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 A past verified announcement can now be outdated.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.
SN66 Conjectures reports verified DEFLATE parsers
The dated digest reports three DEFLATE submissions released as a verified Rust crate, with Lean proofs of parser correctness and claimed speed/compression gains over named comparators. This is attributed release evidence, not a benchmark we reran or proof every compression operation is universally superior. A separate public report describes requiring AI chat logs for new proof bounties; its detailed bonus amounts were not verified against original rules and are not presented as confirmed payouts.
A reported usable verified-parser artifact is distinct from casual benchmark chatter.
SN78 UMI previews its product path; paid weights are not payouts
SN78 UMI’s October 16 trailer and caption describe pre-orders as opening soon. Inspected frames show SignRush, miner/model competition and a reward-curve discussion; the full audio was not transcribed, so further media-only specifications are not inferred. Separately, the dated digest reports certified C5 paid-service allocation weights, while explicitly keeping on-chain reward activation pending. A published allocation table is not an executed alpha payment, and a product trailer is not completed preorder availability.
New product context and a scoring milestone matter while the payment gate remains unresolved.
SN81 Reliquary adds 1,893 supervised-training problems
SN81 Reliquary’s primary environment release appends 1,893 competitive-code problems for supervised fine-tuning. Its release notes keep previous training indices and evaluation/qualification sets unchanged and describe decontamination against those held-out sources. This expands training material without claiming a new benchmark score; data-policy implementation is not an independent contamination audit.
A specific training-data release is substantive and its evaluation boundary is documented.
SN86 was removed and re-registered; sale terms are unverified
Late-found direct chain evidence places NetworkRemoved at block 9,289,316 on October 15 at 14:26:48 UTC, in extrinsic 18, alongside a new registration queue entry. NetworkAdded follows at block 9,289,319 at 14:27:24 UTC; the registration-height storage and new owner key match. The queued lock is 493.992245801 TAO, a protocol registration lock—not a negotiated purchase price. The screener now labels the slot OnSale, but that label does not identify a buyer or prove a sale. These events precede yesterday’s stated cutoff by minutes but occurred after its recorded node head, so they are explicitly backfilled today. The identity of any new operating team remains unconfirmed.
Decoded events and before/after storage establish a real slot replacement rather than inferring a sale from price or a name.
SN90 KubeTEE publishes a new hardware acceptance policy
SN90 KubeTEE’s October 16 documentation change says new GPU nodes must use Blackwell or Vera Rubin and new CPU nodes sixth-generation Xeon or EPYC with the specified confidential-compute support. Existing Hopper capacity is deprecated, with removal targeted as new capacity comes online around November. This is published operator policy, not evidence every older node was removed today or that every advertised hardware configuration has been independently validated.
A published acceptance/deprecation rule materially changes node planning.
SN91 Cascade reports first place in a paper-trading competition
SN91 Cascade’s October 15 post says its Gnomon time-series harness with GLM 5.2 finished first of 26 entries across five weeks: +14.8% on a $10,000 paper account over 23 trades. It contrasts GLM 5.2 alone at tenth and −7.7%. The relevant development is a dated application result for its forecasting tooling. This is a team-reported simulated competition, not live investment performance, a causal controlled experiment or a promise of future returns. Connecting the harness to time-series inference is described as the next step.
A specific dated applied result merits coverage while paper-account and independent-reproduction limits prevent overstatement.
SN92 REAX revises scoring while competition awards remain gated
SN92 REAX’s October 16 digest reports active inference mining and a capacity-scoring revision following a delegation fix. Primary release notes for 1.2.0rc5 retain owner-signed approval before competition awards and hold the engine track in code. This prevents the scheduled October 16 competition time being mistaken for independently verified reward execution. Current identity is REAX; commits in the legacy MicroTensor repository are not automatically attributed to today’s slot. The 12:49 UTC alpha/TAO snapshot is −16.04% over 24 hours and +2,485.58% over seven days, which shows volatility rather than proving one catalyst.
