September 3, 2026 · Published daily
Subnet News - September 3, 2026
Capcomp opened more of its audit trail, Ditto staged a private coding canary on Hippius, MicroTensor changed its emission ladder, Cascade scheduled new duel-only rules, and several teams shipped scoring and infrastructure fixes.
- Coverage window
- Sep 2, 2026, 10:00 AM – 10:00 AM ET
- Subnets in coverage roster
- 128
- Subnets with updates
- 23
What matters first
Top signals
Capcomp removes privileged early access and publishes replay evidence
Capcomp bound run archives to on-chain commitments, published per-instance traces and removed the validator early-access path. Validators now read the same published run after the fact, giving outside reviewers a clearer audit trail.
Ditto stages a private coding canary on Hippius
Ditto added a default-off, single-validator coding canary that encrypts inputs stored on Hippius and includes custody-recovery controls. The code is staged behind hard gates, so this is not a network-wide coding rollout.
MicroTensor moves to a fixed top-eight emission ladder
MicroTensor committed a fixed reward ladder for the top eight positions and now carries each system source through settlement into the control plane. These are verified code changes; this review did not independently replay an on-chain payout.
Organized by SN#
Subnet desk
Apex
Apex publishes v4.4.1 after the v4.4.0 feature release
Apex tagged v4.4.0 and a follow-up v4.4.1 during the reporting window. The repository changes add richer artifact and evaluation metadata; the commits verify the software release, not adoption by every validator.
DSperse
DSperse begins a validator IP-rotation rollout
A dated ecosystem brief reports that v14.14.0 adds security patches and tolerance for validator IP rotation, with rotation beginning during this window. The linked Discord post could not be independently retrieved, so rollout status remains developing.
Numinous
Numinous adds a baseline hurdle to forecasting rewards
Numinous announced that a baseline forecaster now serves as a difficulty proxy alongside market price, so miners must outperform the baseline to earn rewards. The post describes an active scoring change; independent reward-output verification was not available.
Pareton
Pareton hardens Docker cleanup without discarding useful build cache
Pareton merged storage-pruning changes and a guard that preserves the native build cache. The code targets accumulated container storage while reducing the chance that cleanup forces unnecessary rebuilds.
AdTAO
AdTAO stops charging miners for operator-side execution failures
AdTAO changed scoring so missing images, disk pressure, sandbox problems and exhausted operator budgets no longer count as miner absence. It also added a public admission-verdict feed for image digests and queue decisions.
UR
UR completes its release 1.0 adversarial testnet campaign
UR committed the release record for an adversarial testnet campaign, including pinned artifacts, evidence and supervised logs. This is testnet release evidence and should not be read as a mainnet launch.
SayGM
SayGM announces Gemini 3.8 Flash support
SayGM said Gemini 3.8 Flash is live and added image-generation API support, while advertising pricing below the model vendor list price. Availability and pricing are project claims from the official account.
Nepher Robotics
Nepher reports a completed Leatherback LiDAR maze run
Nepher says its Phase 2 Leatherback robot completed a real maze without wall contacts and reported 97 miners with 337 submissions. Those performance figures are project-reported and were not independently reproduced.
lium.io
Lium tightens GPU capacity and overlay checks
Lium raised matrix-multiplication VRAM headroom for confidential-computing GPUs and corrected its cluster-overlay guard to evaluate UDP rather than treating unrelated TCP use as a conflict. Both changes are present in the public repository.
True Performance Network
TPN opens its third model competition
True Performance Network announced TPN-003, a competition for a Gemma 4 12B uncensored model constrained to run under 8 GB and evaluated across five benchmarks. The official post establishes the competition terms; results are not yet available.
Gittensor
Gittensor reduces long-prompt prefill memory pressure
SparkInfer now processes long prompts in position windows instead of allocating scratch space for the entire prompt. A separate commit adds an OpenRouter service-level profile; it is configuration support, not evidence of a commercial listing.
OpenRoboto
OpenRoboto publishes real-robot track rules and prize-pool verification
OpenRoboto documented its real-robot track, including challenge and claim mechanics, then published the SN80 UID 2 prize-pool hotkey and verification steps. The commits establish the rules and address, not the value of future awards.
