King-of-the-hill trajectory-distillation subnet on Bittensor (netuid 97, finney).
SN 97. Rank 6 of 32 by emission
Albedo
Emission5.2%of the block reward · rank 6 of 32
Participants5 miners · 3 validators
Registration0.2096 TAO
Verified2026-09-09links checked and scores merged
The programtraining · github.com@1604916 · read 2026-09-19
The work
A fine-tuned checkpoint of Qwen3.6-35B-A3B, an open 35 billion parameter model, trained further to act as a coding agent: read a real repository, run shell commands, make edits over several turns. The miner uploads it to HuggingFace or Hippius and commits a v7 reveal on chain. What is paid for is a checkpoint that beats the reigning king in a duel, twice.
own docs · albedo.tech
King-of-the-hill trajectory-distillation subnet on Bittensor (netuid 97, finney).
How it is scored
A duel on sampled coding-task prefixes. Per task an evaluator model rolls three reference runs, reduces them to milestones and writes yes-or-no questions. King and challenger each roll every task twice; judge models answer each question, and a side's score is its mean yes-rate over its runs. A run that collapses into a command loop scores 0 unjudged. The challenger wins when its score beats the king's by 0.025.
code · raw.githubusercontent.com · src/albedo_eval_service/judge_core.py
CHALLENGER_WIN_MARGIN = 0.025
How the pay splits
Equal shares across a chain of at most five kings. A new winner takes slot 1 and the sitting kings shift down; the sixth drops off. The weight of 10,000 basis points is split evenly among the members, so the reigning king earns the same as a king it displaced. A member that has been deregistered, or an empty reign, sends its share to burn UID 0.
code · raw.githubusercontent.com · src/set_reign_worker/service.py
new_challenger_slot_1; shift_existing_down; split_weight_evenly
Who judges
Three validators were active on the 2026-09-09 chain snapshot, of 64 permitted. Commit-reveal is off; weights may change at most once per 100 blocks, and the code refreshes them every 101. The judge is a backend pipeline of services on one Postgres state machine: it ingests commits, validates files, runs the whole duel on remote GPUs with judge models, crowns, and sets weights. Nothing is sampled from miners.
chain · hyperparams-2026-09-19.json
commit_reveal_weights_enabled False, weights_rate_limit 100, max_validators 64; chain-2026-09-09.json active_validators 3
How often it turns
Emission settles every 360 blocks, about 72 minutes. The puzzle turns per duel, not per epoch: a challenger is evaluated when it reaches the front of the queue and must win two separate evals, so a first win only re-queues it at priority 0. A king stands until displaced, then fades out over four further coronations. A new registration is immune from deregistration for 5,000 blocks, about 16 hours.
chain · hyperparams-2026-09-19.json
tempo 360, immunity_period 5000
Who buys
Nobody is named. The front page offers a Chat with the King link to the current champion and nothing else: no price, no API, no customer and no unit of sale, on the site, in llms.txt or in the repository docs. The miner's income is emission.
our reading · albedo.tech
Chat with the King
What has been gamed
Re-rolling the duel: a hotkey is written to a used_hotkeys ledger once its submission finishes eval and cannot submit again, and a first win only re-queues the challenger, so the win must repeat. Copies: a secret-seeded sketch of each tensor rejects earlier entries and noised versions. Looping: 0.61 duplicate commands or a run of 5 scores 0 unjudged. Verdict-shaped text in a candidate's output is stripped before judging.
code · raw.githubusercontent.com · src/chain_guard/db.py
SELECT 1 FROM used_hotkeys WHERE hotkey = $1
5 of 7 fields rest on the chain or the code. A manifest is the site's reading of the subnet's program, not a review; each field links its source. All programs · What a field means · Correct a field
What it says
The subnet in its own words: the line the index took from its front door, and what the team has said on the record.
What it shows
Explained from zero by the index, from the subnet's own materials, for someone who has never heard of Bittensor.
SN97 Albedo is a playground game of king of the hill for coding models, where one fine-tuned model stands on top and earns most of the prize until a challenger knocks it off by a clear margin, twice.
The commodity, explained from zero
The output is a model checkpoint: a saved copy of an AI model's weights, the file needed to run it. Every entrant climbs from the same flat ground, the open model Qwen3.6-35B-A3B, and fine-tunes it, meaning they train it a little further on their own data for one job. The job here is acting as a coding agent: reading a real software repository, running commands, and making edits over several turns. The materials call the recorded runs "trajectories" and the method "trajectory distillation", a term they do not define (inferred from the scoring doc: learning from example runs of a stronger model).
A better coding-agent model could have a price, but Albedo's materials name no buyer, no price, and no unit of sale. The only public use is a "Chat with the King" link to the current champion. For someone to pay, the champion would have to beat the models they already rent; the site's benchmark panel should show that but rendered only "loading" when fetched.
Why it is on Bittensor at all
The honest answer is: an open contest with a public ledger. Anyone can register, upload a fine-tuned model to a file host, and commit a fingerprint on chain; the king is chosen by a fixed procedure rather than a lab's release schedule. The materials do not compare this with a centralized provider on cost, ownership, or censorship. Here the comparison breaks: on the playground, holding the hill is the whole point, but here the model on top is supposed to be the product, and nobody has said who wants it.
How the work gets done
Miners fine-tune the base model, keep its configuration files identical to the original, and publish it; a duplicate check rejects copies and lightly noised copies of earlier entries. Validators, the judges who set pay, run the challenger and the reigning king on the same hundred coding tasks from real repositories, then ask a judge model (GLM 5.2) a checklist of yes-or-no questions about each run; a run that loops on one command scores zero. A challenger takes the hill only if it beats the king by 2.5 points in two separate evaluations. Emission, the subnet's share of new tokens, is split across the current king and the previous four in a rolling chain, so a king pushed off the hill fades out rather than dropping to zero.
