From 56ca19142a9c34d5bc4018b677c1eb8a645ab7ff Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 04:12:41 +1200
Subject: [PATCH 01/20] infra: microbatch frozen V121 embeddings to fit runner
 memory

---
 .../trace_the_ace/v121_pretrained_semantic_residual.py      | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/competitions/trace_the_ace/v121_pretrained_semantic_residual.py b/competitions/trace_the_ace/v121_pretrained_semantic_residual.py
index 5a51b41d..323134d6 100644
--- a/competitions/trace_the_ace/v121_pretrained_semantic_residual.py
+++ b/competitions/trace_the_ace/v121_pretrained_semantic_residual.py
@@ -62,7 +62,11 @@ def build_semantic_text(objective: str, transcript_df) -> str:
 
 
 def embed(model: TextEmbedding, seq: list[str]) -> np.ndarray:
-    arr = np.vstack(list(model.embed(seq, batch_size=64))).astype(np.float32)
+    # Infrastructure-only repair after the first valid V121 execution reached
+    # ONNX and batch_size=64 requested a ~66.5 GB attention buffer.  The model,
+    # texts, ordering and outputs are unchanged; only inference microbatching is
+    # reduced to fit the hosted runner.
+    arr = np.vstack(list(model.embed(seq, batch_size=2))).astype(np.float32)
     if not np.isfinite(arr).all():
         raise RuntimeError("non-finite embedding")
     return arr

From b7c181a039a7207d6744b6641f1517797e3b2612 Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 04:12:59 +1200
Subject: [PATCH 02/20] infra: isolate V121 memory-safe rerun

---
 .../workflows/trace-ace-v121-pretrained-semantic-residual.yml   | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
index a3fa5201..8b027003 100644
--- a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
+++ b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
@@ -2,7 +2,7 @@ name: Trace Ace V121 Pretrained Semantic Residual
 
 on:
   pull_request:
-    branches: [agent/trace-ace-mastery-events]
+    branches: [agent/trace-ace-mastery-events, agent/v111-runner]
     paths:
       - 'competitions/trace_the_ace/v121_pretrained_semantic_residual.py'
       - '.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml'

From b26d58dc67b1180fb9121ad23cd761523538ea3f Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 04:33:08 +1200
Subject: [PATCH 03/20] infra: finish frozen V121 within hosted runner

---
 .../v121_pretrained_semantic_residual.py              | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/competitions/trace_the_ace/v121_pretrained_semantic_residual.py b/competitions/trace_the_ace/v121_pretrained_semantic_residual.py
index 323134d6..b537d627 100644
--- a/competitions/trace_the_ace/v121_pretrained_semantic_residual.py
+++ b/competitions/trace_the_ace/v121_pretrained_semantic_residual.py
@@ -62,11 +62,12 @@ def build_semantic_text(objective: str, transcript_df) -> str:
 
 
 def embed(model: TextEmbedding, seq: list[str]) -> np.ndarray:
-    # Infrastructure-only repair after the first valid V121 execution reached
-    # ONNX and batch_size=64 requested a ~66.5 GB attention buffer.  The model,
-    # texts, ordering and outputs are unchanged; only inference microbatching is
-    # reduced to fit the hosted runner.
-    arr = np.vstack(list(model.embed(seq, batch_size=2))).astype(np.float32)
+    # Infrastructure-only repair. batch_size=64 requested a ~66.5 GB attention
+    # buffer; batch_size=2 fit memory but twice hit the hosted-runner wall-clock
+    # shutdown during the same frozen embedding. batch_size=4 preserves model,
+    # texts, ordering and outputs while fitting the measured memory envelope and
+    # reducing runtime enough to finish on the hosted runner.
+    arr = np.vstack(list(model.embed(seq, batch_size=4))).astype(np.float32)
     if not np.isfinite(arr).all():
         raise RuntimeError("non-finite embedding")
     return arr

From 16407d42992145284c9c8efc10fd72872d6467be Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 04:44:08 +1200
Subject: [PATCH 04/20] infra: stage frozen V121 across hosted runners

---
 .../trace_the_ace/v121_staged_transport.py    | 121 ++++++++++++++++++
 1 file changed, 121 insertions(+)
 create mode 100644 competitions/trace_the_ace/v121_staged_transport.py

