Skip to content

0031 — Caller-enrolled baseline execution

Date: 2026-10-20. Scope: W15, not scientific admission or final study evaluation.

Hypothesis: one executable path can bind the reviewed study, verified source import and frozen candidate configuration, then fit/choose candidates with only training rows and retain every threshold before scoring test. Baseline: those APIs are separate and only the authored comparison fixture joins them. Mutable paths are the joining module/CLI, authored dossier fixture, focused tests, package checks and documentation. Existing data grouping, numerical evaluators, study targets and historical results remain unchanged.

The command requires an expected admission-review identity supplied outside its input directory. The caller authenticates that review; document names, marker approvals or the command itself cannot provide human authority. Check the exact review documents before reading source data or constructing a model runtime. Bind the run plan, complete installed Python source/version inventory, dataset identity, import-plan identity, source approvals, split bytes, guard assets, seeds, input bounds and fixed study targets to those reviewed terms. Reproduce the imported dataset from its original source/lineage evidence.

Default execution rejects fixture dossiers. An explicit fixture flag admits only fixture imports and labels every result accordingly; it cannot turn fixtures into real-data evidence. The fixture uses invented review markers, local authored rows and a tiny ONNX marker graph. It supplies no genuine human label audit.

All selector/optimizer calls receive training rows only. Refit every selected family/seed, then calibrate every retained fit and the fixed guard. Persist all fit/calibration identities before the first test-scoring call. Preserve failed selection attempts, nonconverged refits and unavailable comparisons. Controls are fixed pass/block decisions. The compact encoder remains explicitly unavailable in this W15 command; W16 supplies it and W17 supplies the full study reports. The result never claims task-quality success or independent verification.

Use a new private output directory, exclusive synchronized writes and a retained failure record. Refuse existing output directories. Check elapsed time between stages and reserve output space before writes; these are cooperative limits, not an OS sandbox or a way to interrupt a blocked native operation. The foreground caller must enforce the reviewed total time/memory/storage budget and retain its resource receipt. Test reuse across separate output roots also remains a caller obligation. A complete fixture run does not waive the real study's label, contamination, final review, statistical power or evidence gates.