Skip to main content

3.4 Anti-Gaming Mechanisms

1. Synthetic Ground Truth Tasks (15–20%)​

At a 17.5% injection rate, validators secretly replace a real benchmark task with a synthetic variant derived at runtime. The variant has the same description and reference answer as the original task, but a new opaque task ID (syn_<8-hex>) and tighter budget/latency constraints.

Miners cannot distinguish a synthetic from a real task. Validators score it using the same deterministic methods (ROUGE-L, test pass rate, schema validation) — no LLM judge.

The derivation is fully reproducible given the CSWON_SYNTHETIC_SALT:

import hashlib
seed = f"{SALT}:{validator_hotkey}:{block}".encode()
task_id = "syn_" + hashlib.sha256(seed).hexdigest()[:8]

Full synthetic task protocol →

2. VRF-Keyed Per-Validator Task Schedule​

Copy resistance: Miners cannot serve a cached plan because the task they receive is derived from hash(validator_hotkey + block_height). Copying the current top miner's last submission will not reproduce the correct response to a different validator's query.

import hashlib
seed       = f"{validator_hotkey}:{current_block}".encode()
h          = hashlib.sha256(seed).digest()
task_index = int.from_bytes(h, 'big') % len(benchmark_tasks)
task       = benchmark_tasks[task_index]

Different validators query different tasks at the same block height. Cross-validator score comparison uses distributional statistics over the rolling window, not identical-task point comparisons.

3. Scoring Version Enforcement​

Validators encode SCORING_VERSION as an integer in __spec_version__ and as a human-readable string in axon.info.description. Mismatches are detectable from the live metagraph.

4. Dynamic Benchmark Rotation​

Anti-overfitting: When >70% of miners score above 0.90 for 3 consecutive tempos, the task is deprecated and rotated. Miners cannot memorise the benchmark.

5. Execution Sandboxing​

Validators execute all workflows in isolated Docker containers, tracking actual TAO costs, latency, retries, and step completions.

6. Temporal Consistency Checks​

Sudden unexplained performance jumps trigger a manual audit flag in the validator dashboard.

7. Completion Ratio Enforcement​

Submitting a single-step workflow for a multi-step task always results in a proportionally penalised S_success.

8. Budget Ceiling​

budget_ceiling = min(constraints["max_budget_tao"],
                     1.5 * workflow_plan["total_estimated_cost"])

Validators abort execution when cumulative cost reaches the ceiling, preventing malicious miners from forcing expensive sandbox executions.

Incentive Alignment and Penalties​

C-SWON does not introduce on-chain slashing — Bittensor does not support automatic stake slashing at the protocol level. Instead:

  • Yuma Consensus bonds: Validators whose scores deviate from consensus earn progressively less
  • Delegation flow: Delegators move stake away from misbehaving validators
  • Governance control: Subnet owner can adjust validator limits and prune permits
← Previous3.3 Output Quality Scoring
→ Next3.5 Scoring Version Control
IndexDocumentation Index