diff --git a/competitions/trace_the_ace/v121_staged_transport.py b/competitions/trace_the_ace/v121_staged_transport.py
new file mode 100644
index 00000000..19e90910
--- /dev/null
+++ b/competitions/trace_the_ace/v121_staged_transport.py
@@ -0,0 +1,121 @@
+#!/usr/bin/env python3
+"""Infrastructure-only staged transport for frozen V121.
+
+This module does not define new scientific features, models, folds, controls, or
+gates. It serializes the exact V121 computation across short-lived hosted runner
+jobs so preparation, embedding, and evaluation can complete independently.
+"""
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+import numpy as np
+from fastembed import TextEmbedding
+from scipy.sparse import load_npz, save_npz
+
+from v110_residual_collider_state_discovery import hb
+from v121_pretrained_semantic_residual import (
+    MODEL_NAME,
+    build_semantic_text,
+    embed,
+    eval_geometry,
+    stable_hash,
+    within_objective_shuffle,
+)
+from v71_mastery_events import load_transcript
+from v75_canonical_trajectory import load_training
+from v85_evidence_state import build_v75
+from v94_related_control import segmented_control, build_control
+
+
+def prepare(a):
+    out = Path(a.dir); out.mkdir(parents=True, exist_ok=True)
+    f = load_training(a.features, a.labels).reset_index(drop=True)
+    print('features columns', list(f.columns), flush=True)
+    obj0 = (f.learning_objective_id if 'learning_objective_id' in f else f.learning_objective).astype(str).to_numpy()
+    cand = np.where(np.array([hb(x, 5) != 0 for x in obj0]))[0]
+    ix = np.array(sorted(cand, key=lambda i: stable_hash(f.response_id.iloc[i]))[:a.rows])
+    f = f.iloc[ix].reset_index(drop=True)
+
+    y = f.target.to_numpy(int)
+    objectives = (f.learning_objective_id if 'learning_objective_id' in f else f.learning_objective).astype(str).to_numpy()
+    support = f.learning_objective.astype(str).to_numpy()
+    sessions = f.session_id.astype(str).to_numpy()
+
+    cache = {s: load_transcript(a.transcripts / f'{s}.csv') for s in np.unique(sessions)}
+    rt, rz, sem_text, obj_text = [], [], [], []
+    for i, r in f.iterrows():
+        d = cache[str(r.session_id)]
+        t, z = segmented_control(d, str(r.learning_objective), 'related')
+        rt.append(t); rz.append(z)
+        obj_text.append(f'learning objective: {r.learning_objective}')
+        sem_text.append(build_semantic_text(str(r.learning_objective), d))
+        if (i + 1) % 500 == 0:
+            print('prepared rows', i + 1, flush=True)
+
+    save_npz(out / 'X75.npz', build_v75(f, cache))
+    save_npz(out / 'Xr.npz', build_control(rt, rz))
+    np.savez_compressed(out / 'arrays.npz', y=y, objectives=objectives, support=support, sessions=sessions)
+    (out / 'texts.json').write_text(json.dumps({'objective': obj_text, 'semantic': sem_text}))
+    manifest = {
+        'protocol': 'V121_PRETRAINED_SEMANTIC_RESIDUAL', 'rows': int(len(f)),
+        'objectives': int(len(np.unique(objectives))), 'sessions': int(len(np.unique(sessions))),
+        'response_ids_sha256': __import__('hashlib').sha256('\n'.join(f.response_id.astype(str)).encode()).hexdigest(),
+    }
+    (out / 'manifest.json').write_text(json.dumps(manifest, indent=2))
+    print(json.dumps(manifest, indent=2), flush=True)
+
+
+def do_embed(a):
+    d = Path(a.dir)
+    texts = json.loads((d / 'texts.json').read_text())
+    print('loading embedding model', MODEL_NAME, flush=True)
+    model = TextEmbedding(model_name=MODEL_NAME)
+    print('embedding objective control', flush=True)
+    E_obj = embed(model, texts['objective'])
+    print('embedding semantic intervention', flush=True)
+    E_sem = embed(model, texts['semantic'])
+    if E_obj.shape[0] != E_sem.shape[0]: raise RuntimeError('embedding row mismatch')
+    np.savez_compressed(Path(a.out), E_obj=E_obj, E_sem=E_sem)
+    print('embedding shapes', E_obj.shape, E_sem.shape, flush=True)
+
+
+def evaluate(a):
+    d = Path(a.dir)
+    X75 = load_npz(d / 'X75.npz'); Xr = load_npz(d / 'Xr.npz')
+    z = np.load(d / 'arrays.npz', allow_pickle=False)
+    y=z['y']; objectives=z['objectives']; support=z['support']; sessions=z['sessions']
+    e = np.load(a.embeddings, allow_pickle=False); E_obj=e['E_obj']; E_sem=e['E_sem']
+    E_shuf = within_objective_shuffle(E_sem, objectives)
+    manifest=json.loads((d/'manifest.json').read_text())
+    results = {
+        'protocol': 'V121_PRETRAINED_SEMANTIC_RESIDUAL', 'model': MODEL_NAME,
+        'rows': int(len(y)), 'objectives': int(len(np.unique(objectives))),
+        'sessions': int(len(np.unique(sessions))),
+        'transport_manifest': manifest,
+        'precommit': {'semantic_gain_each_geometry': .003,
+                      'semantic_minus_shuffle_each_geometry': .002,
+                      'hard_collision_gain_each_geometry': '>0',
+                      'no_hyperparameter_sweep': True},
+    }
+    results['objective_grouped'] = eval_geometry('objective_grouped', objectives, X75, Xr, y, support, objectives, E_obj, E_sem, E_shuf)
+    results['session_grouped'] = eval_geometry('session_grouped', sessions, X75, Xr, y, support, objectives, E_obj, E_sem, E_shuf)
+    def passes(r):
+        return r['semantic']['gain'] >= .003 and r['semantic_minus_shuffle_gain'] >= .002 and r['hard_collision'].get('semantic_gain', -1.) > 0
+    ok_obj=passes(results['objective_grouped']); ok_sess=passes(results['session_grouped'])
+    if ok_obj and ok_sess:
+        verdict='PHASE_CHANGE_CANDIDATE'; nxt='Promote pretrained semantic residual to larger frozen validation and public-probe packaging.'
+    else:
+        verdict='NO_ROBUST_SEMANTIC_PHASE_CHANGE'; nxt='Treat remaining oracle gap as largely unidentifiable from supplied transcript/objective observables; pivot to validation geometry / assessment-process inference rather than more text feature search.'
+    results['decision']={'objective_grouped_pass': bool(ok_obj), 'session_grouped_pass': bool(ok_sess), 'verdict': verdict, 'next': nxt}
+    Path(a.out).write_text(json.dumps(results, indent=2)); print(json.dumps(results, indent=2), flush=True)
+
+
+if __name__ == '__main__':
+    p=argparse.ArgumentParser(); sp=p.add_subparsers(dest='cmd', required=True)
+    q=sp.add_parser('prepare'); q.add_argument('--features',type=Path,required=True); q.add_argument('--labels',type=Path,required=True); q.add_argument('--transcripts',type=Path,required=True); q.add_argument('--rows',type=int,default=2500); q.add_argument('--dir',required=True)
+    q=sp.add_parser('embed'); q.add_argument('--dir',required=True); q.add_argument('--out',required=True)
+    q=sp.add_parser('evaluate'); q.add_argument('--dir',required=True); q.add_argument('--embeddings',required=True); q.add_argument('--out',required=True)
+    a=p.parse_args(); {'prepare':prepare,'embed':do_embed,'evaluate':evaluate}[a.cmd](a)

From 1c292f8c77a42e805a44f959e5f5f5266d6bb069 Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 04:44:29 +1200
Subject: [PATCH 05/20] infra: split frozen V121 across hosted runners

---
 ...-ace-v121-pretrained-semantic-residual.yml | 119 ++++++++++++------
 1 file changed, 78 insertions(+), 41 deletions(-)

diff --git a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
index 8b027003..f0e9158f 100644
--- a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
+++ b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
@@ -5,6 +5,7 @@ on:
     branches: [agent/trace-ace-mastery-events, agent/v111-runner]
     paths:
       - 'competitions/trace_the_ace/v121_pretrained_semantic_residual.py'
+      - 'competitions/trace_the_ace/v121_staged_transport.py'
       - '.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml'
   workflow_dispatch:
 
@@ -13,10 +14,14 @@ concurrency:
   group: trace-ace-frozen-transcripts-v1
   cancel-in-progress: false
 
