MODELS_DEFERRED.json: the worklist for the later checkpoint port

Standing policy through this consolidation has been evals and scores now, weights later. This records what was deferred, where it lives and what of it is already here -- 2,368 GiB across 8 rows -- so the port starts from a worklist instead of a re-survey.

Two rows are not merely weight-deferred and are the real outstanding work: the gemma4_26b charter-graft s2 pilot (set B, 96.9 GiB, holds the charter x2 = 0.746 vs 0.423 held-out finding) and the two template-diversity studies. Neither has scores here.

Sizes came from list_repo_tree, not repo_info(files_metadata=True): that call silently truncates its sibling list on large repos and under-reported scimt-dispatch-models as 8,486 files / 649 GiB against the true 11,515 / 1,352 GiB. Recorded in the file because it reads exactly like data loss.

MODELS_DEFERRED.json

@@ -0,0 +1,137 @@
+{
+ "schema_version": "models_deferred_v1",
+ "updated": "2026-09-14",
+ "what_this_is": "The worklist for the later model-checkpoint port. Standing policy across this consolidation has been: take evals and scores now, defer weights. This file records WHAT was deferred, WHERE it lives and WHAT of it is already here, so the port is a worklist rather than a re-survey.",
+ "read_this_first": [
+  "Two rows are NOT merely weight-deferred -- they have no scores here either, and are the real outstanding work: the gemma4_26b s2 pilot (set B) and template_diversity_v1 / template_response_diversity_v1.",
+  "Sizes were measured with list_repo_tree, NOT repo_info(files_metadata=True): that call silently TRUNCATES its sibling list on large repos and under-reported arcadia-impact/scimt-dispatch-models as 8,486 files / 649 GiB against the true 11,515 / 1,352 GiB. A shrinking file count from repo_info is not evidence of deletion.",
+  "MODELS_DEFERRED in build_clean_repo.py excludes a deferred profile from the generic planner. A unit listed here but missing from that frozenset will be quietly re-planned in full on the next pass."
+ ],
+ "total_deferred_gib": 2368.0,
+ "rows": [
+  {
+   "unit": "gemma3_27b_190m_clause_asym",
+   "repo": "arcadia-impact/scimt-dispatch-gemma27b-clause-asym-v1",
+   "prefix": "gemma3_27b_190m_clause_asym/",
+   "files": 494,
+   "gib": 236.6,
+   "stages": [
+    "midtrain",
+    "dolci",
+    "aft (agreement + charter_only, step 512)",
+    "data"
+   ],
+   "scores_here": "scores/gemma27b_clause_asym_v1/",
+   "batteries_here": "batteries/gemma27b-clause-asym-v1/",
+   "note": "Gemma-3-27B clause-asymmetric replication, COMPLETE 2026-09-14. Code is untracked on no branch; run_record/launch/source.bundle is the only durable copy."
+  },
+  {
+   "unit": "glm45_air_190m_clause_asym",
+   "repo": "sidbaines/scimt-dispatch-harder-episodes-glm + campaign repos",
+   "prefix": "glm45_air_190m_clause_asym/",
+   "files": null,
+   "gib": 647.0,
+   "stages": [
+    "midtrain 199",
+    "dolci 199 + a sharded duplicate",
+    "four AFT cells"
+   ],
+   "scores_here": "scores/ (campaign grid)",
+   "batteries_here": "batteries/v5-glm/glm45_air_190m_clause_asym/",
+   "note": "Evals and scores taken 2026-09-14; models deferred. Excluded from the generic planner via MODELS_DEFERRED so a re-plan does not silently pull 647 GiB."
+  },
+  {
