August 20, 2026 · Published daily
Subnet News — August 20, 2026
A subnet slot changed hands for 501.15 TAO while only 31% of alpha markets rose. Perturb activated stake-weighted consensus, BASE promoted PRISM 2.1, and security or integrity changes landed across NeverPlayAlone, Harnyx, Lium, Prometheon, Minotaur, and other active networks.
- Coverage window
- Aug 19, 2026, 10:00 AM – 10:00 AM ET
- Subnets checked
- 128
- Subnets with updates
- 22
What matters first
Top signals
SN36 re-registers for 501.15 TAO as market breadth falls to 31%
At block 8,884,359, SN36 changed from owner 5GLGT…LUxc to 5DCjv…JxpZ in a 501.15 TAO registration. At the cutoff scan, only 31% of subnet markets were higher; the newly reset SN36 market showed the largest 24-hour move at about +279.7%, while Instant was down about 24.5%. The SN36 move follows a slot reset and should not be read like an uninterrupted incumbent-market return.
TAO rebounds while 42 subnet markets remain at near-full miner burn
TaoStats showed TAO near $207.58 at the morning scan, up about 8.3% over 24 hours with market capitalization near $2.35 billion. SubnetRadar counted 42 of 128 reporting subnets at or above 90% miner burn, so the root-token rebound did not translate into broad alpha-market participation.
Organized by SN#
Subnet desk
Apex
Apex closes scoring-path manipulation routes in Humanoid Scramble
Apex released versions 4.3.4 and 4.3.5 during the window. In the associated Humanoid Scramble competition repository, the referee and history-record boundaries were hardened so competitor-controlled records cannot influence scoring through untrusted summary fields or forged history state. The code and release commits are verified; production rollout beyond the published releases was not separately observed.
Pareton
Pareton replaces calibration with round-based benchmark ranking
Pareton removed its calibration and promotion model in favor of explicit benchmark rounds, added a pure ranking seam, and changed the harness to score one batched round at a time. A follow-up prevents one bad candidate from voiding the whole round or escaping disqualification. The round watcher and public read API were also merged before cutoff; the team still had not announced a firm public launch date.
Cacheon
Cacheon launches a submission dashboard and warns of support impersonators
Cacheon says its new dashboard is live with queue state and evaluation output visibility. The team also warned users that scammers are impersonating support in private messages, and said submissions made before the 1 TAO fee change must be resubmitted under the current validator rules.
Perturb
Perturb activates stake-weighted validator consensus and a top-heavy payout curve
Perturb now weights each validator’s reported miner scores by validator stake instead of averaging reports equally. Its miner pool allocates 70% to rank 1, 15% to rank 2, 10% to rank 3, and the final 5% across ranks 4–10; miners below rank 10 receive none. The repository change and the team’s activation notice both landed inside the window.
ReadyAI
ReadyAI investigates anomalous validator weight and task traffic
ReadyAI is investigating a validator reported to be updating weights multiple times per second and receiving an unusually large share of task requests. The team has not established database access, score manipulation, or another root cause. Separately, miners using the default stack reported OpenAI rate limits on the new coverage-evaluation task, which remains capped at 5% of traffic.
Subnet 36
SN36 changes hands in a 501.15 TAO re-registration
The former Leoma slot re-registered at block 8,884,359. Ownership moved from 5GLGT…LUxc to 5DCjv…JxpZ, and the chain event reset the market identity to the generic Subnet 36 label. The registration is finalized on-chain; no project identity or operating plan from the new owner was verified by cutoff.
ChronoLLM
ChronoLLM adds layered duplicate and benchmark-leak detection
ChronoLLM merged a multi-layer deduplication path and weighted leak evaluation. The new checks combine exact and normalized matches with similarity evidence, then weight suspected leakage instead of treating every overlap as identical. The change is verified in the public subnet repository; production activation was not independently observed.
Ralph
Ralph removes an uncollectible bond and caps submissions per tier
Ralph’s v2 incentive work disabled a bond the code could not actually collect, changed submission caps from coldkey-wide to per-tier, and began preserving bait-and-switch evidence instead of silently dropping it. The repository changes are verified; the public branch does not independently establish when validators adopted them.