InfiniteQuant
InfiniteQuant fixes a mismatch between displayed and paid weights
InfiniteQuant changed its validator and audit query to carry exit timing into the causal scoring rule. The project says the old path could show dashboard share while paying zero; the repository verifies the fix, while the stated affected-miner count remains project-reported.
cascade
Cascade schedules duel-only scoring changes for block 8,992,800
Cascade committed a block-gated switch to duel-only rounds, a scored-horizon ladder and a fresh-king margin reduction from 2% to 1%. The publication snapshot was still below the activation block, so these rules were scheduled rather than active.
MicroTensor
MicroTensor moves to a fixed top-eight emission ladder
MicroTensor committed a fixed reward ladder for the top eight positions and now carries each system source through settlement into the control plane. These are verified code changes; this review did not independently replay an on-chain payout.
Bitcast
Bitcast reports that emissions and payouts have resumed
A dated brief says Bitcast restored emissions after a temporary interruption and that miner payouts continued. The live screener showed positive emission at publication, but the cause and exact recovery path were not independently verified.
Cortex
Cortex makes bounty scoring fail closed on public evidence
Cortex changed bounty scoring to accept only its public backend feed, added checks against mismatched leaderboard and report revisions, and pinned a proven CUDA image. The commits strengthen reproducibility and consistency without establishing downstream validator adoption.
Capcomp
Capcomp removes privileged early access and publishes replay evidence
Capcomp bound run archives to on-chain commitments, published per-instance traces and removed the validator early-access path. Validators now read the same published run after the fact, giving outside reviewers a clearer audit trail.
Beam
Beam completes its first R2-to-Hugging Face transfer
Beam announced a successful transfer from R2 storage to Hugging Face. The official post verifies the project report, but no independent throughput or durability test accompanied it.
Minos
Minos remains at zero emission while auditability is reviewed
Minos showed zero emission in the publication snapshot. A dated brief attributes the pause to a broader review of auditability and validation, but no primary explanation was accessible, so the cause remains developing.
Claims
Claims updates reward buckets and tightens validator requirements
Claims added revised reward-bucket logic and a newcomer guarantee, then enforced Silver-tier mainnet validator requirements. A follow-up registration-burn fix landed in the same window.
Ditto
Ditto stages a private coding canary on Hippius
Ditto added a default-off, single-validator coding canary that encrypts inputs stored on Hippius and includes custody-recovery controls. The code is staged behind hard gates, so this is not a network-wide coding rollout.
CookingTAO
CookingTAO opens notification beta v2
CookingTAO announced a second notification beta with eight event types, mandatory billing and security notices, one active subnet and five more planned. Those deployment counts are the project’s current report.
No published update: SN3, SN4, SN5, SN7, SN8, SN9, SN11, SN12, SN13, SN14, SN15, SN16, SN17, SN18, SN19, SN20, SN22, SN23, SN24, SN26, SN27, SN29, SN30, SN31, SN32, SN33, SN34, SN35, SN36, SN37, SN38, SN39, SN40, SN41, SN42, SN43, SN44, SN45, SN46, SN47, SN48, SN50, SN52, SN53, SN54, SN55, SN56, SN57, SN58, SN59, SN60, SN61, SN62, SN63, SN64, SN66, SN67, SN68, SN69, SN70, SN71, SN72, SN73, SN75, SN76, SN77, SN78, SN79, SN81, SN82, SN83, SN84, SN85, SN86, SN87, SN88, SN90, SN94, SN95, SN96, SN97, SN98, SN99, SN101, SN102, SN104, SN106, SN108, SN109, SN110, SN112, SN113, SN114, SN115, SN116, SN117, SN119, SN120, SN121, SN123, SN124, SN125, SN126, SN127, SN128. See coverage notes for source-access gaps.
Audit trail
Coverage notes
Searched
- All 128 active subnet slots and names from the ShizzyUnchained finalized-chain screener snapshot
- Dated September 3 subnet briefs and their available source links
- Public GitHub repositories for changes committed from September 2 at 10:00 AM ET through September 3 at 10:00 AM ET
- Official public X posts surfaced during the same fixed reporting window
Public-data gaps
- Discord posts that required membership or were unavailable outside the app were not treated as independently verified.
- Repositories that were private, missing, organization-only listings, or had no public repository link could not be fully code-reviewed.
- No independently verified forward alpha-unlock calendar was available, so upcoming unlocks are left empty rather than inferred.
- On-chain values cited from the screener are a publication snapshot and may differ from the 10:00 AM ET coverage cutoff.