PFP VIII: Conservative Grammar Growth and Source-Bound Evidence in the MSS Pipeline

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

PFP VIII: Conservative Grammar Growth and Source-Bound Evidence in the MSS Pipeline This is a source-pinned methods and audit study of the Proposal Fidelity Protocol (PFP, historically the Symbol Game). It separates extending syntax, checking a finite instance, and proving semantic preservation. The actual archived MSS-0/MSS-1 parser and maps recover the one old packet and preserve new packets in typed refusals; the left-inverse fact establishes injectivity on that domain, not a conservative theory extension. A finite flow fixture has zero objective gap but fails balance. Retained source-bound Lean receipts cover 39 selected theorems and a broader census of 576 local environment declarations, including ExactRat and generated declarations. These scopes do not certify the separate 52 external Level Engine payloads. The present review verifies the retained source bindings and reports its fresh local Lean replay as blocked before startup. The historical C001–C050 census has 40 solver packages with two checker modules, two specifications calling for model measurements, and eight other specifications. Package presence is not a correctness receipt. The C005 results retain two separate runs (36/50 and 38/50 for Arm 2), corrected exact McNemar calculations, missing per-case-data limits, and the 405B model/budget protocol deviations. Eight archived K4 traces document one run per model and condition; correct repairs appear in answer-bearing assisted prompts. Historical terminal scores remain the retained report's replay. The package includes thirteen exact archived source/report/trace files with full revision and hash bindings, the Lean sources and reports, a strict offline graph-consistency runner with complete version-2 execution receipts, 52 offline tests, and a 16-question Landau-style quiz with worked answers and reproducible changed-input exercises. No live model measurement, independent fresh kernel pass, validated educational outcome, external deposit or DOI is claimed.