The first LLM without lookahead bias. Each model is trained only on data available up to its cutoff date, preserving backtest integrity for quant research.
SN 38. Rank 22 of 32 by emission
ChronoLLM
Emission1.4%of the block reward · rank 22 of 32
Participants10 miners · 6 validators
Registration0.74 TAO
Verified2026-09-09links checked and scores merged
The programtraining · github.com@a1f1aa7 · read 2026-09-22
The work
A set of twelve language models, one per vintage year from 2013 to 2024, each trained only on text available up to its cutoff so it cannot know later events. Each is at most 2.2 billion parameters in safetensors, uploaded to Hugging Face and registered on chain. What is paid for is a slate that passes the leak test and wins text-completion duels.
own docs · raw.githubusercontent.com
Each model must be trained only on data available up to its cutoff year. A model for 2018 must know nothing about 2019 or beyond.
How it is scored
Stage 1 is a gate: per year the model must recognise at least 70 percent of pre-cutoff items and at most 10 percent of post-cutoff ones, the year average must beat -3.0. Up to 30 survivors then duel round-robin: each writes completions for prompts a judge model invents fresh that round, and GPT-5.4 picks the winner per pair. Final score is leak_weight x normalised leak plus quality_weight x win rate.
code · raw.githubusercontent.com · sn38/neurons/validator.py
final_scores[uid] = leak_weight * normalized_leak[uid] + quality_weight * win_rates[uid]
Sources disagree The code's fallback defaults at the pinned commit are leak_weight 0.7, quality_weight 0.3 and top_n_for_quality 10; api.chronollm.com/config, re-fetched 2026-09-22, still returns leak_weight 0.0, quality_weight 1.0 and top_n_for_quality 30, so the README's final_score = quality_win_rate is what runs.
How the pay splits
Winner-heavy exponential decay. The top ten by final score are paid exp(-0.8 x rank^0.9) normalised to one, which gives about 52, 23, 12, 6, 3, 2, 1 percent and under, and rank eleven and below get nothing. An emission_pct below 1 sends the remainder to the owner UID; the live config has it at 1.0. The owner UID itself is excluded from the ranking.
code · raw.githubusercontent.com · sn38/neurons/validator.py
rewards = np.exp(-0.8 * (np.arange(n) ** 0.9))
Who judges
Six 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. A validator runs inside an Intel TDX enclave on Phala Cloud with an H200, fetches private evaluation data from the team's backend over attested TLS, downloads every model, re-runs the leak test and duels itself, sets weights once, and exits. The compose hash is publicly attestable.
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 6
How often it turns
Emission settles every 1,440 blocks, about 4.8 hours, four times the usual tempo. The puzzle turns weekly: a round opens every Monday at 12:00 UTC, all prior submissions are wiped, miners resubmit, and validators evaluate the previous round once. A new registration is immune from deregistration for 50,400 blocks, about a week, ten times the usual 5,000.
chain · hyperparams-2026-09-19.json
tempo 1440, immunity_period 50400
Who buys
Quants and researchers, on three tiers with no price. Individual: full weights free to download and self-host. Academic research: free hosted API for non-commercial work behind a request form. Institutional: on-premises deployment fine-tuned on the buyer's data, with support and SLAs, priced to the deployment via a contact form. No customer of the subnet's output is named and no usage figure appears.
own docs · chronollm.com
The full model, free to download and run on your own hardware.
What has been gamed
Copying a rival's weights: an identical checksum is rejected and the first submitter keeps priority; near-copies are caught pairwise by SVD spectral distance under 0.001, weight cosine over 0.95 or logit KL under 1.0, invariant to rotation, permutation and scaling. Overfitting the leak set: it stays inside the enclave, and duel prompts are invented fresh each round. Stale models: the slate is wiped every Monday.
code · raw.githubusercontent.com · sn38/template/svd_gate.py
SVD_THRESHOLD = 0.001 ... COSINE_THRESHOLD = 0.95 ... KL_THRESHOLD = 1.0
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.
SN38 ChronoLLM pays miners to train AI models frozen at each year from 2013 to 2024, like a shelf of old encyclopedias, so a finance researcher testing an old strategy consults only what was knowable then.
The commodity, explained from zero
A backtest is how a quantitative investor checks a strategy: run it on past data and see whether it would have made money. If the model reasoning has read tomorrow's newspapers, the backtest lies. The site's line is "Standard models make a backtest look brilliant, then fall apart the moment it trades live." The materials call this lookahead bias.
ChronoLLM's product is language models with a hard knowledge cutoff, each "trained only on data available up to its cutoff date", so the 2018 volume knows nothing of 2019. The unit is a model checkpoint per year: saved weights a researcher downloads and runs. The site offers free download of the full weights for individuals, a free hosted API for academic research, and "use-based" pricing for institutions, with no figure given.
The buyers named are quant researchers, asset managers, and financial institutions. The site says the operator, Crunch, serves "institutional clients including ADIA Lab and the Broad Institute of MIT and Harvard"; it does not say those clients use ChronoLLM.
Why it is on Bittensor at all
On the materials, Bittensor supplies the training labour: instead of one lab training twelve models, miners compete to train them each week. The site names "Subnet 38" only in its header and footer; on what decentralization changes for the buyer on cost, ownership, or censorship, nothing is stated yet. The encyclopedia comparison breaks at the checking step: an old volume's print date is on its spine, but a model's cutoff has to be tested, so validators fire "thousands of post-cutoff leak probes" and a reader can run them too.
