Leadpoet SN71 has merged a submission route for miner-built Python sales programs. The code reached the project’s main branch on September 5 after an isolated testnet exercise moved one submission through upload, execution, judging, persistence and public result retrieval.
The result was technically useful and competitively poor. The miner produced accepted output for 17 of 20 ideal-customer-profile tasks, then scored zero. The baseline scored 1.02.
Leadpoet preserved that failure in its validation record. It is the most important fact in the release because it separates infrastructure progress from model quality.
The submission path is merged to main. It is not enabled in production.
What miners can change
Leadpoet’s existing arena evaluates sales-intelligence agents against ICP tasks. The new path allows a registered miner to submit Python source that exposes the defined run_icp function interface.
Miners can change the model, prompts, tool routing and Python harness. Dependencies may be included under archive and binary-wheel limits. The current runner does not support arbitrary programming-language runtimes.
An accepted upload enters a round queue. Acceptance means the archive and registration checks passed; it does not mean the agent was scored, approved or rewarded.
Results remain private until the round is published. The public response then carries company outputs, per-ICP scores and an aggregate score. Unknown submission IDs and IDs from another round are intended to return 404 rather than disclose unrelated results.
The merged design gives miners more ownership over the agent being tested. It also transfers part of the operating cost to them.
Provider calls are funded by the submission
The route uses miner-supplied runtime credentials for approved model and data-provider calls. Leadpoet says a management credential is checked and discarded, while two runtime credentials are encrypted for later execution and judging. Submitted source and the validator sandbox do not receive the raw values directly.
There is no organizer-key fallback for miner execution. The miner pays the upstream provider charges attached to its submission.
This article does not reproduce the setup procedure. Public documentation should be read by operators who understand secret management, provider billing and the staged deployment requirements.
The repository records how credentials are intended to move. It is not an independent security audit. Tao Outsider did not inspect the production key-management environment, attempt credential extraction or endorse the custody model.
The test used isolated live services
Leadpoet’s validation report describes an isolated testnet-401 round named arena-2026-09-05-e2e2. The submitted agent ran in fresh gVisor sandboxes and was judged across 20 ICP slots.
The test used real provider calls, KMS, object storage, database migrations and registered testnet identity. It also used separate processes, a separate PostgreSQL database and a separate storage prefix on the gateway and validator hosts.
Rewards were disabled. There were no chain writes or production changes. The SQL path used a direct test transport rather than the production PostgREST route.
That boundary prevents two opposite misreadings. The exercise used real services in isolation and stopped short of a production launch.
The report also records a first round that was cancelled after a capacity problem and a later transient HTTP 500 caused by a closed local database connection. A subsequent read recovered without changing state. Leadpoet did not erase those events or claim the production transport had been proven.
Seventeen outputs can still be worth zero
The miner returned accepted output on 17 tasks and exhausted its allowed attempts on three. Schema-compliant completion did not earn a positive score.
Across four ICPs, it proposed companies that the judge rejected for identity, geography or business-stage mismatch. Empty outputs and exhausted attempts also received zero. The final score remained zero while the baseline reached 1.02.
This result shows why reliable completion cannot stand in for usefulness. A model may complete most calls, obey the response schema and still fail the commercial task.
The run did not prove successful downstream intent verification because candidate fit failed earlier gates. It provides no evidence that the submitted model could produce profitable leads, paying customers or better sales outcomes.
Production still has explicit work left
Leadpoet’s operator document requires staged schema and key-management changes before miner intake can be enabled. Production PostgREST wiring remains untested for this path. Older submissions must be allowed to finish under the previous version, and the service must be checked before and after the credential migration.
Those requirements are not footnotes. They are why the correct headline is about a merged workflow and isolated testnet exercise rather than a live miner marketplace.
The repository also notes thousands of passing CI checks alongside six failures reproduced on the main-branch base. They remained visible, so the suite cannot be described as entirely green.
Why the change matters for Bittensor DeMkt
Leadpoet is one of the Bittensor projects attempting to turn decentralized model competition into sales work. A credible market needs more than a central baseline chosen by the subnet team. Miners need a defined route for bringing differentiated software into the contest, paying for its dependencies and receiving inspectable results.
PR 174 builds much of that path. Its own zero-score result shows why infrastructure is only the beginning.
The next proof should come from production enablement, several independently operated submissions and scores that survive scrutiny across fresh ICP sets. Revenue or customer outcomes would require separate evidence.
At the September 5 checkpoint, TaoSwap showed 51 active miners on Leadpoet SN71 and an emission_value of 0.000001228. That is point-in-time subnet context. It does not show how many miners used the new submission system, because production intake is still disabled.
Leadpoet has merged a route for miners to compete with their own sales programs. The test result refuses to turn that engineering step into hype. The isolated system completed its run; the submitted program earned zero.
Sources
Leadpoet miner-model submission pull request 174
Miner submission design and production boundary
September 5 isolated testnet validation record
Was this article useful?
One tap feedback helps us improve each post.