Substantial scoring/award-gate changes deserve coverage without turning a scheduled competition into confirmed payouts.
SN95 Actual publishes DGX Station inference measurements
SN95 Actual’s October 15 case study introduces its e engine on NVIDIA DGX Station with Nemotron 3.5 Super VL. It reports preliminary decode measurements of 455 tokens/second for one stream and over 8,000 across 128 streams, with 0.58-second first-token latency for a fresh 8K-token prompt. The stated setup uses FP8 weights and MTP. This is a concrete effort to serve department-scale agent workloads on owned hardware. These are team measurements, not an independently reproduced comparison; a hosted model and a local open model are not interchangeable quality/cost benchmarks. The confidential relay claim concerns the described relay architecture, not proof of privacy on every endpoint.
A primary product case study supplies a named engine, hardware/model setup and measurable serving results.
SN97 Albedo reportedly halted evaluation
The October 16 digest attributes confirmation of a halt at 16:00 UTC on October 15 to SN97 Albedo. Further changes were teased, with no verified specification or restart time. This is an availability update and open follow-up, not a completed relaunch or evidence the slot has been deregistered.
Reported evaluation availability directly affects participants.
SN102 ConnitoAI’s owner coldkey swap is finalized
At block 9,294,106 on October 16, our node decodes ColdkeySwapped in extrinsic 14 and swap_coldkey_announced execution; SN102 ConnitoAI’s owner storage changes from the prior key to the named successor. This verifies execution, not merely a scheduled announcement. A coldkey rotation alone does not establish an off-chain sale, buyer identity or payment. Separately, the dated channel digest reports v0.7.0’s residual-expert task switch and temporary RPC-related fallback weights; that fleet/recovery description remains attributed rather than an independently reconstructed reward history.
A decoded executed key change resolves ownership-state monitoring without inventing transaction economics.
SN105 Beam publishes audit visibility; the global fix remains open
SN105 Beam’s October 16 UTC post advertises endpoints for transfers, assignments, scores and reward calculations. This adds a path for participants to inspect recorded outcomes. It does not resolve the underlying room/S3 assignment allegations: the dated digest attributes continued investigation and says the global fix is not yet deployed. The endpoint demo was not fully reviewed or reconciled against historical transfers, so neither universal audit completeness nor recovery of disputed rewards is asserted. A secondary Box-integration lead lacked a verified counterpart announcement and is not promoted as a new confirmed partnership.
A transparency feature and a material unresolved incident belong together, avoiding an unsupported all-clear.
SN108 ChipForge closes Quest1 and previews a larger NPU task
The October 16 digest reports Quest 1’s October 15 21:00 UTC deadline, with timely submissions evaluated and a further 24-hour reward period for the winner. Quest 2 is described as a larger edge-AI NPU with a custom compiler; full specification and timing remain upcoming. The 12:49 UTC alpha/TAO snapshot is −21.19% daily, which is a research trigger and not evidence the scheduled transition caused the move.
A specific competition transition affects participation; the next task is accurately labelled planned.
SN117 Everyframe prepares v0.5.7 with clearer provider failures
SN117 Everyframe’s primary miner change preserves provider failure categories and prepares v0.5.7. Distinguishing a provider fault from a generic worker error helps operators diagnose failed service. This follows yesterday’s balance/recovery release; merged preparation is not proof that all workers upgraded or that all external provider errors have disappeared.
A bounded operational repair improves diagnosis without overstating rollout.
SN118 Ditto describes a new development-funding allocation
The October 16 digest attributes a plan to direct 75% of miner emissions to product development to SN118 Ditto, alongside stuck-agent work and a future source reveal. This is different from yesterday’s reported 70% burn/gamma settings: the two descriptions are not silently equated or added together. Finalized weights and funding transfers were not decoded, so exact implementation and reconciliation remain open. The team’s separate roughly-$1 pull-request-review example is a workload claim, not a universal cost guarantee.
