Publish nested-reference MP-OPD probe 01c7bf4d

MP-OPD-MULTIREF.md

@@ -0,0 +1,49 @@
+# Nested 1/4/8-reference mechanics probe
+
+Prepare4 groups of13 unique prompts: one rollout,8 select,4 eval. Select pools
+are nested prefixes1/4/8, predetermined by a preparation seed, never chosen by
+eval. Mean outer NLL equally weights references; each reference is token-mean.
+One rollout and teacher credit computation per group, reused across counts.
+Frozen adapter and model are unchanged across candidates. Weighting networks
+start identically and train independently per count on earlier select groups.
+Eval never trains or selects; all count outcomes are reported. Four groups
+remain four independent units, not12. The ordinary summary describes the
+largest count; paired-select-counts.json and results.jsonl include all counts.
+Checkpoint weighting.pt is the largest-count network; weighting-select-N.pt
+files retain each count. This is not a resume-complete training checkpoint.
+
+## Company run, only after a new HF bundle pull
+
+From the new isolated source checkout returned by the existing HF pull flow:
+
+```bash
+bash "$ABC_WORK/source/experiments/runai/transfer/04-prepare-multiref.sh"
+# Explicit GPU boundary; only after the card is free and not reserved by eval:
+bash "$ABC_WORK/source/experiments/runai/transfer/05-run-multiref.sh" 1
+```
+
+Prepare reuses the pinned selected.parquet or downloads its pinned HF revision
+if missing, verifies SHA256, excludes all conflicting-reference prompt groups,
+and writes a new exclusive file. This default source potentially appeared in
+SFT: results are mechanics only, not unseen-data generalization evidence.
+All52 selected prompts are disjoint across roles/groups, but no claim of
+semantic deduplication or exclusion from external model pretraining is made.
+For a scientific test, prepare a separate authorized train/dev reference source
+and use --exclude-prompts for SFT/train/prior probe sets. Never use benchmark
+test references. Keep the evaluation reference set fixed across select counts.
+
+Runner uses exact company SFT Gemma and Qwen7B paths, FP32, maxnew64,
+virtualLR.1, maxref1024,15-minute timeout. Long references fail rather than
+truncate. It refuses the eval queue's GPU lock or >=1GiB occupied VRAM, so do
+not stop eval to make room. A timeout preserves partial logs and returns124;
+partial groups are not efficacy evidence. No automatic W&B upload.
+
+## Validation
+
+CPU fake-HF end-to-end tests cover legacy and1/4/8 modes, unchanged real
+parameters, identical atomic eval NLL across counts, fixed eval IDs, nested
+select counts, independent group accounting, overlap rejection, and analytical
+gradients of the mean-reference objective. The multi-reference path completed on Modal A100-80GB (lhtu05), app
+ap-uR42YdE7fVsJg3tbfZYE5h, source ee28346: four synthetic groups, base
+Gemma-2-2b-it / Phi-4-mini, FP32, virtual LR 0.1. Company SFT/Qwen remains
+unexecuted. See MP_OPD_OVERNIGHT_20260910.md for results and limitations.

simct-b200-portable.bundle

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6aafb042e138f63cf5d558a9d9bda737665330f194b9f3e894c12b45949faa2b
-size 4163162
+oid sha256:4872eae417f41ad55a8a46e03aba975d9362f9023b977825cce3cf86030cf6e7
+size 4184862

source-manifest.json

@@ -1,9 +1,9 @@
 {
   "source_branch": "vdt/ops/b200-portable",
-  "source_commit": "a7ed9610c2de02bab040d30136d2bfe8f574f8e7",
+  "source_commit": "01c7bf4d591271643f9ab9397cc0b4916059ee31",
   "bundle_file": "simct-b200-portable.bundle",
-  "bundle_sha256": "6aafb042e138f63cf5d558a9d9bda737665330f194b9f3e894c12b45949faa2b",
-  "bundle_bytes": 4163162,
+  "bundle_sha256": "4872eae417f41ad55a8a46e03aba975d9362f9023b977825cce3cf86030cf6e7",
+  "bundle_bytes": 4184862,
   "history": "complete",
   "format": "git-bundle",
   "reason": "HF Git receive hook rejects historical PNG/JPG blobs; bundle preserves original Git object IDs",
@@ -21,5 +21,11 @@
   "eval_queue": {
     "pull-eval-queue.sh": "d20d18a6b22bae31313f13b26252f7032a290e4b229af842ddfe3a0d77ad8aa8",
     "EVAL-QUEUE-RUNBOOK.md": "e249ed6750fb4264496fd6ec7bf2a5dcc67ccfc3be4e7f85be05cb00d8446f2b"
+  },
+  "multiref_probe": {
+    "runbook": "MP-OPD-MULTIREF.md",
+    "sha256": "ae03b4e845d4dfe2dbb060a7015529d99cc837248926373da71d65441aef40cf",
+    "gpu_canary_source": "ee28346",
+    "company_gpu_run": false
   }
 }