0004 — Return the December study to frozen one-shot robustness

Date: 2026-09-29. Status: final planning direction requested by the researcher; implementation, capacity verification, and final-test lock remain pending.

The researcher requested a final v3 plan based on the assessment of the v3 draft and its fit to the existing architecture. The controlling December execution design is now the final v3 plan. It supersedes the December scope of decision 0003 and the v3 draft while preserving the v1/v2 protocols, existing implementation, history, and artifacts. This decision does not relabel v2 as complete, mutate an existing protocol hash, deploy the tracker, or authorize novel-test access.

The study returns to the final prospectus's one-clean-exemplar-per-unseen-action question with frozen generic MotionBERT, an identical 8,704-to-2,048 linear head, physical P=8/K=4 batches, seeds 7/17/29, three required objectives, class-disjoint clean development selection, and all-100-action final training. The confirmatory contrast is full contextual recipe versus standard contrastive recipe. Its different margin and regularizer prevent attribution to contextual neighborhoods alone. SupCon remains descriptive.

The fixed duration is 20 epochs, including three complete engineering pilots. There is no performance-dependent horizon or best-epoch selection. Each arm receives the same three learning rates and three paired seeds, followed by nine fresh final heads. The optional regularizer-only control is omitted now for bounded December scope and capacity, before new pilot/selection scores. No post-opening extension is permitted.

Retain the validated development episode and official novel anchors. Adopt nested corruption severities and a fixed auxiliary-training torso fallback. Use a conditional within-action performance-cluster bootstrap with seeds, actions, anchors and realized corruptions held fixed. Report absolute corrupted scores, every seed, every cell and leave-one-action-out sensitivity. A null or negative contextual result is a successful scientific outcome.

Build on the clean committed v2/Week 3 successor architecture while preserving the dirty local checkout. Version the retained frozen pipeline for v3; keep v2 benchmark internals and historical settings. Produce fresh v3-bound caches after extraction/validation code is frozen. All training/extraction versions must respect the same dataset-level opening state.

The historical aggregate loader deserializes all records before selecting auxiliary examples. Existing seal evidence means no novel retrieval preprocessing, embedding, visualization, scoring, or tuning access; it does not establish that novel bytes were never deserialized. Audit this fact and every historical opening location before pilots. The new final extractor must validate the full v3 lock and atomically record opening before loading novel annotations. Any prior outcome access blocks the planned confirmatory opening.

September 29 SCC inspection found 272.80 decimal GB free, leaving about 58.05 GB above the 200 GiB reserve. The final plan budgets 100 decimal GB incrementally, including slack and the existing architecture's separate primary/official artifacts; current capacity is short by approximately 41.95 GB. Additional accessible artifact capacity is a concrete prerequisite. Previous clean extraction success is not full-campaign feasibility. The October 18 gate requires measured complete pilots and corrupted development extraction/scoring, a GPU allocation calendar, actual researcher availability, and the entire workload/storage forecast with at least 25% slack. Failure requires a documented result-blind amendment before selection; do not shrink one arm or change the test after seeing outcomes.

Independent-research governance remains effective; no advisor sign-off is required. Retain source/licensing safeguards and the recorded BU determination. Adopt the matching machine-readable v3 configuration, AGENTS.md, public fallback, and guarded tracker migration during implementation. Preserve live versions, evidence, private ownership/RLS, closed-week audit rules, and actual time records. This planning decision records no invented hours or completed operational gates.