How you would know it works
No artifact a reader can check exists today, in those words. The dashboard at albedo.tech lists benchmarks, reign, queue, history, and fails, but each section shows "loading…" to a fetcher, and the chat link opens a chat app that explains nothing.
What is missing
The chain description is two words, "Alchemical intelligence". The repository has no README; the guides sit in a docs folder nothing links to. The chain contact appears on no page. There is no price, customer, status page, or support path. The win margin is 2.5 percent in the guides and 3 percent in the site's llms.txt, a file written for AI agents. Registration cost and miner hardware are not stated.
Go deeper
Sources for this explainer
Metaphor: a playground game of king of the hill. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.
How legible it is
Four audiences, four questions each, rated on what a first-time reader can find in five minutes. The scale measures findability, not quality.
On-chain identity 6/7- subnet_name
- Albedo
- github_repo
- github.com
- subnet_contact
- arbos@bittensor.com
- subnet_url
- us-east-1.hippius.com
- discord
- @arbos
- description
- Alchemical intelligence
- additional
- not set
Stakers and validators●●●●● 2.3
Should I allocate here? · rank 27 of 32 for this audience
Q1What it is●●●●● 3
King-of-the-hill trajectory-distillation subnet on Bittensor (netuid 97, finney).
Output is a crowned fine-tuned Qwen3.6-35B-A3B checkpoint, no unit or price, stated in a machine facing llms.txt linked from the front page nav, chain description reads Alchemical intelligence
Q2Who it is for●●●●● 1
nothing on users, customers, revenue or usage in own materials, the only sign of use is a Chat with the King link that opens an Open WebUI instance, third-party pages add nothing on demand
Q3How it resists gaming or fails●●●●● 3
SCORING.md names anti-gaming steps (reference leak filtering, prompt injection stripping, looped trajectories scored zero) and a failure rule (eval fails below 0.8 valid samples), lives in the repo docs folder with no README pointing to it, needs the mechanism vocabulary, llms.txt says a 3 percent margin while SCORING.md and MINING.md say 2.5
Q4Identity and documentation●●●●● 2
identity 6 of 7 but the discord field is the handle @arbos not a link, subnet_url is a hippius redirect page to albedo.tech, github resolves to a repo with no README, no staker or validator page
Miners●●●●● 2.8
Can I compete, and what wins? · rank 27 of 32 for this audience
Q1What it is●●●●● 3
Albedo is a king-of-the-hill subnet for Qwen3.6-35B-A3B language models.
Miner fine-tunes that model, uploads to HuggingFace or Hippius and commits a v7 reveal on chain, model constraints stated (F16 or BF16, no py files, metadata byte for byte), no worked example, three clicks from the front page via a repo with no README
Q2Who it is for●●●●● 2
Registration costs recycled TAO
with no figure (chain burn 0.21 TAO today), no hardware statement for miners anywhere, competitiveness stated only as beating the king by the 2.5 percent margin in two separate evals
Q3How it resists gaming or fails●●●●● 3
duel scoring by checklist with four equal weighted tag categories, a 2.5 point absolute margin and two SUCCEEDED runs to promote, rolling 5 slot king chain for emissions in llms.txt, no example numbers, no immunity or deregistration policy, margin contradicts the 3 percent in llms.txt
Q4Identity and documentation●●●●● 3
step by step guide with albedo CLI commands (validate, upload, check, register, commit), repo pushed 2026-09-09, no releases, no dates on the guide, no README to reach it from
Buyers and enterprises●●●●● 1.3
Can I use this today? · rank 28 of 32 for this audience
Q1What it is●●●●● 3
Chat with the King
link to chat.albedo.tech which loads an Open WebUI app, no product page, no API, no example, the rest of the front page renders as loading placeholders
Q2Who it is for●●●●● 0
no sourcesecondhandunchecked
no pricing and no customer type anywhere in own or third-party materials
Q3How it resists gaming or fails●●●●● 0
no sourcesecondhandunchecked
no status page, SLA, support address or support path on the site or in the repo, the eval-service-status.md file is an internal runbook, the discord field is a handle not a link
Q4Identity and documentation●●●●● 2
no API reference, subnet_contact arbos@bittensor.com is set on chain but appears on no page fetched
Newcomers●●●●● 1.8
What is this and why does it matter? · rank 31 of 32 for this audience
Q1What it is●●●●● 2
Trajectory-distillation king-of-the-hill Bittensor subnet
is the repo blurb and the only prose a reader reaches from the front page, chain description Alchemical intelligence is a tag, no plain paragraph on the site
Q2Who it is for●●●●● 1
own materials contrast the duel with a benchmark run, not with a centralized service, third-party search snippets supply the only plain claim about permissionless competition
Q3How it resists gaming or fails●●●●● 2
sections named benchmarks, reign, queue, history and fails exist on the front page but render as loading placeholders to a fetcher, the chat app loads, no dated benchmark, no explanation of the numbers
Q4Identity and documentation●●●●● 2
subnet_name Albedo matches the site header ALBEDO, description is not a sentence, url resolves through a redirect page, no about or learn page, discord field is not a link
Provenance
How this rating came to be. Verified means the scorer fetched the page and the words are on it; secondhand means concluded from code, absence, or a third party. The link badge is a separate automated check made before publication.
scored by agent b 2026-09-09, rubric v1.0merged 2026-09-09links verified 2026-09-09: 30 live, 0 unreachable, 0 manual
Something wrong? Corrections of fact are applied as they arrive during the window; rating disputes are batched at its close. How to file one · GitHub issue · email.