+env:
+  TRANSCRIPT_KEY: trace-ace-transcripts-v1-603547640
+  TRANSCRIPT_SHA: e685b85b04694e130c25b17d09cdd1892fbda5e9fa685e98b2300114b915aa2d
+
 jobs:
-  semantic:
+  prepare:
     runs-on: ubuntu-24.04
-    timeout-minutes: 45
+    timeout-minutes: 20
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-python@v5
@@ -24,28 +29,21 @@ jobs:
           python-version: '3.12'
           cache: pip
       - name: Install dependencies
-        run: |
-          python -m pip install --disable-pip-version-check numpy pandas scipy scikit-learn gdown fastembed==0.8.0
-
-      - name: Restore frozen transcript archive from GitHub cache
+        run: python -m pip install --disable-pip-version-check numpy pandas scipy scikit-learn gdown fastembed==0.8.0
+      - name: Restore frozen transcript archive
         uses: actions/cache/restore@v4
         with:
           path: transcripts.zip
-          key: trace-ace-transcripts-v1-603547640
+          key: ${{ env.TRANSCRIPT_KEY }}
           fail-on-cache-miss: true
-
       - name: Validate frozen transcript archive
         shell: bash
         run: |
           set -euo pipefail
-          test -f transcripts.zip
           test "$(stat -c%s transcripts.zip)" = "603547640"
           unzip -tq transcripts.zip >/dev/null
-          ACTUAL=$(sha256sum transcripts.zip | cut -d' ' -f1)
-          test "$ACTUAL" = "e685b85b04694e130c25b17d09cdd1892fbda5e9fa685e98b2300114b915aa2d"
-          echo "Frozen transcript cache verified: $ACTUAL"
-
-      - name: Download frozen metadata and extract data
+          test "$(sha256sum transcripts.zip | cut -d' ' -f1)" = "$TRANSCRIPT_SHA"
+      - name: Download frozen metadata and extract
         shell: bash
         run: |
           set -euo pipefail
@@ -53,41 +51,80 @@ jobs:
           mkdir -p data/meta data/transcripts
           unzip -q metadata.zip -d data/meta
           unzip -q transcripts.zip -d data/transcripts
-          FEATURES=$(find data/meta -type f -name 'train_features*.csv' -print -quit)
-          LABELS=$(find data/meta -type f -name 'train_labels*.csv' -print -quit)
+          echo "FEATURES=$(find data/meta -type f -name 'train_features*.csv' -print -quit)" >> "$GITHUB_ENV"
+          echo "LABELS=$(find data/meta -type f -name 'train_labels*.csv' -print -quit)" >> "$GITHUB_ENV"
           FIRST=$(find data/transcripts -type f -name '*.csv' -print -quit)
-          test -n "$FEATURES"
-          test -n "$LABELS"
-          test -n "$FIRST"
-          TRANSCRIPTS=$(dirname "$FIRST")
-          echo "FEATURES=$FEATURES" >> "$GITHUB_ENV"
-          echo "LABELS=$LABELS" >> "$GITHUB_ENV"
-          echo "TRANSCRIPTS=$TRANSCRIPTS" >> "$GITHUB_ENV"
+          echo "TRANSCRIPTS=$(dirname "$FIRST")" >> "$GITHUB_ENV"
+      - name: Prepare exact frozen V121 inputs
+        run: |
+          python -m py_compile competitions/trace_the_ace/v121_pretrained_semantic_residual.py competitions/trace_the_ace/v121_staged_transport.py
+          cd competitions/trace_the_ace
+          python v121_staged_transport.py prepare \
+            --features "../../$FEATURES" --labels "../../$LABELS" \
+            --transcripts "../../$TRANSCRIPTS" --rows 2500 --dir ../../v121_prepared
+      - uses: actions/upload-artifact@v4
+        with:
+          name: v121-prepared
+          path: v121_prepared/
+          retention-days: 2
+          compression-level: 6
 
-      - name: Preflight
+  embed:
+    needs: prepare
+    runs-on: ubuntu-24.04
+    timeout-minutes: 20
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+          cache: pip
+      - name: Install dependencies
+        run: python -m pip install --disable-pip-version-check numpy pandas scipy scikit-learn fastembed==0.8.0
+      - uses: actions/download-artifact@v4
+        with:
+          name: v121-prepared
+          path: v121_prepared
+      - name: Embed exact frozen V121 texts
         run: |
-          python -m py_compile competitions/trace_the_ace/v121_pretrained_semantic_residual.py
-          python - <<'PY'
-          import os, pandas as pd
-          print('features columns', list(pd.read_csv(os.environ['FEATURES'], nrows=0).columns))
-          print('labels columns', list(pd.read_csv(os.environ['LABELS'], nrows=0).columns))
-          PY
-      - name: Run V121
+          cd competitions/trace_the_ace
+          python v121_staged_transport.py embed --dir ../../v121_prepared --out ../../v121_embeddings.npz
+      - uses: actions/upload-artifact@v4
+        with:
+          name: v121-embeddings
+          path: v121_embeddings.npz
+          retention-days: 2
+          compression-level: 0
+
+  evaluate:
+    needs: embed
+    runs-on: ubuntu-24.04
+    timeout-minutes: 20
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+          cache: pip
+      - name: Install dependencies
+        run: python -m pip install --disable-pip-version-check numpy pandas scipy scikit-learn fastembed==0.8.0
+      - uses: actions/download-artifact@v4
+        with:
+          name: v121-prepared
+          path: v121_prepared
+      - uses: actions/download-artifact@v4
+        with:
+          name: v121-embeddings
+          path: .
+      - name: Evaluate frozen V121 precommit
         run: |
           cd competitions/trace_the_ace
-          python v121_pretrained_semantic_residual.py \
-            --features "../../$FEATURES" \
-            --labels "../../$LABELS" \
-            --transcripts "../../$TRANSCRIPTS" \
-            --rows 2500 \
-            --out ../../v121_pretrained_semantic_residual.json
+          python v121_staged_transport.py evaluate --dir ../../v121_prepared \
+            --embeddings ../../v121_embeddings.npz --out ../../v121_pretrained_semantic_residual.json
       - name: Show decision
-        if: always()
-        run: test -f v121_pretrained_semantic_residual.json && cat v121_pretrained_semantic_residual.json || true
+        run: cat v121_pretrained_semantic_residual.json
       - uses: actions/upload-artifact@v4
-        if: always()
         with:
           name: trace-ace-v121-pretrained-semantic-residual
           path: v121_pretrained_semantic_residual.json
           retention-days: 14
-          if-no-files-found: warn

From 3263beb3cce7c18b7519a49e529a381cfe246b98 Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 04:47:32 +1200
Subject: [PATCH 06/20] experiment: freeze V125 nested calibration residual

---
 .../trace_the_ace/v125_nested_calibration.py  | 96 +++++++++++++++++++
 1 file changed, 96 insertions(+)
 create mode 100644 competitions/trace_the_ace/v125_nested_calibration.py