+   "unit": "wave era (true midtrain / SDF-late / graft / dose)",
+   "repo": "arcadia-impact/scimt-dispatch-models",
+   "prefix": "multiple",
+   "files": 11515,
+   "gib": 1352.1,
+   "stages": [
+    "aft_wave_v2 571.23",
+    "aft_wave_retrain 150.92",
+    "sdf 147.64",
+    "deconfound_sdf_v1 123.26",
+    "graft_dose_v1 61.42",
+    "midtraining_4epoch 49.21",
+    "gate2_midtrain4 49.21",
+    "sft_4epoch 49.21",
+    "sft 49.21",
+    "midtraining 49.21",
+    "aft 20.14",
+    "rl_grpo 15.70",
+    "aft_elicitation_v1 6.54",
+    "aft_wave_x0p5 6.53",
+    "grafting_v1 2.51"
+   ],
+   "scores_here": "scores/wave_v1/, scores/grafting_v1/, scores/graft_dose_v1/, scores/three_way_midtrain_sdf_graft/",
+   "batteries_here": "batteries/dispatch-wave-v1/, batteries/dispatch-grafting-v1/",
+   "note": "Measured 2026-09-14 with list_repo_tree. The wave PARENTS are additionally in jbostock/scimt-dispatch-midtrained-sft-v1 @ 527f0b6c (1033.5 GiB), which is the pin the wave itself used; grafting_v1 and graft_dose_v1 exist ONLY in the arcadia repo."
+  },
+  {
+   "unit": "gemma4_26b charter-graft s2 pilot (set B)",
+   "repo": "sidbaines/scimt-dispatch-gemma4-26b-charter-graft-s2-pilot-v1",
+   "prefix": "",
+   "files": 121,
+   "gib": 96.88,
+   "stages": [
+    "grafts",
+    "aft"
+   ],
+   "scores_here": null,
+   "batteries_here": null,
+   "note": "NOT YET MOVED AT ALL -- neither scores nor batteries. Holds the charter x2 = 0.746 vs 0.423 held-out finding. This is the one outstanding set, not just deferred weights."
+  },
+  {
+   "unit": "elicitation_ablation_v1 part 2",
+   "repo": "sidbaines/scimt-elicitation-ablation-v1",
+   "prefix": "elicitation_ablation_v1/part2/",
+   "files": 395,
+   "gib": 27.89,
+   "stages": [
+    "adapters (every save) for 3 of 6 framed cells"
+   ],
+   "scores_here": "scores/elicitation_ablation_v1/",
+   "batteries_here": null,
+   "note": "Study paused at 3/6 cells; absence of the rest is an interruption, not a result."
+  },
+  {
+   "unit": "template_response_diversity_v1",
+   "repo": "arcadia-impact/scimt-prior-coins-template-response-diversity-v1",
+   "prefix": "",
+   "files": 65,
+   "gib": 3.6,
+   "stages": [
+    "LoRA step 256 + 512 on plain gemma-3-12b-it"
+   ],
+   "scores_here": null,
+   "batteries_here": null,
+   "note": "NOT MOVED. Parser-methods study on a NON-midtrained base; not a dispatch result."
+  },
+  {
+   "unit": "dispatch_v5 / harder-episodes GLM adapters",
+   "repo": "sidbaines/scimt-dispatch-harder-episodes-glm",
+   "prefix": "<profile>/<arm>/aft/",
+   "files": 195,
+   "gib": 3.92,
+   "stages": [
+    "AFT adapters"
+   ],
+   "scores_here": "scores/costsweep_v2/",
+   "batteries_here": "batteries/v5-glm/",
+   "note": "Cost-sweep responses and scores are here; the adapters are not."
+  },
+  {
+   "unit": "template_diversity_v1",
+   "repo": "sidbaines/scimt-prior-coins-dispatch-sdf-aft-v1",
+   "prefix": "extensions/template_diversity_v1/",
+   "files": null,
+   "gib": null,
+   "stages": [
+    "3 substrates x {baseline, step512}"
+   ],
+   "scores_here": null,
+   "batteries_here": null,
+   "note": "NOT MOVED (scores or weights). Its 100 templates ARE already in the repo indirectly: the campaign's held-out-surface eval prompts are rendered from them."
+  }
+ ]
+}