kadubon/Proof-Carrying-Skills--PCS-Core-: v.1: PCS-Core PoC2: Operational Evidence Release (S2/S3, Gemini API, Fail-Closed)

Proof-Carrying Skills (PCS-Core): PoC2 Operational Evidence Release (S2/S3, Fail-Closed, Gemini API) This record publishes the PoC2 operational evidence package for Proof-Carrying Skills (PCS-Core), including staged runs, aggregated reports, reproducibility fingerprints, and audit logs. BackgroundThis release is an empirical operational follow-up to the core PCS theory:Takahashi, K. (2026). Stop Recomputing for AI/LLMs: Proof-Carrying Skills for Compute-Saving Inference Reuse. Zenodo. https://doi.org/10.5281/zenodo.18490939 ObjectiveThe objective is to evaluate whether PCS-based reuse improves operational latency and cost versus full recomputation under fail-closed constraints, using real LLM API calls. Scope of this releaseRepository: Proof-Carrying Skills (PCS-Core)Model/API used in this evidence set: gemini-2.5-flash-lite via Gemini APIStages included: S2_PREPROD (poc2_s2_preprod_run1) S3_FULL_EVIDENCE (poc2_s3_full_evidence_run1)Hard gates required and passed: P1_CONFORMANCE_SMOKE = pass P3_STRESS_FAIL_CLOSED = pass P4_DETERMINISM_SAMPLE = pass hard_pass = true Main quantitative results (S2 + S3 combined)rows analyzed: 1260 (pass_rows = 1260)latency improvement ratio: mean 0.773993, 95% CI [0.754563, 0.790780]ms-cost improvement ratio: mean 0.620851, 95% CI [0.593802, 0.645646]token-cost improvement ratio: mean -5.187144, 95% CI [-5.627192, -4.752554]failure rate estimate: 0.0retry average per operation estimate: 0.002227 Operational interpretationWithin the tested envelope, PCS reuse improves latency and ms-based operational cost when reuse density is sufficient.The empirically observed ms-model threshold is stable at N_reuse >= 3 across both S2 and S3.Token-based savings are not established in this evidence set and are treated as a non-claim. Claim boundarySupported by current evidence:Under the tested settings in PoC2, PCS reuse provides positive latency and ms-cost gains when N_reuse >= 3, with hard safety gates passed.Not claimed: Universal gains outside measured workloads/settings/models Model quality improvement Token-cost savings under current telemetry completeness Direct transferability of numeric thresholds to other model families without replication Reproducibility and auditabilityRe-run commands:python PoC2/poc2_operational_runner.py --config PoC2/staged-configs/poc2_s2_preprod.yaml --out-dir PoC2/runs --tag poc2_s2_preprod_run1python PoC2/poc2_operational_runner.py --config PoC2/staged-configs/poc2_s3_full_evidence.yaml --out-dir PoC2/runs --tag poc2_s3_full_evidence_run1python PoC2/generate_poc2_results_report.py --runs-dir PoC2/runs --tags poc2_s2_preprod_run1 poc2_s3_full_evidence_run1 --out-json PoC2/poc2_s2s3_results_report.json --out-md PoC2/poc2_s2s3_results_report.md Primary evidence files:PoC2/poc2_s2s3_results_report.jsonPoC2/poc2_s1s2s3_results_report.json Key audit artifacts include protocol/workload fingerprints, immutable SHA256 fingerprints (config_sha256, runner_sha256, gemini_driver_sha256, plan_sha256), and per-call logs (*_gemini_requests.jsonl, *_gemini_responses.jsonl, *_gemini_errors.jsonl).Any post-hoc code/config change invalidates direct comparability unless re-run under updated fingerprints. Relation to the theory paperThis record provides operational evidence and reproducibility artifacts for a deployment-oriented evaluation path derived from:Takahashi, K. (2026). Stop Recomputing for AI/LLMs: Proof-Carrying Skills for Compute-Saving Inference Reuse. Zenodo. https://doi.org/10.5281/zenodo.18490939