Retrospective map of the frozen campaign inputs

This file maps retained work; it is not a newly invented pre-request frozen plan. The protected campaign planned research, feasibility, scouting, focused quality, restoration and publication stages. The current product's complete serving-comparison/v1 metric-coverage contract was not populated by that older launcher. Full serving comparison remains incomplete.

The public scope is the seven-task coding fixture and configuration-specific synthetic context, tool-use, image and short-output diagnostics. See workloads, configuration and native artifacts.

Nineteen coding cohorts include historical controls, rejected policies, reversal controls and fresh larger-context profiles. The 1,024-to-4,096 budget follow-up was selected after failures; it is post-hoc tuning, not held-out validation. Original strict 128-word capacity scouts had zero eligible requests. Separate 16-word capacity populations were planned before their requests.

Restored-reference context work was reduced before traffic to 15 cases with a 900-second limit; tool-use was reduced before traffic to a 180-second limit. Unsupported native reasoning-control mode was rejected before traffic and replaced with default/no override while the launch remained thinking-off. No result was converted from failed to passed.