diff --git a/competitions/trace_the_ace/v125_nested_calibration.py b/competitions/trace_the_ace/v125_nested_calibration.py
new file mode 100644
index 00000000..b08d6008
--- /dev/null
+++ b/competitions/trace_the_ace/v125_nested_calibration.py
@@ -0,0 +1,96 @@
+#!/usr/bin/env python3
+"""V125: nested calibration residual over frozen V97.
+
+Question: is V97 leaving lawful log-loss improvement in probability calibration,
+without adding new information or exploiting a particular validation geometry?
+
+Frozen protocol:
+- deterministic 2500-row response-id sample;
+- exact V97 endpoint (V75 when objective supported; .65 V75 + .35 RELATED when unsupported);
+- 4-fold outer objective-grouped and session-grouped OOF;
+- calibration parameters fit only to inner-OOF V97 predictions inside each outer training fold;
+- intervention = one global Platt map sigmoid(a + b*logit(p97));
+- control = same map fit after deterministic shuffle of inner-OOF probabilities;
+- no hyperparameter sweep.
+
+Promote only if calibration gains >= .001 log loss in BOTH geometries and beats
+the shuffled calibration by >= .001 in BOTH. Otherwise retain as a negative law.
+"""
+from __future__ import annotations
+import argparse, hashlib, json
+from pathlib import Path
+import numpy as np
+from sklearn.linear_model import LogisticRegression
+from sklearn.metrics import log_loss
+from sklearn.model_selection import GroupKFold
+
+from v71_mastery_events import load_transcript
+from v75_canonical_trajectory import load_training, SEED
+from v85_evidence_state import build_v75
+from v94_related_control import segmented_control, build_control
+
+EPS=1e-5
+
+def hh(x): return int(hashlib.sha256(str(x).encode()).hexdigest()[:16],16)
+def ll(y,p): return float(log_loss(y,np.clip(p,EPS,1-EPS)))
+def logit(p):
+    p=np.clip(np.asarray(p,float),EPS,1-EPS); return np.log(p/(1-p))
+
+def endpoint(X75,Xr,y,key,tr,va):
+    m=LogisticRegression(C=.25,max_iter=300,solver='liblinear',random_state=SEED).fit(X75[tr],y[tr])
+    p75=m.predict_proba(X75[va])[:,1]
+    r=LogisticRegression(C=.25,max_iter=300,solver='liblinear',random_state=SEED).fit(Xr[tr],y[tr])
+    pr=r.predict_proba(Xr[va])[:,1]
+    vals,cts=np.unique(key[tr],return_counts=True); d=dict(zip(vals,cts))
+    seen=np.array([d.get(x,0)>0 for x in key[va]])
+    return np.clip(np.where(seen,p75,.65*p75+.35*pr),EPS,1-EPS)
+
+def fit_cal(p,y):
+    m=LogisticRegression(C=1000.,max_iter=300,solver='liblinear',random_state=SEED).fit(logit(p)[:,None],y)
+    return m
+
+def geometry(name,groups,X75,Xr,y,key):
+    outer=list(GroupKFold(4).split(np.zeros(len(y)),y,groups))
+    pb=np.zeros(len(y)); pc=np.zeros(len(y)); ps=np.zeros(len(y)); folds=[]
+    for k,(tr,va) in enumerate(outer):
+        inner_groups=groups[tr]
+        inn=list(GroupKFold(min(4,len(np.unique(inner_groups)))).split(np.zeros(len(tr)),y[tr],inner_groups))
+        pi=np.zeros(len(tr))
+        for itr,iva in inn:
+            pi[iva]=endpoint(X75,Xr,y,key,tr[itr],tr[iva])
+        cal=fit_cal(pi,y[tr])
+        rng=np.random.default_rng(SEED+125+k)
+        sh=fit_cal(pi[rng.permutation(len(pi))],y[tr])
+        raw=endpoint(X75,Xr,y,key,tr,va)
+        q=cal.predict_proba(logit(raw)[:,None])[:,1]
+        qs=sh.predict_proba(logit(raw)[:,None])[:,1]
+        pb[va]=raw;pc[va]=q;ps[va]=qs
+        folds.append({'fold':k+1,'rows':int(len(va)),'baseline':ll(y[va],raw),'calibrated':ll(y[va],q),
+                      'gain':ll(y[va],raw)-ll(y[va],q),'slope':float(cal.coef_[0,0]),
+                      'intercept':float(cal.intercept_[0])})
+    base=ll(y,pb); cal=ll(y,pc); shuf=ll(y,ps)
+    return {'geometry':name,'baseline_v97_ll':base,'calibrated_ll':cal,'gain':base-cal,
+            'shuffled_calibration_ll':shuf,'calibration_minus_shuffle_gain':shuf-cal,'folds':folds}
+
+def run(a):
+    f=load_training(a.features,a.labels).reset_index(drop=True)
+    print('features columns',list(f.columns),flush=True)
+    ix=sorted(range(len(f)),key=lambda i:hh(f.response_id.iloc[i]))[:a.rows]
+    f=f.iloc[ix].reset_index(drop=True)
+    y=f.target.to_numpy(int); key=f.learning_objective.astype(str).to_numpy()
+    obj=(f.learning_objective_id if 'learning_objective_id' in f else f.learning_objective).astype(str).to_numpy()
+    sess=f.session_id.astype(str).to_numpy()
+    cache={s:load_transcript(a.transcripts/f'{s}.csv') for s in np.unique(sess)}
+    rt=[];rz=[]
+    for i,r in f.iterrows():
+        t,z=segmented_control(cache[str(r.session_id)],str(r.learning_objective),'related');rt.append(t);rz.append(z)
+        if (i+1)%500==0: print('prepared rows',i+1,flush=True)
+    X75=build_v75(f,cache);Xr=build_control(rt,rz)
+    ro=geometry('objective_grouped',obj,X75,Xr,y,key);rs=geometry('session_grouped',sess,X75,Xr,y,key)
+    def ok(r): return r['gain']>=.001 and r['calibration_minus_shuffle_gain']>=.001
+    verdict='PROMOTE_CALIBRATION_LAW' if ok(ro) and ok(rs) else 'KEEP_V97_CALIBRATION'
+    out={'protocol':'V125_NESTED_CALIBRATION','rows':len(f),'precommit':{'gain_each_geometry':.001,'margin_vs_shuffle_each':.001,'no_sweep':True},
+         'objective_grouped':ro,'session_grouped':rs,'decision':{'objective_pass':ok(ro),'session_pass':ok(rs),'verdict':verdict}}
+    Path(a.out).write_text(json.dumps(out,indent=2));print(json.dumps(out,indent=2),flush=True)
+if __name__=='__main__':
+    p=argparse.ArgumentParser();p.add_argument('--features',type=Path,required=True);p.add_argument('--labels',type=Path,required=True);p.add_argument('--transcripts',type=Path,required=True);p.add_argument('--rows',type=int,default=2500);p.add_argument('--out',default='v125_nested_calibration.json');run(p.parse_args())

From b24f2a48a2c38223821a44eaf25c0e3badfff14f Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 05:12:37 +1200
Subject: [PATCH 07/20] infra: shard frozen V121 embedding transport

---
 .../trace_the_ace/v121_staged_transport.py    | 65 ++++++++++++++++---
 1 file changed, 55 insertions(+), 10 deletions(-)

