Save-Panda - Subnet Design Proposal
Save-Panda is a subnet on Bittensor designed to build the world's most accurate, continuously updated, and openly accessible food allergen database. Our corevision is to eliminate a critical global information asymmetry: over 520 million people living with food allergies cannot reliably verify whether a product is safe to consume, because allergen information is scattered across thousands of incompatible sources—manufacturer websites, regulatory filings, and crowdsourced platforms—with no single machine-readable, trustworthy reference.
Save-Panda addresses this by creating a decentralized network of AI agents (miners) that autonomously research, cross-reference, and verify allergen information for food products, competing to produce the most accurate structured records. Each record follows the EU 14 allergen standard—the strictest food labeling regulation globally—and is keyed by Universal Product Code (UPC/EAN barcode), making it universally addressable without geographic dependency.
The subnet's incentive mechanism rewards miners who produce verified, evidence-backed records validated against Open Food Facts (4-star completeness tier) and regulatory databases. By combining Bittensor's competitive dynamics with the global food safety imperative, Save-Panda creates a self-sustaining ecosystem that continuously improves the world's food allergen data infrastructure—one that no single company, government, or volunteer community can build alone.
Save-Panda uses a proportional performance-based reward system. Miners earn TAO proportional to their composite score in each validation epoch. Proportional distribution (rather than winner-takes-all) ensures sustained participation across a broad miner base while preserving competitive pressure to improve accuracy and coverage.
Emission and Reward Logic: A Winner-Takes-All System
Score = 0.60 × Allergen_F1 + 0.25 × Label_Accuracy + 0.15 × Calibration
| Component | Definition | How Measured |
|---|---|---|
| Allergen_F1 | Precision × Recall on EU 14 allergen fields | vs. validator ground truth labels |
| Label_Accuracy | Exact match on binary dietary labels (vegan, vegetarian, gluten-free, halal, kosher) | Binary: 0 or 1 per field |
| Calibration | Expected Calibration Error (ECE) of per-field confidence scores | 90% confidence must map to ≥90% accuracy |
Incentive Alignment for Miners and Validators
For Miners: Higher evidence quality and accuracy → higher composite score → more TAO. Miners are incentivized to run more capable agents, consult more independentsources, and maintain fresh records. The 90-day freshness decay means passive miners who stop updating lose rewards—continuous effort is economically required.
For Validators: Validators maintain the ground truth dataset (Open Food Facts 4-star records + 500 manually audited products). Their economic stake is tied to subnet credibility: a validator that manipulates scores undermines the subnet's value and their own TAO holdings. Ground truth versioning is public and auditable by any participant.
Mechanisms to Discourage Adversarial Behavior
Several layers of defense are built into the mechanism to reduce gaming, improve data quality, and protect validator efficiency.
Hidden Evidence Bundles. Validators maintain a private set of canary products whose correct allergen data does not appear in Open Food Facts or any commonly crawled source. These products can only be answered correctly by querying manufacturer regulatory filings or regional databases that require active research. Canary products are rotated quarterly and never publicly disclosed. A miner that consistently answers canary products correctly provides strong evidence of genuine multi-source research. A miner that fails canary tasks disproportionately signals surface-level scraping rather than deep verification.
Hard Gate for High-Risk Fields. The six highest-risk allergens under EU 14 — peanut, tree nuts, shellfish, fish, milk, and egg — are the most common triggers of anaphylaxis. For these fields specifically, two conditions must both be satisfied: confidence ≥ 0.95 and ≥ 3 independent sources in the evidence chain. If either condition fails, the field is scored as 0 regardless of whether the value is factually correct. This makes it impossible to obtain a passing score on high-risk fields through lucky guessing, overclaiming, or single-source citation.
Pass-Rate + Score Floor.
Each epoch, a miner must satisfy both thresholds simultaneously:
If either threshold is violated, the miner's epoch weight is set to 0. This eliminates cherry-picking strategies — miners cannot selectively submit only easy products to inflate their average while skipping difficult or high-risk products.
For each product in each epoch, all miner submissions with field-level similarity > 97% are grouped into a single cluster. The cluster receives the score of its earliest-submitted record only; all later members of the cluster receive 0, regardless of their individual accuracy. This is stronger than per-pair copy detection: it neutralizes coordinated collusion rings that produce near-identical records without directly copying each other's CIDs, and removes any economic incentive for sharing evidence collection outputs across miners.
Any miner found to have fabricated URLs, cited pages whose content does not support the claimed data, or submitted evidence chains with fetch timestamps after the submission block (post-hoc construction) receives an immediate epoch score of 0 and a permanent fraud strike on their hotkey. After 3 fraud strikes, the hotkey is permanently deregistered from the subnet. The expected value of systematic fraud is therefore: guaranteed 0 score on detected epochs plus permanent loss of all future mining capacity, making fraud economically irrational at any scale.