PART 1
0/3Tokenizer
Build a character-level vocabulary and encode/decode strings to and from token ids.
VeriLLM
VeriLLM
Wrap cached GPT inference in Merkle commitments and spot checks, then simulate a decentralized committee with rewards and slashing.
Every source step is its own lesson with intuition, concepts, correctly rendered MathJax mathematics, implementation, tests, mistakes, and a checkpoint.
PART 2
0/13Embeddings and Causal Self-Attention with KV Cache
Implement token and positional embeddings, the linear projection primitive, and a single-head causal self-attention module that reads and writes a KV cache.
004embed tokens+5 XP005add positional embeddings+5 XP006linear projection+5 XP007compute attention scores+5 XP008scale attention scores+5 XP009apply causal mask+5 XP010softmax attention weights+5 XP011weighted value sum+5 XP012project qkv+5 XP013append kv cache+5 XP014scaled dot product attention with cache+5 XP015apply output projection+5 XP016single head causal self attention+5 XP
PART 3
0/7Feed-Forward, LayerNorm, and Transformer Block
Assemble the position-wise feed-forward network, layer normalization, residual add-and-norm sublayer, and the full transformer block that updates the KV cache.
PART 4
0/5LM Head, Prefill, and Autoregressive Decoding
Project hidden states to vocabulary logits, run prefill over a prompt, and generate tokens one decode step at a time while recording per-step state.
PART 5
0/8Merkle Commitments over Decode Steps
Hash tensors and decode-step states into leaves and build a Merkle tree with inclusion-proof generation and verification.
PART 6
0/8Prover Transcript and Spot-Check Verification
Run the prover to produce outputs and commitments, assemble the public transcript, and implement seeded spot-check verification that re-executes audited decode steps.
PART 7
0/4Tampering, Detection Probability, and Verifier Cost
Tamper with transcripts, derive the detection probability under k audits and a corruption fraction, and quantify verifier cost as a fraction of full re-execution.
PART 8
0/9Decentralized Committee, Incentives, and End-to-End Evaluation
Sample verifier committees, aggregate votes by majority, implement rewards and slashing, and run honest and malicious rounds reporting end-to-end verification cost.