Pay for proof, not for a private checkpoint.
Cortex Subnet 100 is an autonomous research network on Bittensor. It pays for reproducible findings and real bugs. Nothing pays unless it reproduces. Every gate fails closed.
This is a human reading of the live protocol and the source paper. The PDF stays the dated source at the bottom. Live counts come from the Site API this minute, or they say unknown.
An autonomous research network on Bittensor.
Cortex exists so useful methods accumulate. Independent contributors submit a claim with the recipe to reproduce it. The network keeps the finding, not only the winning weights. Bittensor subnet 100 is the incentive layer. People still publish topics, pin the judge, and adjudicate bugs. Autonomous describes the research loop being built, not a claim that no one is at the console.
SN100 is verified research, a shared harness, and two live challenges: Proof and Bounty. Those methods feed Cortex App: Chat, Code, Bot, Image, and the other product surfaces. Subnet 100 is the research kernel, not the product people open.
The unit of work is a finding plus the code, a data manifest, a declared compute budget, and measured evidence. One miner might improve training speed. Another might improve data filtering. Keeping only a checkpoint can hide both recipes. Keeping the artifacts lets later work reuse either method after it is measured again, not because two good scores were added together.
A finding, the code, a data manifest, a compute budget, and measured evidence.
Proof and Bounty, as the paper proposes them.
Missing evidence refuses. A ready light is permission to try, not a payment guarantee.
Proof and OpenType are no longer offered on this site; the live challenge is Bounty and Hypertrain is a preview.
A recursive controller plans the next check.
The research loop uses a recursive controller to plan, read, and propose the next check. It steers investigation. It is not a second scoring judge, and it does not invent a Proof verdict.
In the control loop, the controller breaks a signed topic into checks, gathers prior artifacts and logs, and proposes the next experiment. A miner or operator still has to run that experiment under the declared budget. The digest-pinned judge and the hidden-holdout harness sit outside this loop. If reproduction fails, the controller cannot open the gate.
Break a topic into checks, gather context, propose the next experiment. Control, not settlement.
Override a digest-pinned judge, fill a missing holdout, or mint weight. If reproduction fails, the gate stays closed.
Recursive language model used for control: plan, read, propose. Settlement stays with the digest-pinned judge and the harness.
A planned check is not a score. Settlement stays with the pinned judge and the hidden-holdout harness.
Two live challenges. One sealed bundle.
Operators publish signed research topics. A miner answers with a claim, the recipe, and the FLOPs spent. A digest-pinned judge image re-runs the recipe. A separate harness measures a hidden holdout against a baseline sealed before the topic opened. Bounty is the other door: real, reproducible product bugs, priced by an operator.
Proof divides its allocation equally among open topics. A topic either pays the best qualifying result (winner-takes-topic, with exact ties shared) or uses a discovery split: a pass floor plus an improvement pool on the primary metric. Duplicate artifact digests do not earn a second time. The miner pays for the judge pod with their own Lium key. No open topic, or no pinned image: submissions are refused.
Trust-root split is 8000 / 2000 bps. A configured share is not revenue, equity, or a promised return.
A signed topic ships with a sealed baseline and a private holdout. You submit the claim, the code, and the declared compute; the same pinned judge runs for everyone, and you pay the judge pod. Winner-takes-topic or a discovery share pays a primary metric plus a pass floor, not the paper's full multi-metric novelty model; scores sum over open topics, and submissions are refused when no topic is open or no image is pinned.
Pair a dedicated Cortex Chat mining account and accept the terms, then file a real bug with exact reproduction steps. An operator reproduces it and prices severity from trivial to critical; pay is precision times mean severity, and noise scores zero. Duplicates, cosmetics, already-fixed reports, and fabrication burn weight; validators consume the signed adjudication and do not re-break the product.
The judge is still partial. Proof submissions live in memory. Automatic Proof reward emission is not wired into the service. A ready status is not an end-to-end payment guarantee.
How weight moves. What this page will not promise.
Bittensor emits to subnet 100. The trust-root split is Proof / Bounty at 8000 / 2000 bps. A miner's score becomes on-chain weight only after a sealed bundle is published and validators accept the board. Unmatched emission burns to UID 0. Retired challenges (relearn*, design, prism) have no trust-root row and earn nothing.
Validators verify signatures, the reward calculation, and that the bundle matches the published board. They do not re-run every recipe and they do not re-adjudicate every bug. An unsealed fallback is never a valid submit path. Missing prerequisites return a refusal, never an invented score.
A configured share is a mechanism number, not a cashflow. This page will not write an APY, a TAO price, a payout calendar, or a Cortex market-cap. If a figure did not come from the network or a named public source, it is omitted.
Unverifiable research. Static tests. Closed loops.
The source paper models a familiar contest: miners train separate checkpoints, recipes stay private, and a visible static evaluator picks a winner. The network keeps the chosen model and loses the methods behind the other runs. Repeated tuning to a test you can see improves that test without proving the work holds anywhere else.
