scimt-dispatch-harder-episodes-glm

LoRAs and responses of the harder episodes study (science-of-midtraining, branch sid/dispatch-harder-episodes, experiments/prior_coins/dispatch_v5/), served 2026-09-14 on five published GLM-4.5-Air parents of the dispatch campaign. Renamed from scimt-dispatch-v5-glm on 2026-09-14; the old name redirects and the tree is unchanged.

Layout: <profile>/<arm>/

• aft/<cell>/adapter/ -- the three AFT LoRAs trained on the harder tables (agreement, mixed_coin, charter_only), one per cell
• eval/<battery>/<endpoint>/responses.jsonl -- responses of every endpoint (pre_aft, v5-<cell>, campaign-<cell>) on the batteries v5, canonical and costsweep_v2, with an EVAL_COMPLETE.json receipt per directory
• *.json -- per-parent run receipts (host, restore, staging, training, eval, publish)

Parents: glm45_air_190m/{charter,coin,control}, glm45_air_1b/charter, glm45_air_190m_clause_asym/charter. The campaign-* endpoints are the campaign's own published step-512 adapters (the corrected #1c 2% draw for the 190M arms). Scoring and the written results live on the branch above (dispatch_v5/results/); the cost-sweep scores also on sid/dispatch-final-v1 (dispatch_final_v1/costsweep_v2/).