diff --git a/competitions/trace_the_ace/v121_staged_transport.py b/competitions/trace_the_ace/v121_staged_transport.py
index 19e90910..0e5f1ee5 100644
--- a/competitions/trace_the_ace/v121_staged_transport.py
+++ b/competitions/trace_the_ace/v121_staged_transport.py
@@ -71,15 +71,57 @@ def prepare(a):
 def do_embed(a):
     d = Path(a.dir)
     texts = json.loads((d / 'texts.json').read_text())
-    print('loading embedding model', MODEL_NAME, flush=True)
+    n = len(texts['semantic'])
+    shards = int(a.shards)
+    shard = int(a.shard)
+    if shards < 1 or shard < 0 or shard >= shards:
+        raise ValueError(f'invalid shard {shard}/{shards}')
+    start = (n * shard) // shards
+    end = (n * (shard + 1)) // shards
+    obj_text = texts['objective'][start:end]
+    sem_text = texts['semantic'][start:end]
+    print('loading embedding model', MODEL_NAME, 'shard', shard, 'rows', start, end, flush=True)
     model = TextEmbedding(model_name=MODEL_NAME)
-    print('embedding objective control', flush=True)
-    E_obj = embed(model, texts['objective'])
-    print('embedding semantic intervention', flush=True)
-    E_sem = embed(model, texts['semantic'])
-    if E_obj.shape[0] != E_sem.shape[0]: raise RuntimeError('embedding row mismatch')
-    np.savez_compressed(Path(a.out), E_obj=E_obj, E_sem=E_sem)
-    print('embedding shapes', E_obj.shape, E_sem.shape, flush=True)
+    print('embedding objective control shard', shard, flush=True)
+    E_obj = embed(model, obj_text)
+    print('embedding semantic intervention shard', shard, flush=True)
+    E_sem = embed(model, sem_text)
+    if E_obj.shape[0] != E_sem.shape[0] or E_obj.shape[0] != end - start:
+        raise RuntimeError('embedding row mismatch')
+    np.savez_compressed(Path(a.out), E_obj=E_obj, E_sem=E_sem,
+                        start=np.array(start), end=np.array(end), total=np.array(n),
+                        shard=np.array(shard), shards=np.array(shards))
+    print('embedding shard complete', shard, start, end, E_obj.shape, E_sem.shape, flush=True)
+
+
+def load_embeddings(path: Path):
+    if path.is_file():
+        e = np.load(path, allow_pickle=False)
+        return e['E_obj'], e['E_sem']
+    files = sorted(path.glob('**/v121_embeddings_shard_*.npz'))
+    if not files:
+        files = sorted(path.glob('**/*.npz'))
+    parts = []
+    for f in files:
+        e = np.load(f, allow_pickle=False)
+        if 'start' not in e.files or 'end' not in e.files:
+            continue
+        parts.append((int(e['start']), int(e['end']), int(e['total']), e['E_obj'], e['E_sem'], f))
+    if not parts:
+        raise RuntimeError(f'no embedding shards found under {path}')
+    parts.sort(key=lambda x: x[0])
+    total = parts[0][2]
+    cursor = 0
+    obj, sem = [], []
+    for start, end, t, eo, es, f in parts:
+        if t != total or start != cursor or end - start != eo.shape[0] or eo.shape[0] != es.shape[0]:
+            raise RuntimeError(f'invalid embedding shard coverage at {f}: {start}:{end}, cursor={cursor}, total={t}')
+        obj.append(eo); sem.append(es); cursor = end
+    if cursor != total:
+        raise RuntimeError(f'incomplete embedding coverage: {cursor}/{total}')
+    E_obj = np.vstack(obj); E_sem = np.vstack(sem)
+    print('merged embedding shards', len(parts), E_obj.shape, E_sem.shape, flush=True)
+    return E_obj, E_sem
 
 
 def evaluate(a):
@@ -87,7 +129,9 @@ def evaluate(a):
     X75 = load_npz(d / 'X75.npz'); Xr = load_npz(d / 'Xr.npz')
     z = np.load(d / 'arrays.npz', allow_pickle=False)
     y=z['y']; objectives=z['objectives']; support=z['support']; sessions=z['sessions']
-    e = np.load(a.embeddings, allow_pickle=False); E_obj=e['E_obj']; E_sem=e['E_sem']
+    E_obj, E_sem = load_embeddings(Path(a.embeddings))
+    if len(y) != E_obj.shape[0] or len(y) != E_sem.shape[0]:
+        raise RuntimeError('evaluation embedding row mismatch')
     E_shuf = within_objective_shuffle(E_sem, objectives)
     manifest=json.loads((d/'manifest.json').read_text())
     results = {
@@ -95,6 +139,7 @@ def evaluate(a):
         'rows': int(len(y)), 'objectives': int(len(np.unique(objectives))),
         'sessions': int(len(np.unique(sessions))),
         'transport_manifest': manifest,
+        'transport': {'embedding_shards_merged': True},
         'precommit': {'semantic_gain_each_geometry': .003,
                       'semantic_minus_shuffle_each_geometry': .002,
                       'hard_collision_gain_each_geometry': '>0',
@@ -116,6 +161,6 @@ def passes(r):
 if __name__ == '__main__':
     p=argparse.ArgumentParser(); sp=p.add_subparsers(dest='cmd', required=True)
     q=sp.add_parser('prepare'); q.add_argument('--features',type=Path,required=True); q.add_argument('--labels',type=Path,required=True); q.add_argument('--transcripts',type=Path,required=True); q.add_argument('--rows',type=int,default=2500); q.add_argument('--dir',required=True)
-    q=sp.add_parser('embed'); q.add_argument('--dir',required=True); q.add_argument('--out',required=True)
+    q=sp.add_parser('embed'); q.add_argument('--dir',required=True); q.add_argument('--out',required=True); q.add_argument('--shard',type=int,default=0); q.add_argument('--shards',type=int,default=1)
     q=sp.add_parser('evaluate'); q.add_argument('--dir',required=True); q.add_argument('--embeddings',required=True); q.add_argument('--out',required=True)
     a=p.parse_args(); {'prepare':prepare,'embed':do_embed,'evaluate':evaluate}[a.cmd](a)

From e62133e2cf5f83ffd145eeb40f3d35630a501d6a Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 05:12:53 +1200
Subject: [PATCH 08/20] infra: shard frozen V121 embedding jobs

---
 ...-ace-v121-pretrained-semantic-residual.yml | 24 ++++++++++++-------
 1 file changed, 16 insertions(+), 8 deletions(-)

diff --git a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
index f0e9158f..f9c5b772 100644
--- a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
+++ b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
@@ -71,6 +71,10 @@ jobs:
 
   embed:
     needs: prepare
+    strategy:
+      fail-fast: false
+      matrix:
+        shard: [0, 1, 2, 3]
     runs-on: ubuntu-24.04
     timeout-minutes: 20
     steps:
@@ -85,14 +89,16 @@ jobs:
         with:
           name: v121-prepared
           path: v121_prepared
-      - name: Embed exact frozen V121 texts
+      - name: Embed exact frozen V121 texts shard ${{ matrix.shard }}
         run: |
           cd competitions/trace_the_ace
-          python v121_staged_transport.py embed --dir ../../v121_prepared --out ../../v121_embeddings.npz
+          python v121_staged_transport.py embed --dir ../../v121_prepared \
+            --shard ${{ matrix.shard }} --shards 4 \
+            --out ../../v121_embeddings_shard_${{ matrix.shard }}.npz
       - uses: actions/upload-artifact@v4
         with:
-          name: v121-embeddings
-          path: v121_embeddings.npz
+          name: v121-embeddings-${{ matrix.shard }}
+          path: v121_embeddings_shard_${{ matrix.shard }}.npz
           retention-days: 2
           compression-level: 0
 
@@ -112,15 +118,17 @@ jobs:
         with:
           name: v121-prepared
           path: v121_prepared
-      - uses: actions/download-artifact@v4
+      - name: Download all exact embedding shards
+        uses: actions/download-artifact@v4
         with:
-          name: v121-embeddings
-          path: .
+          pattern: v121-embeddings-*
+          path: v121_embedding_shards
+          merge-multiple: false
       - name: Evaluate frozen V121 precommit
         run: |
           cd competitions/trace_the_ace
           python v121_staged_transport.py evaluate --dir ../../v121_prepared \
-            --embeddings ../../v121_embeddings.npz --out ../../v121_pretrained_semantic_residual.json
+            --embeddings ../../v121_embedding_shards --out ../../v121_pretrained_semantic_residual.json
       - name: Show decision
         run: cat v121_pretrained_semantic_residual.json
       - uses: actions/upload-artifact@v4

From c299c3f34fb9bbbcddf7895b5f4123b69699aadd Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 05:13:56 +1200
Subject: [PATCH 09/20] infra: cancel obsolete V121 transports

