# Development before optimization

The latest user instruction authorizes a uniquely named new learned placement study, superseding the earlier baseline-only scope for this new milestone. Earlier status files remain complete; no duplicate or modified frozen earlier run is authorized. The supplied public report could not be accessed through the web tool; the completed local report, handoff, configuration, source, closeout, attempts and resource ledger were read. No separate primary state was inspected.

Reuse is source-based: unchanged host.py and growth program.py; legacy.py is an exact copy of the previous construct.py, preserving B_original. local.py duplicates the prior growth scheduling algorithm to retain its local states, then schedules a new pure local placement program. Runtime changes add evaluator-only write commit/delivery observations. They neither send acknowledgments nor alter state/transitions. The new workload/event generator is task V3: initialization plus 64 unforced weighted events, finite IDs recycled after settlement. The episode cap rises from 100000 to 110000 to cover 72 events at at most 1394 ticks per gap; the inherited transaction physics and 1292-tick physical bound are unchanged.

A 25-parameter four-unit tanh scorer uses raw charged path distances and workload weights, with fixed scalings and one explicit boundary preference. THETA0 has authored hidden weights, zero readouts, and boundary coefficient -.25. This faithfully reproduces the previous feasible boundary allocation through the same bidding/reservation shell. No learned coefficients exist at first freeze. No development optimization scores are used to hand-tune the scorer. The analytical comparator explicitly computes immediate arithmetic; the exact solver and per-site cost matrix are evaluator-only and never supplied to the policy.

Development fixtures are manual K2 and two L4/K4 hosts only, plus finite tiny solver matrices and sixteen workload-frequency draws. They test sensitivity, actual local actions, feasibility, objective/trace agreement, baselines and replay. All attempted fixture jobs and errors remain in attempts.json and numbered logs. No extra agent, scheduler, GPU, external numerical service or global installation is used.

First runner startup failed at sandbox-denied ps preflight before any worker launch; retained and conservatively charged one second. Pre-freeze source catch improved to preserve a future preflight exception. The inherited monitor runs through bounded command escalation; no scientific scores informed this correction.