A reported major funding policy warrants notice while conflicting mechanisms require verification.
SN120 Affine weight complaints remain unresolved
Yesterday’s primary recovery work does not establish a fleet-wide all-clear. The October 16 SN120 Affine digest continues to report unchanged weights and active miners missing rewards, with a team account asking for action. Its exact duration and individual losses were not independently reconstructed. A separate public proof-of-work registration report is relevant to access, but does not settle the weight-update incident.
A continuing reported reward failure remains material despite published fixes.
SN121 Sundae Bar’s bundle challenge has a reported scoring issue
The October 16 digest says SBC30 opens with folder-based submissions and reports an over-strict cross-file repetition check, with a fix scheduled for SBC31. It also reports removal of exposed construction_notes from the challenge API. These are attributed operational findings, not independently replayed exploit or grading results. Yesterday’s bundle support does not imply the new format is free of scoring faults.
A new challenge and specific reported scoring limitation affect submissions.
SN122 CookingTAO separates total revenue from monthly revenue
The October 16 UTC post from SN122 CookingTAO reports $1,064.84 in total beta-v2 revenue, including $505.44 this month, and a $17.69 alpha buyback test. This clarifies yesterday’s approaching-$1,000 language: the new figure is cumulative, not verified monthly recurring revenue. Automated real buybacks are still planned after beta. Its mining-management product provides the business context; revenue and the test transaction were not independently audited.
A new concrete commercial milestone also corrects an easy-to-confuse revenue period.
Null Consensus offers larger competitions—with a single weight-setting validator
Openτensor’s October 15 release announcement and primary documentation describe an owner-selected alternative to Yuma. Null permits capacity up to floor(2,500/mechanism count), versus floor(256/mechanism count) under Yuma; switching does not automatically expand a subnet. The highest unrounded-stake registered UID becomes the sole validator permitted to submit weights, and consensus/bond calculations are skipped. That creates room for bigger competitions and precise integer reward ratios, while concentrating weight-setting authority. Yuma remains the default and root cannot select Null. Optional proof-of-work registration is independently available for Null or Yuma: a direct zero-tip submission needs no registration burn, initial collateral purchase or transaction fee, but still requires computational work and obeys admission rules. Const’s October 15 quote welcomes the added scale and freedom; that is commentary, not evidence every subnet has switched. Direct finalized storage at block 9,296,035 confirms SN120 Affine already using Null with capacity 1,000 and PoW enabled; SN25 UR and SN128 byteleap.ai remain Yuma with capacity 256. The release index separately labels a 16,000-capacity Null version upcoming, conflicting with the guide/post’s 2,500 description. We do not assume that larger version is active; exact per-subnet configuration takes precedence.
A primary protocol release materially changes competition capacity, admission options and trust assumptions across Bittensor.
eTAO emission-weight code is published; finalized runtime remains 479
The primary repository contains spec-480 eTAO per-subnet emission-weight work. Our finalized archive-node check at block 9,296,035 still reports runtime 479. Published implementation and active-chain deployment are separate facts: this report does not call spec 480 live or assume its new controls are already available. The current immunity and registration-cost decay parameters match yesterday’s checked values. This matters because subnet teams should follow actual activation gates, not treat every merged or branch commit as a mainnet upgrade.
Substantial protocol implementation is relevant, with the observed activation boundary retained.
New public Shizzy episode added to the video library
The October 15 public upload “TAO RISES FROM HELL! 🔥 Bittensor Is Making HUGE Moves!” is now the default video, with 18:44 duration confirmed by player metadata. Both Videos and Live tabs were checked. SN48 Quantum Compute remains a completed replay, the October13 members video stays correctly restricted, and no new upcoming stream was verified. Removed or postponed episodes remain disabled.
Verified channel metadata keeps the public player and library current.
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.
