# Current primary priority — 2026-10-05 Europe/Moscow

First active task: accelerate and finish the complete native generator, C++/CUDA compute + Rust control,1000 epochs x10000 generated roots, seeded normative schedules and all correctness/certificate/exact-label/transform/last-occurrence suffix/durable resume/close/HF contracts. Repetitions are allowed; dev/test exclusion is disabled by the dated human changes below.

Current isolated acceleration candidate: random_repeat_typed_portfolio_candidate_20261004. Append reuses its private validated root; four reader modes reuse one stored-row validation; suffix and transform builders reuse immutable checked input portfolios. Every offered raw/mapped trajectory and each actual returned batch retain validation.18 actual CPU tests passed, including56 contiguous normative positions/all14sizes/all4family certificate fixtures/actual lifted source, exact JSON parity against frozen pre-reuse helpers, malformed input/schema refusals and durable crash/close. This is bounded correctness evidence, not native compiler quality or a measured GPU speedup.19 Linux checks and matched full10000 target timing remain required.

Molab reports kernel not found after a reconnect attempt. Current remote native process state is unknown, last visible generated frontier401; do not infer completion, termination or authorize a competing runtime from stale output. No runtime restart performed. Primary remains sole remote/cleanup owner; expert ACK received for the limited typed reuse boundaries.

Historical priorities and dated replacements retained below; they do not replace the first active task above.

# Current primary priority — 2026-10-04 22:47 Europe/Moscow

First active task: simplest complete native random train path, C++/CUDA compute + Rust control, fixed1000 epochs x10000 generated roots, seeded normative size/family schedules, correctness/certificates/exact labels/scheduled transforms/last-occurrence suffix rules, durable per-epoch resume/close, full targetMolab timings and HF artifact storage.

User requirement verified in project chat "Проверь работу агента", human message01a10873-593f-7e60-ac62-15a9dd39a1db: "Лучше просто случайные и всё. Похуй на повторы". New train policy allows repetitions both within and across epochs. Remove global identity registry, duplicate rejection, UNIQUE(matrix_json), registry config/payload/registration/reconciliation overhead. Dev/test exclusion remains disabled as previously authorized. Keep seeded random draws; no extra entropy assumption. Version manifest/contract/digest/store explicitly: repetitions_allowed=true, global_train_matrix_uniqueness_required=false. Do not claim old unique-policy tests verify this new policy.

History retained: uniqueness-preserving correctness-only implementation superseded2026-10-04 22:47, source/log pins in unique_policy_superseded_20261004_2248/history_manifest.json (21 actual CPU tests under that old policy). Bank128 screening optimization superseded2026-10-04 22:14. Reduced smoke subset cancelled2026-10-01 16:31. These are not current first tasks.

Legacy recovery v2 remains terminalexit1 (retained train has sidecars), preserved. Do not restart generation or finish legacy archive machinery ahead of the new simple full10000 target run. No subagents, no competing cleanup/GPU launch, no D writes, no unapproved runtime restart. Primary sole remote owner. New policy source remains isolated until compiled/tested and target run verified; no speculative speedup claims. Only service Python cell explicitly authorized earlier.