Instant
Instant opens production miner and validator onboarding
Instant launched a one-line installer and a deterministic scoring contract weighted 40% success, 40% TOPLOC verification, 10% latency, and 10% throughput. Proof generation is bound to the transcript, model, and nonce, and verification runs asynchronously so customer responses are not blocked. The team says emissions remain fully burned until the complete cycle is verified.
lium.io
Lium withholds idle incentives when executors cannot enforce GPU power limits
Lium now denies the unrented-node incentive to an executor that provably cannot apply a GPU power cap, while leaving rented work eligible and avoiding penalties when capability cannot be measured. The gate ships enabled by default, and miner logs identify the required executor-stack or container-capability fix.
engy
Engy onboarding failures continue beyond the earlier attestation blocker
Multiple miners reported worker onboarding failures, including missing capacity-probe requests needed to establish concurrent gateway capacity. Standard Qwen3.6 onboarding was reported working, while GLM5.2 and K3 still require a TEE path or whitelist access. The team was troubleshooting, but no single root cause or completed fix was confirmed by cutoff.
NIOME
NIOME reports Scottish Enterprise backing
NIOME and ecosystem reporting say the subnet has secured support from Scottish Enterprise, Scotland’s national economic development agency. The announcement is material, but no matching Scottish Enterprise release or agreement details were found in the public index by cutoff, so the scope, funding, and claim that this is Bittensor’s first government partnership remain developing.
conjectures
Conjectures repairs live TMC Pay webhook verification
Conjectures found that TMC Pay signs the timestamp plus raw request body, not the body-only form described in the processor documentation. The validator now verifies the live signature shape, preserves omitted invoice fields instead of treating them as retractions, and adds opt-in redacted diagnostics for rejected deliveries. Multi-currency redirect handling also merged in the same window.
Harnyx
Harnyx v9 rewrites champion retention and novelty rewards
Harnyx’s version-9 emission policy raises a newly selected champion’s maximum share to 50%, then steps successful retention to 25% and later to a 12.5% floor, scaled by score. Version 9 also doubles the novel-artifact multiplier from 5 to 10. Failed batches neither advance nor break the retention sequence; historical versions keep their prior rules.
OpenRoboto
OpenRoboto adds 393.6K alpha of new conviction
One OpenRoboto hotkey added two conviction locks of approximately 298.2K and 95.4K alpha during the window, for about 393.6K alpha combined. These are finalized lock events; they do not establish the locker’s motivation or guarantee that the position will remain perpetual.
MicroTensor
MicroTensor confirms launch and burns emissions before public mining
MicroTensor now identifies as SN92 and says the subnet is live. The team is directing emissions to a burn hotkey until the planned Friday public-mining activation. Its repository also added release manifests, round-cutoff accounting, miner telemetry, and unattended training/submission paths during the window. The launch and current burn state are verified team statements; Friday activation remains scheduled.
NeverPlayAlone
NeverPlayAlone bans command-injection gameplay exploits
NeverPlayAlone now automatically rejects agents that abuse server commands such as /give, spoof positions, or self-grant items. The team said three hotkeys were banned retroactively under the new enforcement. The exploit affected competition integrity rather than the Bittensor chain itself.
BASE
BASE promotes PRISM 2.1 and fixes its winner-take-all sealing path
BASE split PRISM 2.1 into a new scoring generation, ignored old 2.0 harvests, and promoted that release to production before cutoff. The new competition burns until its first eligible 2.1 submission. A later fix strips zero-share Design leaves so PRISM’s live champion can be sealed instead of hidden behind burn; that repair was merged and staged, but production promotion of the fix was not independently confirmed by cutoff.
Capcomp
Capcomp cuts the champion share to 90% and opens one resubmission per run
Capcomp reduced the leader’s share of the payable miner pool from 95% to 90%; ranks 2–10 now divide the other 10% by rank instead of grade. Because 80% of each run still burns, the winner receives 18% of total run emission and runners-up share 2%. The code also permits one resubmission per hotkey per run. Separately, miners reported dashboard and signed-report discrepancies that remained unresolved at cutoff.
Prometheon
Prometheon removes Chutes and moves model execution to validators
Prometheon changed commitments to a repository-plus-revision format. Validators now resolve the submitted SHA, download the checkpoint, and run the miner’s model themselves, removing the Chutes deployment and the checks that previously verified it. This is a breaking format-2 migration and requires the database layer to drop the old chute identifier.
