performance Network says its benchmark product is already being used and stress-tested by miners, and it is starting outreach to grow the user base. Its September 9 update also describes an upcoming consumer product focused on AI sovereignty and further refinement of the subnet’s model-compression service. The consumer product remains in development; the post does not disclose a completed launch or new revenue figure.
A specific go-to-market update connects subnet output to users, while current products and future releases stay distinct.
Conjectures replaces miner-hotkey signing with coldkey-linked submissions
Conjectures’ new schema removes the miner hotkey from submission identity and CLI-session signing. The documented model separates the proved submission coldkey from the payout destination and continues to allow scoped sessions for submissions. This is a source-level account and submission change, not a request to share or upload a private key. Production migration and payout activation were not independently established in this scan.
The identity and signing model materially changes how contributors interact with the proof market.
Harnyx rejects responses with broken citation references
Harnyx says its default research path now validates every citation index and rejects a response containing an invalid reference. A matching September 9 repository update rejects unresolved inline citations. The separate fast mode already existed; today’s change tightens the traceable path rather than introducing fast mode for the first time. Valid citation syntax alone does not prove that a cited source supports every claim.
A checkable reliability improvement matters for research users who need answers they can trace back to sources.
NOVA brief caps Blueprint bounties ahead of the Boltz-2 transition
The September 8 NOVA brief reports a 23,000-alpha bounty cap until the planned Boltz-2 transition, with one initial Blueprint target and server-side evaluation. It says the existing PSICHIC path continues in the meantime and delayed payouts help limit abuse. This adds operating detail to yesterday’s transition plan; a primary activation notice was not located, so the cap and timing remain developing.
Bounty size and evaluation timing affect contributor participation while the model transition is pending.
Gittensor’s SparkInfer improves long-context Muse inference
SparkInfer’s published measurements show a shared key/value decode tile raising Muse throughput from 71.91 to 86.49 tokens per second at a 64k context, about a 20% improvement in that test. A separate fix corrects the cache datatype used during prompt processing, and the newest change reduces repeated storage of a layer operand. These are narrowly specified implementation results, not a general 20% speedup for every model or workload.
Measured open-source inference improvements are relevant to builders, with the exact workload keeping the performance claim meaningful.
Investing88 says its root validator excludes its own subnet
Investing88 confirms that 88 Root does not allocate to SN88 itself, describing the exclusion as a conflict-of-interest precaution. The team says its weight vector is generated from miner outputs. The announcement confirms the stated policy; current full validator allocations were not independently reconstructed here.
A clearly stated allocation policy adds useful transparency, with a narrower impact than a launch.
InfiniteQuant expands custom market sizing while keeping new scoring gates on testnet
InfiniteQuant’s source expands custom sizing to the recorded Hyperliquid market universe, with the commit naming 329 pairs. It also adds a per-wash 24-hour emission cut on the points path, explicitly armed for testnet only. The wider market coverage and experimental scoring rule are development changes, not evidence of a new live mainnet penalty.
A larger supported market universe changes what can be modeled; the testnet restriction prevents misreporting an active emission rule.
Verathos confirms its MicroTensor partnership and testnet model deployment
Following yesterday’s MicroTensor announcement and our model-listing check, Verathos has now published its own confirmation. It describes a partnership to serve certified subnet-native models and identifies mt-code-3g-r1238-5ED3LzEW as the first deployment. The coding model is routed through Verathos’ testnet, with its arena certificate linked to the MicroTensor result. This adds receiving-team confirmation; it does not announce a new mainnet launch.
The other side of a cross-subnet integration has confirmed the deployment. The lower rating reflects a follow-up to yesterday’s lead rather than a second launch.
NeverPlayAlone’s next benchmark adds a wave-based combat mission
The September 9 channel brief describes a new CombatMission round with ten minutes of preparation followed by eight minutes of combat across eight waves. Agents start with an empty inventory and must prepare equipment before fighting; scoring balances three objectives with a death penalty. This is a dated competition lead from the channel digest, with a matching primary release not retrieved during the scan.
A new benchmark tests planning, crafting and combat together, rather than another routine infrastructure change.
Beam publishes transfer-audit access and plans an OpenClaw integration
Beam highlights public routes for inspecting transfer assignments, tasks, attempts, results and overseer interventions. Separately, it says an OpenClaw skill is being built to give agents access to Beam’s data movement and streaming, with deeper Rooms and Tunnels integration planned alongside Beam Studio. The audit documentation is public; the OpenClaw integration is described as upcoming, not a completed partnership or live release.
Public verification tools and a named agent integration are useful product developments, with current access separated from the roadmap.
Claims tightens scientific-claim eligibility and adds adjudication fallbacks
Claims’ September 8–9 changes unify eligibility adjudication, strengthen the checks for its Silver workflow, enable rigor validation in mainnet profiles and add fallback handling when a model-based adjudication fails. These are changes to how submitted research claims are assessed. They are not independent validation of the medical conclusions in the project’s promotional examples.
Better eligibility checks and recovery from failed judging affect the quality and reliability of the scientific-claim pipeline.
TensorUSD closes a route exposing scoring ground truth to the evaluation environment
TensorUSD’s merged change restricts the evaluation environment to the supplied data file and removes access to ground truth. That protects the distinction between the information contestants receive and the answers used to judge them. The project also reiterates that TUSDT is native to Bittensor; its September 9 post does not announce an EVM deployment.
Keeping answers out of the contestant environment is material to the integrity of a forecasting competition.
Ditto confirms V2 is still in preparation, with a redesigned app and onboarding
Ditto’s September 9 product update says the team is preparing V2 with new features, a redesigned application and improved onboarding. This is a roadmap-status update, not confirmation that V2 has launched. Repository work also adds native coding-runtime qualification controls and excludes runs affected by platform inference errors from scoring. Those implementation changes do not establish general availability of every planned feature.
The first-party launch status resolves ambiguity in circulating summaries, while the code changes improve evaluation reliability.
Affine activates its stricter chat-model rules for September 9
Affine’s published fork record moves weight version 12 to 13, requiring a proper closing reasoning tag, enforcing a chat-protocol admission test with at least a 90% pass rate, and scoring the visible final reply. IDE-agent prompts also enter the dataset. The project’s updated fork history explicitly dates the rules as effective September 9, advancing the previously announced change from a future plan to the current rule set.
An effective scoring and admission change can determine whether a miner is accepted or forfeits a turn.
Sundae’s SBC26 challenge window runs September 8–10
The dated Sundae channel brief describes SBC26 as open from September 8 through September 10, with prescreening intended to filter poor-quality submissions before private judging. Separately, Sundae’s public account explains how its agents review completed challenges and open follow-up tickets when gaps remain. The specific competition window is still sourced to the brief.
A short competition deadline is actionable for miners and belongs in the day’s schedule.
Swarm randomizes its simulated lab as it develops a drone interceptor
Swarm says its simulated lab now rearranges the same 41 objects on every run, making a memorized map less useful. The team is building an open-source drone interceptor and testing it in a real lab, but describes the interceptor as still coming. The new environment detail adds substance to the project’s development without implying a finished deployment.
Changing the environment tests whether a model generalizes beyond a fixed room layout.
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.