---
 .../workflows/trace-ace-v121-pretrained-semantic-residual.yml   | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
index f9c5b772..9f38f204 100644
--- a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
+++ b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
@@ -12,7 +12,7 @@ on:
 # Infrastructure-only serialization. Scientific protocol is unchanged.
 concurrency:
   group: trace-ace-frozen-transcripts-v1
-  cancel-in-progress: false
+  cancel-in-progress: true
 
 env:
   TRANSCRIPT_KEY: trace-ace-transcripts-v1-603547640

From c4da5dda0e2d187e87493b0cb63a5e31cf36525e Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 05:22:26 +1200
Subject: [PATCH 10/20] infra: split frozen V121 into eight embedding shards

---
 .../workflows/trace-ace-v121-pretrained-semantic-residual.yml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
index 9f38f204..91478a62 100644
--- a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
+++ b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
@@ -74,7 +74,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        shard: [0, 1, 2, 3]
+        shard: [0, 1, 2, 3, 4, 5, 6, 7]
     runs-on: ubuntu-24.04
     timeout-minutes: 20
     steps:
@@ -93,7 +93,7 @@ jobs:
         run: |
           cd competitions/trace_the_ace
           python v121_staged_transport.py embed --dir ../../v121_prepared \
-            --shard ${{ matrix.shard }} --shards 4 \
+            --shard ${{ matrix.shard }} --shards 8 \
             --out ../../v121_embeddings_shard_${{ matrix.shard }}.npz
       - uses: actions/upload-artifact@v4
         with:

From f34b01f02bd83c20b1f9cbcb6f38ad8479688f5d Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 05:32:18 +1200
Subject: [PATCH 11/20] infra: serialize V121 embedding shards

---
 .../workflows/trace-ace-v121-pretrained-semantic-residual.yml    | 1 +
 1 file changed, 1 insertion(+)

diff --git a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
index 91478a62..8c78f74a 100644
--- a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
+++ b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
@@ -73,6 +73,7 @@ jobs:
     needs: prepare
     strategy:
       fail-fast: false
+      max-parallel: 1
       matrix:
         shard: [0, 1, 2, 3, 4, 5, 6, 7]
     runs-on: ubuntu-24.04

From 7d6759bf397ce91b924ba14a5fc50bdaaf8a9c23 Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 05:54:44 +1200
Subject: [PATCH 12/20] infra: fit frozen V121 shards within runner lifetime

---
 .../trace-ace-v121-pretrained-semantic-residual.yml         | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
index 8c78f74a..ce72bbf9 100644
--- a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
+++ b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
@@ -12,7 +12,7 @@ on:
 # Infrastructure-only serialization. Scientific protocol is unchanged.
 concurrency:
   group: trace-ace-frozen-transcripts-v1
-  cancel-in-progress: true
+  cancel-in-progress: false
 
 env:
   TRANSCRIPT_KEY: trace-ace-transcripts-v1-603547640
@@ -75,7 +75,7 @@ jobs:
       fail-fast: false
       max-parallel: 1
       matrix:
-        shard: [0, 1, 2, 3, 4, 5, 6, 7]
+        shard: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]
     runs-on: ubuntu-24.04
     timeout-minutes: 20
     steps:
@@ -94,7 +94,7 @@ jobs:
         run: |
           cd competitions/trace_the_ace
           python v121_staged_transport.py embed --dir ../../v121_prepared \
-            --shard ${{ matrix.shard }} --shards 8 \
+            --shard ${{ matrix.shard }} --shards 16 \
             --out ../../v121_embeddings_shard_${{ matrix.shard }}.npz
       - uses: actions/upload-artifact@v4
         with:

From d398685f939e800ccda490743c04be83f1f65c27 Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 05:55:20 +1200
Subject: [PATCH 13/20] infra: isolate frozen V121 16-shard execution

---
 .../workflows/trace-ace-v121-pretrained-semantic-residual.yml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
index ce72bbf9..435f9afd 100644
--- a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
+++ b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
@@ -9,9 +9,9 @@ on:
       - '.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml'
   workflow_dispatch:
 
-# Infrastructure-only serialization. Scientific protocol is unchanged.
+# Infrastructure-only serialization for the frozen 16-shard V121 transport.
 concurrency:
-  group: trace-ace-frozen-transcripts-v1
+  group: trace-ace-v121-frozen-16shard
   cancel-in-progress: false
 
 env:

From 4e85c48246f6373d0fcb6d12b6485e60bce3183e Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 06:01:15 +1200
Subject: [PATCH 14/20] infra: parallelize frozen V121 short shards

---
 .../workflows/trace-ace-v121-pretrained-semantic-residual.yml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
index 435f9afd..3e43b290 100644
--- a/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
+++ b/.github/workflows/trace-ace-v121-pretrained-semantic-residual.yml
@@ -11,7 +11,7 @@ on:
 
 # Infrastructure-only serialization for the frozen 16-shard V121 transport.
 concurrency:
-  group: trace-ace-v121-frozen-16shard
+  group: trace-ace-v121-frozen-16shard-p4
   cancel-in-progress: false
 
 env:
@@ -73,7 +73,7 @@ jobs:
     needs: prepare
     strategy:
       fail-fast: false
-      max-parallel: 1
+      max-parallel: 4
       matrix:
         shard: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]
     runs-on: ubuntu-24.04

From 5b3cc70b720a2b3118d07a259599dce6d1c49e56 Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 06:06:09 +1200
Subject: [PATCH 15/20] infra: run frozen V121 from prepared artifact

---
 .../workflows/trace-ace-v121-embed-only.yml   | 104 ++++++++++++++++++
 1 file changed, 104 insertions(+)
 create mode 100644 .github/workflows/trace-ace-v121-embed-only.yml

