October 14 label and execution-boundary review

简体中文

Each new field access and counter in the AST inventory was checked individually. The inventory is a review aid, not proof of capability.

Source	New access or counter	Allowed scope and behavior evidence
agent_research/memcon.py	Policy update / episode counter	Update only after real memory-backend operations and terminal feedback; counters are not agent capability. Separate 60-episode observable-record retrieval diagnostic and persistence tests for maintenance and feedback. No old Task answer/plan reads.
system_one/regional_transfer.py	Prediction gold / fallback gold	Input-file consistency, scoring public predictions, and validation temperature/route selection only. Test gold does not select. Regression confirms changing test labels does not alter temperature, threshold, or validation route.
system_one/request_audit.py	Scorer-only positive_labels	Request compiler reads displayed evidence and current question only. Independent request bytes do not change with other questions, route labels, or scoring labels. Labels enter metrics, not the model.
WeGR	Training SID / logged reward	Fit tokenizer on train items/sessions only; validation selects arm, test only reports. Scores and query logs are public MovieLens proxies, not production labels.
EmbeddingGemma 2	Caption–image pair IDs	Encoder sees actual images/text. IDs reach the multi-positive scorer after encoding. All four preset MRL dimensions are reported without held-out selection.

No new agent path returns gold answers. MemCon is a memory operator; JEV request audit is a protocol. Neither automatically qualifies for Evolve promotion. Five deferred papers have no executable adapters or metrics.