[0z5a]

RFC465 starts with ordinary Ouro GRPO through the standard train.py and RolloutManager, with separate trainer and rollout resources. The existing qualification recipe always selects resident colocation and the Thinking/RLTT packet, so it cannot directly establish that initial support profile.

Add an identity-selected ordinary-Ouro K4/FP32 GRPO startup/recovery packet and an explicit resource-layout choice while preserving the resident default and all existing Thinking inputs. The profile keeps three updates, 96 new training completions per logical run and a continuous-versus-fresh-process comparison. No reference, critic, KL, early exit or speculation is included. The small launcher change reuses the existing source/model/state/output audit chain.

Check	Result	Speed comparison
Actual recipe boundary, resident/separate × phases	30 CPU tests pass	Not a timing comparison
Real ordinary model configuration plus legacy commands	Six CPU-MOCK probes pass	Not measured
Public training input preparation	Two preparations byte-identical; official test split untouched	Not measured
Actual Dataset/Tokenizer import audit	Original reader pending; resident import exit1 for missing local Torch source; no package updates	Not measured
Thor source preparation	1,205 source files hash-verified; authentication required; CUDA NOT_RUN	Not measured
Standard real GPU engine, useful updates and fresh resume	NOT_RUN; requires two physical GPUs and admitted complete runtime	Not measured
Full RFC465 fault/default/lifecycle gates, reward convergence and scaling	Remain open	Not measured

Validation retains exact source/input/metadata and wheel-RECORD hashes, XML and original local wait records. No model weight bodies or remote environment changes are added. CPU command checks are explicitly separated from backend acceptance and convergence evidence.

Stacked on draft #36 (ops/scalerlt-initial-paper-v01 at a6c0b51). Only this acceptance-preparation increment is included. The current experimental RLT pin and pending upstream dependency acceptance are documented.