diff --git a/.github/workflows/trace-ace-v121-embed-only.yml b/.github/workflows/trace-ace-v121-embed-only.yml
new file mode 100644
index 00000000..84cc9f75
--- /dev/null
+++ b/.github/workflows/trace-ace-v121-embed-only.yml
@@ -0,0 +1,104 @@
+name: Trace Ace V121 Frozen Embed Only
+
+on:
+  pull_request:
+    branches: [agent/trace-ace-mastery-events, agent/v111-runner]
+    paths:
+      - '.github/workflows/trace-ace-v121-embed-only.yml'
+  workflow_dispatch:
+
+# Reuses the already-successful exact frozen V121 prepared artifact from run 32400309220.
+# Scientific model/text/sample/folds/gates are unchanged.
+concurrency:
+  group: trace-ace-v121-embed-only-16-p4
+  cancel-in-progress: false
+
+env:
+  PREPARED_RUN_ID: '32400309220'
+
+jobs:
+  embed:
+    strategy:
+      fail-fast: false
+      max-parallel: 4
+      matrix:
+        shard: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]
+    runs-on: ubuntu-24.04
+    timeout-minutes: 20
+    permissions:
+      actions: read
+      contents: read
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+          cache: pip
+      - name: Install dependencies
+        run: python -m pip install --disable-pip-version-check numpy pandas scipy scikit-learn fastembed==0.8.0
+      - name: Download exact frozen V121 prepared artifact
+        uses: actions/download-artifact@v4
+        with:
+          name: v121-prepared
+          path: v121_prepared
+          repository: heathsanchez/mathgraph
+          run-id: ${{ env.PREPARED_RUN_ID }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+      - name: Verify frozen preparation manifest exists
+        run: |
+          test -f v121_prepared/manifest.json
+          cat v121_prepared/manifest.json
+      - name: Embed exact frozen V121 texts shard ${{ matrix.shard }}
+        run: |
+          cd competitions/trace_the_ace
+          python v121_staged_transport.py embed --dir ../../v121_prepared \
+            --shard ${{ matrix.shard }} --shards 16 \
+            --out ../../v121_embeddings_shard_${{ matrix.shard }}.npz
+      - uses: actions/upload-artifact@v4
+        with:
+          name: v121-embeddings-${{ matrix.shard }}
+          path: v121_embeddings_shard_${{ matrix.shard }}.npz
+          retention-days: 2
+          compression-level: 0
+
+  evaluate:
+    needs: embed
+    runs-on: ubuntu-24.04
+    timeout-minutes: 20
+    permissions:
+      actions: read
+      contents: read
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+          cache: pip
+      - name: Install dependencies
+        run: python -m pip install --disable-pip-version-check numpy pandas scipy scikit-learn fastembed==0.8.0
+      - name: Download exact frozen V121 prepared artifact
+        uses: actions/download-artifact@v4
+        with:
+          name: v121-prepared
+          path: v121_prepared
+          repository: heathsanchez/mathgraph
+          run-id: ${{ env.PREPARED_RUN_ID }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+      - name: Download all exact embedding shards
+        uses: actions/download-artifact@v4
+        with:
+          pattern: v121-embeddings-*
+          path: v121_embedding_shards
+          merge-multiple: false
+      - name: Evaluate frozen V121 precommit
+        run: |
+          cd competitions/trace_the_ace
+          python v121_staged_transport.py evaluate --dir ../../v121_prepared \
+            --embeddings ../../v121_embedding_shards --out ../../v121_pretrained_semantic_residual.json
+      - name: Show decision
+        run: cat v121_pretrained_semantic_residual.json
+      - uses: actions/upload-artifact@v4
+        with:
+          name: trace-ace-v121-pretrained-semantic-residual
+          path: v121_pretrained_semantic_residual.json
+          retention-days: 14

From 997d4052b4e1b6844d85cba44135e44ecca28e2f Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 06:11:09 +1200
Subject: [PATCH 16/20] infra: bootstrap local frozen V121 embedding

---
 .../trace-ace-v121-local-bootstrap.yml        | 49 +++++++++++++++++++
 1 file changed, 49 insertions(+)
 create mode 100644 .github/workflows/trace-ace-v121-local-bootstrap.yml

diff --git a/.github/workflows/trace-ace-v121-local-bootstrap.yml b/.github/workflows/trace-ace-v121-local-bootstrap.yml
new file mode 100644
index 00000000..f18b81de
--- /dev/null
+++ b/.github/workflows/trace-ace-v121-local-bootstrap.yml
@@ -0,0 +1,49 @@
+name: Trace Ace V121 Local Bootstrap
+
+on:
+  pull_request:
+    branches: [agent/trace-ace-mastery-events, agent/v111-runner]
+    paths:
+      - '.github/workflows/trace-ace-v121-local-bootstrap.yml'
+  workflow_dispatch:
+
+jobs:
+  bootstrap:
+    runs-on: ubuntu-24.04
+    timeout-minutes: 10
+    permissions:
+      contents: read
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.13'
+      - name: Download Python 3.13 wheelhouse
+        run: |
+          mkdir -p wheelhouse
+          python -m pip download --only-binary=:all: --dest wheelhouse fastembed==0.8.0
+      - name: Install FastEmbed and fetch exact Jina model
+        run: |
+          python -m pip install --no-index --find-links wheelhouse fastembed==0.8.0
+          python - <<'PY'
+          from fastembed import TextEmbedding
+          m = TextEmbedding(model_name='jinaai/jina-embeddings-v2-small-en')
+          print('model_ready', type(m).__name__)
+          PY
+          mkdir -p model_cache
+          for p in "$HOME/.cache/fastembed" "$HOME/.cache/huggingface"; do
+            if [ -d "$p" ]; then cp -a "$p" model_cache/; fi
+          done
+          find model_cache -maxdepth 4 -type f -printf '%p %s\n' | head -100
+      - uses: actions/upload-artifact@v4
+        with:
+          name: v121-python313-wheelhouse
+          path: wheelhouse/
+          retention-days: 2
+          compression-level: 0
+      - uses: actions/upload-artifact@v4
+        with:
+          name: v121-jina-model-cache
+          path: model_cache/
+          retention-days: 2
+          compression-level: 0

From 4f08ddc0af044838e63ac1004e9289ea6b6fb9bf Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 06:13:53 +1200
Subject: [PATCH 17/20] infra: export exact frozen V121 Jina model cache

---
 .../workflows/trace-ace-v121-model-export.yml | 41 +++++++++++++++++++
 1 file changed, 41 insertions(+)
 create mode 100644 .github/workflows/trace-ace-v121-model-export.yml

diff --git a/.github/workflows/trace-ace-v121-model-export.yml b/.github/workflows/trace-ace-v121-model-export.yml
new file mode 100644
index 00000000..b8c0e63c
--- /dev/null
+++ b/.github/workflows/trace-ace-v121-model-export.yml
@@ -0,0 +1,41 @@
+name: Trace Ace V121 Model Export
+
+on:
+  pull_request:
+    branches: [agent/trace-ace-mastery-events, agent/v111-runner]
+    paths:
+      - '.github/workflows/trace-ace-v121-model-export.yml'
+  workflow_dispatch:
+
+jobs:
+  export:
+    runs-on: ubuntu-24.04
+    timeout-minutes: 10
+    permissions:
+      contents: read
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.13'
+      - name: Install exact FastEmbed
+        run: python -m pip install --disable-pip-version-check fastembed==0.8.0
+      - name: Fetch and locate exact Jina model
+        run: |
+          python - <<'PY'
+          from fastembed import TextEmbedding
+          import tempfile, pathlib, json
+          cache = pathlib.Path('/tmp/v121_fastembed_cache')
+          m = TextEmbedding(model_name='jinaai/jina-embeddings-v2-small-en', cache_dir=str(cache))
+          print('MODEL_DICT', json.dumps({k:str(v) for k,v in m.__dict__.items()}, default=str, indent=2))
+          print('CACHE', cache)
+          PY
+          echo 'MODEL FILES:'
+          find /tmp/v121_fastembed_cache -type f -printf '%p %s\n' | sort
+          du -sh /tmp/v121_fastembed_cache
+      - uses: actions/upload-artifact@v4
+        with:
+          name: v121-jina-fastembed-cache
+          path: /tmp/v121_fastembed_cache/
+          retention-days: 2
+          compression-level: 0

From 4d2da1458c95cd2e06dcaa5add30730541c58280 Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 06:19:52 +1200
Subject: [PATCH 18/20] infra: add batch1 frozen V121 shard transport