Claims
Claims bounds work per run and hardens missing-assessment handling
Claims changed validation to assess paper claims in one diagnostic pass, bounded the number of eligible claims and adjudication cases, and added safe handling for missing claim assessments. The team separately reported output-schema inconsistencies in recent test runs and said new tasks will wait for contract adjustments, so production stability remains under observation.
minotaur
Minotaur fixes cross-chain benchmarks that scored an empty plan
Minotaur found that cross-chain benchmark scoring could inspect the destination placeholder instead of the source-chain execution leg, leaving the scored plan empty. The validator now selects the source leg for these tasks. A second patch stops one failed Chopsticks health probe from declaring the isolated benchmark sidecar dead.
No material update found: SN2, SN3, SN4, SN5, SN6, SN7, SN8, SN9, SN11, SN12, SN13, SN15, SN16, SN17, SN18, SN19, SN20, SN21, SN22, SN23, SN24, SN25, SN27, SN28, SN29, SN30, SN31, SN32, SN34, SN35, SN37, SN39, SN41, SN42, SN43, SN44, SN45, SN47, SN48, SN49, SN50, SN52, SN54, SN56, SN57, SN58, SN59, SN60, SN61, SN62, SN63, SN64, SN65, SN68, SN69, SN70, SN71, SN72, SN73, SN74, SN75, SN76, SN77, SN78, SN79, SN81, SN82, SN83, SN84, SN85, SN86, SN87, SN88, SN89, SN90, SN91, SN93, SN94, SN95, SN96, SN97, SN99, SN101, SN102, SN104, SN105, SN106, SN107, SN109, SN110, SN113, SN114, SN115, SN116, SN117, SN118, SN119, SN120, SN121, SN122, SN123, SN124, SN125, SN126, SN127, SN128.
Beyond one subnet
Ecosystem desk
Lium draws a 631 TAO whale stake while Chutes loses 615 TAO in two trades
The finalized event feed recorded a 631 TAO stake into SN51 Lium, a 315 TAO stake into ORO, and two separate Chutes exits of 315 and 300 TAO. These are distinct single-wallet transactions. They are verified flows, not proof of coordinated positioning or a project catalyst.
Spec 447 conviction rules are live with no challenger over the takeover gate
The live conviction dashboard now identifies spec 447 and its single-hotkey takeover threshold of more than 18% of eligible alpha. It showed 29.79 million alpha locked across 122 subnets and zero challengers above the gate. The same window added roughly 393.6K alpha of OpenRoboto conviction; no unlock of at least 50,000 alpha was found.
One coldkey swap is scheduled; no additional completed owner change is verified
A wallet scheduled a coldkey swap from a BIP44-style address to an Ethereum-style address at block 8,882,744. The event is a pending key operation, not evidence of an ownership sale. Apart from the finalized SN36 re-registration, no other completed subnet-owner change or new registration was verified in the reporting window.
Audit trail
Coverage notes
Searched
- All 128 active subnet slots and the August 20 SubnetRadar daily brief index, including every subnet with a same-day team brief
- Exact commit activity from 71 Bittensor-linked public GitHub repositories between August 19 14:00 UTC and August 20 14:00 UTC, plus known core repositories
- SubnetRadar live tape, moments archive, conviction ledger, burn pulse, registration events and pruning dashboard
- TaoStats TAO and deregistration snapshots, the Shizzy finalized activity page, official project repositories and linked team announcements
- Public web and X indexing for partnerships, releases, security incidents, launches, registrations, deregistrations, halvings and scheduled alpha unlocks
- Every Subnet News report from August 9 through August 19 for exact and semantic duplicate exclusion
Public-data gaps
- Public X indexing is incomplete; private, deleted, login-gated and rate-limited posts cannot be exhaustively reviewed.
- Many source Discord messages require Bittensor server membership. Discord-derived cards link the public daily brief and are classified according to the evidence available.
- The NIOME–Scottish Enterprise announcement had no matching Scottish Enterprise release or public agreement details in the indexed official source, so its scope remains Developing.
- No verified scheduled alpha unlock of at least 50,000 alpha was found for August 20 or a future dated event. Smaller conviction changes are not listed as Unlock Watch items.
- No additional material subnet registration, completed deregistration, exchange listing, alpha halving, or chain-level security exploit was verified inside the fixed window.
- Empty update arrays mean no new material public signal survived the fixed-window and duplicate checks; they do not imply that the subnet had no private or routine activity. This report is informational and not financial advice.