ChronoLLM is Bittensor subnet 38 (SN38), a project in the Distributed Training lane. One model per year, no peeking ahead
A subnet that pays miners to build language models with amnesia: each one is trained only on data up to a single past year, so it cannot cheat by knowing what happened next.
It is an educational deep dive, not a recommendation to stake. For live price, pool depth, and performance, always open the live subnet page.
What is ChronoLLM (SN38)?
ChronoLLM (Subnet 38) is a Bittensor subnet where miners compete to train language models that are "chronologically consistent." Each model is allowed to learn only from information available up to a specific past year, so a 2018 model knows nothing about 2019 or later. The subnet is run by the team behind CrunchDAO, the data-science competition platform (contact listed on-chain as [email protected]).
In plain English: It is like training a historian sealed off from the future. Ask the 2018 model about 2018 and it answers using only what was known by the end of that year, with no hindsight about what came after.
Centralized reference points: No direct equivalent; ML contest platforms (Kaggle, Numerai) for structure
How miners and validators work
Like every Bittensor subnet, SN38 is a two-sided market: miners produce work, validators score it, and emissions follow those weights under the network’s dTAO economics.
- Miners train a collection of models, one per year, each only on data available up to that year's cutoff, then upload them to HuggingFace and submit on-chain.
- Validators run inside a Trusted Execution Environment (a secure hardware enclave) and score each model in two stages: first checking it has not leaked knowledge from after its cutoff year, then judging answer quality in a round-robin tournament refereed by an LLM judge.
Mechanism and design
Per the project's repository, miners train one language model per year (the README lists a range starting at 2013, and a recent commit caps evaluation at 2024). Each model must use the project's ChronoGPT architecture, be under 2 billion parameters, and ship in safetensors format. Miners upload finished models to HuggingFace and register them on-chain.
Validators do the grading inside a Trusted Execution Environment running on Phala Cloud with Intel TDX hardware. The README gives two reasons for this: the evaluation dataset must stay private so miners cannot overfit to the test, and remote attestation lets anyone verify the validator is running the correct, unmodified code. Scoring runs in two stages. Stage one checks every model for chronological consistency, that it has not absorbed knowledge from after its cutoff year. The top ten miners advance to stage two, where their models answer open-ended questions and an LLM judge compares every pair in a round-robin tournament; the win rate becomes a quality score. The final score combines the two as 0.7 times the consistency score plus 0.3 times the quality win rate. The design is winner-takes-all: only the single highest final score earns emissions in a round, and rounds run weekly, starting each Monday at 12:00 UTC. ChronoLLM cites the ChronoGPT paper (arXiv:2510.11677) and publishes reference models under the manelalab HuggingFace account.
Why the project exists
- The problem it solves: Lookahead bias. Standard LLMs are trained on data from every era at once, so when they are used for financial backtesting or historical analysis they already "know" what happened next. A model that has read about the 2020 crash cannot honestly simulate a 2015 analyst. ChronoLLM aims to produce models that are sealed to a point in time.
- The opportunity: Time-sealed models are useful anywhere you need to reconstruct what was knowable at a past moment: backtesting trading strategies without hindsight contamination, studying how narratives evolved, or evaluating old forecasts fairly.
- The Bittensor advantage: Building one trustworthy model per year across more than a decade is a lot of training work. A competitive subnet crowdsources that effort and pays for the best result, while the secure enclave keeps the test set hidden so miners cannot quietly overfit to it.
- Traction signals: Early. The repository's initial commit is dated May 15, 2026, the README states the subnet is "currently in testing phase on mainnet," and TaoSwap reported no active miners at the time of this snapshot. Net TAO staking flow over the past week was positive.
What stakers should check before buying alpha
Subnet storytelling is cheap. On-chain and product checks are not. Use the same framework as our subnet metrics guide:
| Check | Why |
|---|---|
| Real product / docs / repo | Separates shipping teams from narrative-only slots |
| Pool depth (TAO) | Your exit door: thin pools amplify slippage |
| Net flows (7d / 30d) | Under flow-based emissions, capital votes with its feet |
| Emission share | Whether the market is funding this subnet right now |
| Ownership concentration | Large holders can move a small alpha market |
| Immunity / deregistration | Post-immunity weak slots can lose the race (see immunity guide) |
Pull the live numbers on ChronoLLM · SN38 rather than trusting any static screenshot in a blog post.
Risk factors
- Very early and unproven: The repository's initial commit is from May 2026, the README itself flags a testing phase, and development sits with two contributors. The mechanism is documented but has not been demonstrated at scale.
- Winner-takes-all rewards: By design only the single top miner earns each weekly round. That is a sharp incentive to be best, but it concentrates rewards and may discourage smaller or newer miners from competing.
- Concentration: The top-100 gini sits around 0.72, indicating concentrated ownership and stake distribution. Large positions could swing pool dynamics.
- Enclave dependency: The anti-gaming guarantee rests on the Phala and Intel TDX execution environment and its remote attestation. The integrity of scoring is only as strong as that enclave and the off-chain backend it talks to.
How this fits a TAO portfolio
Many holders keep a root vs subnet split: root for calmer TAO exposure, subnets like ChronoLLM for thematic upside (and higher variance). If you use TaoScale, you can size SN38 from your wallet without handing over custody. See non-custodial staking.
- Live market: SN38 on TaoScale
- Browse and stake: /invest
- More subnet explainers: /blog/subnets
FAQ
What does ChronoLLM do on Bittensor?
It is like training a historian sealed off from the future. Ask the 2018 model about 2018 and it answers using only what was known by the end of that year, with no hindsight about what came after.
What is the SN38 alpha token?
When you stake TAO into subnet 38, you receive that subnet’s alpha. Its price versus TAO moves with the subnet’s pool (see what is dTAO).
How do I research ChronoLLM before staking?
Read the mechanism above, open the official repo/docs when available, then verify pool depth, flows, and activity on SN38. Cross-check with the metrics checklist.
Is this financial advice?
No. TaoScale publishes educational explainers and live tools. Decisions to stake are yours alone.
Not financial advice. Educational overview of ChronoLLM (SN38) compiled from public project materials and ecosystem documentation. Mechanisms and market structure change, so verify live data and official sources before acting.
