P13 pre-output technical amendment: OSF API version 2.2; unchanged scientific design
Abstract
This is a technical pre-output amendment of the approved P13 registration https://osf.io/pu3mb/ (registered 2026-10-04T13:29:38.207859Z). The original record and artifacts remain preserved. A pre-generation verification error was corrected by changing only two OSF request lines to pin API version 2.2, which returns an explicit UTC timestamp. Strict UTC validation remains unchanged. The original generator/protocol/manifest, actual API snapshots, official source references and exact production diff are retained in provenance. No scientific P13 answering output exists. The populations, contexts, model revisions, generation settings, scoring, H1/H2 rules and analysis are unchanged. Both registrations and this technical correction will be disclosed. The run window starts from this final technical registration. P13 tests whether correctness fragility under grade-preserving evidence exchanges predicts harmful losses after real AI system updates, and whether successor plurality voting reduces losses while preserving observed net gain. The incumbent is Qwen3-8B/BM25; updates are Qwen3-32B/BM25 and Qwen3-8B/bge_reranker_large. Confirmatory analyses use the immutable 629-query P5 estimation split (NQ 180, HotpotQA 175, FEVER 180, SciFact 94). The 420 qualification queries and full 1,049-query panel are descriptive. All 12,588 reconstructed contexts and 18,982 planned generations are fixed before P13 answering outputs. Prior P5b GPT outputs, qualification retrieval scores and parent/P7 results are known. Qwen3-32B previously supplied 1,278 semantic-equivalence judgments involving 250 estimation questions; this exposure is disclosed and is distinct from answering from the P5 contexts. No prior Qwen answering run on these contexts was identified in the recovered source set; the author reports no prior P13 answering outputs. This does not assert absence from pretraining. The amended protocol explicitly freezes a new verified context reconstruction rather than claiming recovery of the original context bytes, and declares both scoring corrections before outputs. MOCK tests and feasibility simulations supply no scientific P13 result. The attached PROTOCOL_P13_AMENDED.md is authoritative for frozen H1/H2 thresholds, event-count and identity gates, paired bootstrap, 14-day stopping rule, software/model pins, descriptive analyses and full reporting regardless of outcome. This is a labeled benchmark audit, not a label-free deployment or safety claim. The P13 folder is a distinct extension of the source project https://osf.io/9x35z/; the earlier P7 registration remains separate. Code frozen UTC: 2026-10-04T13:46:22.593905+00:00 PROTOCOL_P13_AMENDED.md SHA-256: da1a43bd8d1e2afe9e3198b6f4ce1bd6117a51a644ff0339386e88455677cefd SHA256SUMS.txt SHA-256: ae340becf2223d9c87b84c155d053d63629f4e1de60491298452274c2125b2c7 Amended contexts SHA-256: 6a6c12e02b4897571b6fa26b2ec9bf143667f3855529f1a2bef547c6bf872696 P13_UTC_Compat_Registration_20261004.zip SHA-256: 040af6edb2b554c492aefa995588396fdeecb0954362582d7aab6fdd6729f25b
