Stop Recomputing for AI/LLMs in practice: this project implements a deterministic verification layer for compute-saving inference reuse.

Instead of recomputing expensive model inference end-to-end, a caller can accept outputs when a compact proof-carrying receipt verifies under explicit bounds.

AI/LLM deployment is increasingly constrained by repeated inference cost (latency, energy, and money). PCS reframes acceptance as a deterministic verification problem:

• compute once with an untrusted provider
• verify many times with a small trusted checker
• accept only if contract + evidence satisfy fail-closed rules

This repository implements that path with a small trusted computing base and no-meta trust assumptions.

Compared with typical caching or heuristic reuse, PCS adds cryptographic and semantic guarantees:

• No-meta trust boundary: providers are untrusted; acceptance is checker-driven
• Deterministic OPVM predicates: bounded gas/size/depth, no unbounded execution
• Replay-resistant invocation binding: receipts bound to checker-issued invocation headers
• Verifiable Trace Receipt (VTR): output commitments plus optional trace evidence
• PCS-Blob support: Merkle inclusion proofs for event logs and blob chunks
• GLUE receipts for composition: deterministic bridge checks across sequential skills

• PCS-Core deterministic checks
• VTR verification (inline trace, Merkle trace, event assertions)
• PCS-Blob verification (included_events, included_chunks)
• GLUE composition verification (from_bundle + glue_receipt + to_bundle)

Out of scope in this repository:

• BRS/SRR runtime semantics (schema slots reserved)

To support operational claims (real deployment cost/latency), this repository includes an optional and non-gating experiment frame:

• runner: PoC/e6_optional_cost_runner.py
• config template: PoC/e6_optional_cost_config.example.json
• protocol: PoC/experiment_master_plan.yaml
• guide: PoC/README.md

E6 follows the paper cost terms:

• CostPCS = Costrun + Costcheck + Costhash + Costregistry
• amortization term: Costcert / E[Nreuse]

Important boundary: E6 results are external operational evidence and do not change PCS-Core verifier correctness/safety claims.

Operational-claim readiness in E6 requires command mode with workload disclosure, sufficient sample size, and positive 95% CI lower bounds for both cost and latency improvements.

• spec/: normative and conformance documents
• schemas/:
  • pcs-core-v1-bundle.schema.json
  • pcs-v1-extensions.schema.json
  • pcs-v1-extensions-placeholder.schema.json (legacy alias)
• reference-checker/: deterministic verifier implementation (Python)
• test-vectors/: accept/reject vectors for bundle and schema modes
• compatibility-suite/: deterministic replay runner (runs vectors twice)
• PoC/: experiment protocol and optional E6 external frame tooling
• PoC2/: cloud-LLM operational benchmark with staged production-load rollout

1. Verify a standard VTR bundle:

python reference-checker/verifier.py --bundle test-vectors/accept/basic_echo/bundle.json

2. Verify a GLUE composition bundle:

python reference-checker/verifier.py --bundle test-vectors/accept/glue_composition/bundle.json

3. Run full compatibility suite:

python compatibility-suite/run_vectors.py

4. Run optional E6 frame (demo config):

python PoC/e6_optional_cost_runner.py --config PoC/e6_optional_cost_config.example.json --out-dir PoC/runs --tag e6_optional_cost

PoC2 provides a cloud-LLM operational benchmark with:

• fail-closed API handling
• deterministic verifier path for reuse checks
• ABBA crossover and audit JSONL artifacts
• ms-based and token-based cost reporting

python PoC2/poc2_operational_runner.py --config PoC2/poc2_operational_config.yaml --out-dir PoC2/runs --tag poc2_gemini

Guide: PoC2/README.md Engineer-facing experiment report: PoC_results.md Production-load staged plan: PoC2/poc2_production_rollout_plan.yaml Stage profiles and generated configs: PoC2/poc2_stage_profiles.yaml, PoC2/staged-configs/

Proof-Carrying Skills, PCS-Core, Stop Recomputing for AI/LLMs, compute-saving inference reuse, deterministic verifier, Verifiable Trace Receipt, GLUE receipt, Merkle inclusion proof, no-meta trust, inference cost reduction, latency reduction.

Takahashi, K. (2026). Stop Recomputing for AI/LLMs: Proof-Carrying Skills for Compute-Saving Inference Reuse. Zenodo. https://doi.org/10.5281/zenodo.18490939