That model is a design argument, not a claim that every other subnet works this way, and not evidence that Cortex already outperforms them. Private holdouts and pinned images reduce some attack surface. They do not guarantee an honest operator, a correct judge, or a finding that generalizes.
A checkpoint without a recipe cannot be rebuilt, audited, or reused.
A fixed, observable bench rewards memorizing the bench.
Private methods die with the round. The next miner starts from zero.
Miners verify. Gates fail closed. Baselines stay sealed.
Cortex changes the unit of work from “here are my weights” to “here is a finding, the recipe, and evidence it works.” Operators commit a baseline before a topic opens. The holdout stays hidden. The judge image is digest-pinned. Validators check signatures and the reward calculation; they do not re-run every experiment. Missing prerequisites return a refusal, never an invented score.
Today the building blocks exist: signed topics, submission intake, payout helpers, and signed bundles. Important gaps remain. The Python judge currently acknowledges and applies static checks rather than investigating arbitrary committed code. Proof state is in memory. Automatic leaf emission is not wired. Treat the schema below as the intended gate sequence, then read the honesty strip before you spend compute.
Validators check signatures and the allocation math. They do not re-run every experiment. Missing evidence refuses.
- 01A signed topic is published with a sealed baseline.
- 02You submit a claim, the recipe, and the compute you spent.
- 03The pinned judge re-runs it. Same image, no favourites.
- 04The harness measures the hidden holdout.
- 05Scores seal. Validators publish the board, or everything burns.
Two loops the paper wants. Neither is a payout promise.
The paper's infrastructure sketch is a continuous miner research loop on dense InfiniBand clusters: research feeds a training job, a checkpoint is measured against the sealed holdout, and a failed checkpoint rolls back to the last good stack. Accepted methods can feed the next job. Rejected ones do not become the shared stack.
Directional infrastructure from the paper: denser eval on InfiniBand clusters. Not a product date, not a revenue model, not a payout promise.
The second loop is how the harness itself improves. Proven artifacts form a research collection. A second, digest-pinned synthesiser may propose a change to the shared harness, training recipe, or data policy. That proposal must pass evaluation against the current stack before adoption. It does not merge raw miner weights. Two complementary findings do not prove they combine. That is another experiment.
Paper end state (§7): a second digest-pinned synthesiser. Not a shipped component. Two good results do not prove they combine.
Both loops are the intended end state in the paper (sections 7 and 8), not shipped product. No date. No revenue model. No APY.
Verified findings flow into Cortex App.
Cortex SN100 is the research network on Bittensor subnet 100. It pays for reproducible findings and real bugs. Verified methods, and the harness improvements they earn, feed Cortex App: Chat, Code, Bot, Image, and the surfaces that come next. People open the apps. They do not open the subnet as a product.
Verified research, a shared harness, Proof, and Bounty. Methods accumulate here.
The product suite people use day to day: Chat, Code, Bot, Image, and the other surfaces.
SN100 improves the research and training substrate. Cortex App inherits that improvement.
A Proof finding that holds on the hidden holdout can enter the research collection. A Bounty report that an operator reproduces can harden the product the apps sit on. When a proposed harness or training change beats the current stack, the shared substrate moves. Chat, Code, Bot, and Image inherit that substrate. Future surfaces inherit it the same way.
The subnet scores research. The apps serve people. A session in Chat is not a Proof submission. A Code edit is not a Bounty report. Weight moves only after the judge, the holdout, and a sealed bundle. The product can get better when those gates pass. It does not mint emission on its own.
SN100 is the research kernel. Cortex App is the product people open. This page does not promise a payout. Check /status before you spend compute.
The AI market is large. That is not a Cortex TAM.
Public “AI market” figures mix chips, cloud, applications, and research labour. They do not map onto subnet emissions and they do not imply a token price. This page does not invent a dollar TAM or a market-cap for Cortex.
For industry scale, read the Stanford Institute for Human-Centered AI AI Index, an annual, public survey of investment, capability, and adoption. Use it as order-of-magnitude context for the sector, not as a forecast for SN100 weight.
Source: Stanford HAI, AI Index (public report series, hai.stanford.edu). Category boundaries change by year. We omit a restated figure on purpose.
InfiniBand clusters. A 24/7 miner research loop.
The intended end state in the paper is a shared research collection plus a second, digest-pinned synthesiser that proposes updates to the shared stack (training recipe, harness, or data policy) and must itself pass evaluation before adoption. That agent is a goal, not a shipped component. The two schemas in chapter 07 are that argument drawn as a flywheel.
On the infrastructure side: denser evaluation on InfiniBand clusters, and a continuous miner research loop that could feed real-time adapted training. Those are directional. They are not a product date, not a revenue model, and not a payout promise.
If it cannot reproduce, it does not pay. If the bundle is not sealed, emission burns. If a number did not come from the network, this site will not write it.
Source, conformité, contact.
Cette page est la lecture principale. Le PDF est le document source daté, pour le conseil juridique et quiconque a besoin de l'artefact statique. Expérience Arcade par echobt. Document : A Proposal for a New Incentive Mechanism on Cortex (Bittensor Subnet 100): Bounty and Proof.