How the work gets done
Miners train one model per year for 2013 to 2024, each at most 2.2 billion parameters, upload them to Hugging Face, and register the set on chain. Validators, the judges who decide pay, reject any model whose weights match an earlier entry, then run a leak test: a model must recognise most pre-cutoff items and miss post-cutoff ones, or it is out. Up to thirty survivors fight round-robin duels on fresh prompts judged by a large model (GPT-5.4), inside a trusted execution environment, a sealed part of a processor that proves the code ran untouched. The top ten share emission, the subnet's share of new tokens, with the winner taking about 52 percent and each rank below roughly half the one above; every Monday the slate is wiped and miners resubmit.
How you would know it works
The nearest artifact is the arXiv paper "Instruction Tuning Chronologically Consistent Language Models", which describes the method and its leak tests. The leaderboard at leaderboard.chronollm.com could not be read by a fetcher. No dated benchmark is on the site.
What is missing
The miner guide gives no registration cost and no hardware beyond "GPU". Institutional pricing has no number and there is no API reference, only a short Python example behind a request form. The chain contact address appears on no page. Discord is a channel in the shared Bittensor server. There is no status page or SLA. Third-party pages say the subnet is in a testing phase; its own pages do not say.
Go deeper
Sources for this explainer
Metaphor: a shelf of encyclopedias, one frozen at each year. 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
- ChronoLLM
- github_repo
- github.com
- subnet_contact
- crew@crunchdao.com
- subnet_url
- chronollm.crunchdao.com
- discord
- discord.com
- description
- Competitive training of chronologically consistent Large Language Models
- additional
- not set
Stakers and validators●●●●● 3.3
Should I allocate here? · rank 12 of 32 for this audience
Q1What it is●●●●● 4
The first LLM without lookahead bias.
Output is point in time model vintages sold as downloads, hosted API or on prem, pricing tiers named with no figure for the institutional tier, chain description matches the README opening
Q2Who it is for●●●●● 2
A crowdsourced ML research platform of 12,000+ data scientists serving institutional clients.
describes the operator Crunch, no usage figure, no named customer, the leaderboard at leaderboard.chronollm.com is JS rendered
Q3How it resists gaming or fails●●●●● 4
README states top 10 emissions on exponential decay (about 52, 23, 12 percent), a leak score gate, LLM judged duels, weight checksums and SVD near duplicate detection, TEE evaluation and a weekly reset, config endpoint at api.chronollm.com/config is live, no subnet failure modes
Q4Identity and documentation●●●●● 3
identity 6 of 7 (additional not set), github resolves, url redirects from crunchdao.com to chronollm.com, discord is a channel in the shared Opentensor server, contact crew@crunchdao.com is on chain only with forms on the site, no staker page
Miners●●●●● 3.3
Can I compete, and what wins? · rank 18 of 32 for this audience
Q1What it is●●●●● 4
Models can use any architecture loadable by HuggingFace AutoModelForCausalLM
one model per year 2013 to 2024, max 2.2B parameters, safetensors, a models.json pinned to commit SHAs and committed on chain, no worked example
Q2Who it is for●●●●● 2
tdx.xlarge (8 vCPU, 16GB) is the recommended minimum.
applies to the self test TEE, training hardware is only GPU for training, registration command with no cost (chain burn 0.74 TAO today), no competitiveness statement
Q3How it resists gaming or fails●●●●● 4
emission split with example percentages, leak gate then round robin duels judged by an LLM, weight hash and SVD anti copy, all submissions clear weekly so miners must resubmit, config endpoint shows leak_weight 0.0 and quality_weight 1.0, split between README and miner.md
Q4Identity and documentation●●●●● 3
three step guide (train, models.json, register and submit), no releases, no dates, repo pushed 2026-09-07, third-party pages say the subnet is in a testing phase
Buyers and enterprises●●●●● 3.0
Can I use this today? · rank 15 of 32 for this audience
Q1What it is●●●●● 4
The full model, free to download and run on your own hardware.
plus a hosted API for academics behind a Request API Key form and a four line Python example, no endpoint URL shown
Q2Who it is for●●●●● 4
Free hosted access for non-commercial research, so you can build on ChronoLLM.
Individual free, Institutional use based with no figure, Academic free, customer types named
Q3How it resists gaming or fails●●●●● 2
contact form only, support mentioned for the institutional tier, no status page, no SLA, Discord is a channel in the shared server
Q4Identity and documentation●●●●● 2
no API reference beyond a Python snippet, chain contact crew@crunchdao.com appears on no page, the site uses forms
Newcomers●●●●● 3.5
What is this and why does it matter? · rank 15 of 32 for this audience
Q1What it is●●●●● 4
Each model is trained only on data available up to its cutoff date, preserving backtest integrity for quant research.
on the front page, consistent with the chain description, backtest and vintage are left to the FAQ
Q2Who it is for●●●●● 3
The financial sector needs a new type of large language model that doesn't cheat.
a concrete difference against standard LLMs, not against a centralized provider, Bittensor appears only in the header and footer
Q3How it resists gaming or fails●●●●● 3
Thousands of post-cutoff leak probes let you verify the model's time boundary yourself.
two arXiv papers linked, leaderboard page is JS rendered, no dated benchmark on the site
Q4Identity and documentation●●●●● 4
subnet_name ChronoLLM matches the site, description is a noun phrase without a finite verb, url resolves via redirect, About and FAQ sections on the front page
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.