TrainingQualityVerifier

Overview

TrainingQualityVerifier.circom is a standalone zk-SNARK circuit that proves a language-model’s training meets or exceeds a required quality threshold without revealing raw training samples, model outputs, or proprietary scores. It serves as the “first gate” in the ZK-AgentMesh verification pipeline and can also be embedded in any project that needs zero-knowledge assurance of model-training quality.

---

Purpose

The circuit enables an agent developer to convince a verifier that:

1. Every individual training sample scores at least min_quality_threshold.

2. The average quality across all samples meets the same threshold.

3. The proof ties back to a prior on-chain training commitment and model weights hash, preventing post-hoc tampering.

All of the above are proven without exposing:

• The underlying data or model responses.

• Exact per-sample scores or private metrics.

---

Inputs

Private Inputs

Group	Name	Description
Training data	training_samples[n_samples × 32]	One Poseidon-hashed field element per token (flattened)
Model outputs	model_responses[n_samples × 32]	Poseidon-hashed response tokens
Quality metrics	quality_scores[n_samples]	Integer score 0–1000 per sample
Integrity	training_seed	Salt used during the commitment phase
Model hash	model_weights_hash	Poseidon hash of final model weights

Public Inputs

Name	Description
agent_id	256-bit unique agent identifier
min_quality_threshold	Minimum acceptable score (0–1000)
training_commitment_hash	Commitment published before training
creator_address	EVM address of developer

Public Outputs

Output	Meaning
quality_verified	1 if all checks pass, otherwise 0
average_quality	Mean score (integer field element)
training_proof_hash	Poseidon hash binding agent, seed, avg. quality, creator
model_integrity_proof	Poseidon hash binding model weights to seed and agent

---

Circuit Logic

1. Per-Sample Check Each quality_scores[i] is constrained with GreaterEqThan to be ≥ min_quality_threshold.

2. Average Calculation AverageCalculator sums all scores and enforces average × n_samples == sum.

3. Average Threshold A second GreaterEqThan gate ensures average_quality ≥ min_quality_threshold.

4. Conjunction of Conditions quality_verified is the product of

   • the average-threshold flag

   • the cumulative product of all individual flags If any single check fails, the product is zero.

5. Hash Commitments

   • training_proof_hash = Poseidon(agent_id, training_seed, average_quality, creator_address)

   • model_integrity_proof = Poseidon(model_weights_hash, training_seed, agent_id)

These hashes are referenced by higher-level circuits or on-chain contracts.

---

Compilation

# Requirements: circom 2.x, snarkjs, pot16_final.ptau
circom TrainingQualityVerifier.circom \
       --r1cs --wasm --sym --c

snarkjs groth16 setup TrainingQualityVerifier.r1cs \
                    pot16_final.ptau \
                    tqv_final.zkey

snarkjs zkey export verificationkey tqv_final.zkey \
                                         tqv_verification_key.json

---

Proof Workflow

# 1. prepare input.json with private + public signals
node TrainingQualityVerifier_js/generate_witness.js \
     TrainingQualityVerifier_js/TrainingQualityVerifier.wasm \
     input.json witness.wtns

# 2. generate proof
snarkjs groth16 prove tqv_final.zkey \
                       witness.wtns \
                       proof.json public.json

# 3. verify locally
snarkjs groth16 verify tqv_verification_key.json \
                       public.json proof.json

public.json will contain the four public outputs, ready for on-chain submission.

---

Security

• Private inputs never leave the prover’s environment.

• Poseidon hashing maintains circuit efficiency on BN128.

• Tampering with model weights is detectable through model_integrity_proof.

• The circuit must be recompiled if n_samples or n_metrics change.

---

Status

• Current version: stable for integer quality metrics (0–1000).

• Planned: floating-point support via fixed-point encoding; integration of additional metrics per sample.