Spark-Hermes
One pinned agent. A competition on how it is instructed. Every result checkable by a stranger. SN74 Gittensor.
A strategy is prose — a SOUL.md and optional skills — and nothing else. The validator runs the same pinned Hermes
agent and model under every strategy, in a sealed sandbox, on task instances whose grading criteria were committed
before submissions opened. A strategy is paid for beating the strategy-less baseline on the same instances. The
crowned strategy is merged as the incumbent; every verified trajectory is exported as training data.
- Live: https://gittensor-model-hub.github.io/Spark-Hermes/live/
- Submit:
submissions/README.md - Pins:
docs/pins.md· Findings:docs/spikes.md - Training data (public):
gittensor-model-hub/spark-hermes-rounds
Layout
| Path | What |
|---|---|
| sh/validator/ | the round loop (orchestrate.py), episode runner and sealed sandbox, grader, inference proxy, batch runner |
| sh/scoring/ | v2.py — payment: paired delta against the baseline over the pooled window, one-sided lower bound; crown.py — the crown, on this round alone |
| sh/predicates/ | the closed predicate vocabulary tasks are graded with; copied into the runner image at build |
| sh/exports/ | the king's SFT rows and DPO pairs from a closed round, leak scan, Hugging Face upload |
| sh/cli/ | miner (tasks / submit / status), lint (the contract CI runs), attest (the hotkey's signature), scorecard, mock_miners |
| sh/web/ | the static page for a closed round |
| | one directory per hotkey, prose only — merged when crowned |
| | every round's tasks (what miners were shown), window, seal, , , , checks, scorecards, and — the scored tasks, when miners were shown previews; — digests of rounds minted ahead |
| | the site served by Pages: landing, live dashboard, round pages; pins and findings |