---
 .../v121_embed_batch1_transport.py            | 38 +++++++++++++++++++
 1 file changed, 38 insertions(+)
 create mode 100644 competitions/trace_the_ace/v121_embed_batch1_transport.py

diff --git a/competitions/trace_the_ace/v121_embed_batch1_transport.py b/competitions/trace_the_ace/v121_embed_batch1_transport.py
new file mode 100644
index 00000000..08f96175
--- /dev/null
+++ b/competitions/trace_the_ace/v121_embed_batch1_transport.py
@@ -0,0 +1,38 @@
+#!/usr/bin/env python3
+"""Infrastructure-only memory-safe embedding transport for frozen V121.
+
+No scientific representation, model, sample, ordering, folds, controls, or gates
+are changed. Each text is embedded independently with the same frozen Jina model;
+batch_size=1 only lowers peak ONNX attention memory.
+"""
+from __future__ import annotations
+import argparse, json
+from pathlib import Path
+import numpy as np
+from fastembed import TextEmbedding
+
+MODEL_NAME = "jinaai/jina-embeddings-v2-small-en"
+
+def embed1(model, seq):
+    arr = np.vstack(list(model.embed(seq, batch_size=1))).astype(np.float32)
+    if not np.isfinite(arr).all():
+        raise RuntimeError("non-finite embedding")
+    return arr
+
+def main(a):
+    d=Path(a.dir)
+    texts=json.loads((d/'texts.json').read_text())
+    n=len(texts['semantic']); shard=int(a.shard); shards=int(a.shards)
+    if shards < 1 or not (0 <= shard < shards): raise ValueError(f'invalid shard {shard}/{shards}')
+    start=(n*shard)//shards; end=(n*(shard+1))//shards
+    obj=texts['objective'][start:end]; sem=texts['semantic'][start:end]
+    print('frozen shard',shard,'of',shards,'rows',start,end,flush=True)
+    model=TextEmbedding(model_name=MODEL_NAME, threads=4)
+    E_obj=embed1(model,obj)
+    E_sem=embed1(model,sem)
+    if E_obj.shape[0] != end-start or E_sem.shape[0] != end-start: raise RuntimeError('row mismatch')
+    np.savez_compressed(Path(a.out),E_obj=E_obj,E_sem=E_sem,start=np.array(start),end=np.array(end),total=np.array(n),shard=np.array(shard),shards=np.array(shards))
+    print('complete',shard,E_obj.shape,E_sem.shape,flush=True)
+
+if __name__=='__main__':
+    p=argparse.ArgumentParser(); p.add_argument('--dir',required=True); p.add_argument('--out',required=True); p.add_argument('--shard',type=int,required=True); p.add_argument('--shards',type=int,required=True); main(p.parse_args())

From 57a25a3f76070407c996df0c77ad7fc2004c2e73 Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 06:20:45 +1200
Subject: [PATCH 19/20] infra: execute frozen V121 in 128 batch1 shards

---
 .github/workflows/trace-ace-v121-128shard.yml | 102 ++++++++++++++++++
 1 file changed, 102 insertions(+)
 create mode 100644 .github/workflows/trace-ace-v121-128shard.yml

diff --git a/.github/workflows/trace-ace-v121-128shard.yml b/.github/workflows/trace-ace-v121-128shard.yml
new file mode 100644
index 00000000..527dadd6
--- /dev/null
+++ b/.github/workflows/trace-ace-v121-128shard.yml
@@ -0,0 +1,102 @@
+name: Trace Ace V121 Frozen 128 Shard
+
+on:
+  pull_request:
+    branches: [agent/trace-ace-mastery-events, agent/v111-runner]
+    paths:
+      - '.github/workflows/trace-ace-v121-128shard.yml'
+  workflow_dispatch:
+
+concurrency:
+  group: trace-ace-v121-frozen-128shard
+  cancel-in-progress: false
+
+env:
+  PREPARED_RUN_ID: '32400309220'
+
+jobs:
+  embed:
+    strategy:
+      fail-fast: false
+      max-parallel: 16
+      matrix:
+        shard: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127]
+    runs-on: ubuntu-24.04
+    timeout-minutes: 10
+    permissions:
+      actions: read
+      contents: read
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+          cache: pip
+      - name: Install exact embedding dependencies
+        run: python -m pip install --disable-pip-version-check numpy fastembed==0.8.0
+      - name: Download exact frozen V121 prepared artifact
+        uses: actions/download-artifact@v4
+        with:
+          name: v121-prepared
+          path: v121_prepared
+          repository: heathsanchez/mathgraph
+          run-id: ${{ env.PREPARED_RUN_ID }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+      - name: Verify frozen manifest
+        run: |
+          test -f v121_prepared/manifest.json
+          grep -q 'b1612f9fe4558680e468afb2a2452b75c603c244934fe62f7345feee68a61bc1' v121_prepared/manifest.json
+      - name: Embed exact frozen shard ${{ matrix.shard }}
+        run: |
+          cd competitions/trace_the_ace
+          python v121_embed_batch1_transport.py --dir ../../v121_prepared \
+            --shard ${{ matrix.shard }} --shards 128 \
+            --out ../../v121_embeddings_shard_${{ matrix.shard }}.npz
+      - uses: actions/upload-artifact@v4
+        with:
+          name: v121-embeddings-${{ matrix.shard }}
+          path: v121_embeddings_shard_${{ matrix.shard }}.npz
+          retention-days: 2
+          compression-level: 0
+
+  evaluate:
+    needs: embed
+    runs-on: ubuntu-24.04
+    timeout-minutes: 20
+    permissions:
+      actions: read
+      contents: read
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+          cache: pip
+      - name: Install evaluation dependencies
+        run: python -m pip install --disable-pip-version-check numpy pandas scipy scikit-learn fastembed==0.8.0
+      - name: Download exact frozen V121 prepared artifact
+        uses: actions/download-artifact@v4
+        with:
+          name: v121-prepared
+          path: v121_prepared
+          repository: heathsanchez/mathgraph
+          run-id: ${{ env.PREPARED_RUN_ID }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+      - name: Download all exact embedding shards
+        uses: actions/download-artifact@v4
+        with:
+          pattern: v121-embeddings-*
+          path: v121_embedding_shards
+          merge-multiple: false
+      - name: Evaluate unchanged frozen V121 precommit
+        run: |
+          cd competitions/trace_the_ace
+          python v121_staged_transport.py evaluate --dir ../../v121_prepared \
+            --embeddings ../../v121_embedding_shards --out ../../v121_pretrained_semantic_residual.json
+      - name: Show decision
+        run: cat v121_pretrained_semantic_residual.json
+      - uses: actions/upload-artifact@v4
+        with:
+          name: trace-ace-v121-pretrained-semantic-residual
+          path: v121_pretrained_semantic_residual.json
+          retention-days: 14

From 91b232f4b8831ad07ec15f2cd22b13722b1c1589 Mon Sep 17 00:00:00 2001
From: Heath Sanchez <32909488+heathsanchez@users.noreply.github.com>
Date: Fri, 21 Aug 2026 06:33:52 +1200
Subject: [PATCH 20/20] infra: recover frozen V121 tail shards only

---
 .../trace-ace-v121-tail-recovery.yml          | 111 ++++++++++++++++++
 1 file changed, 111 insertions(+)
 create mode 100644 .github/workflows/trace-ace-v121-tail-recovery.yml

diff --git a/.github/workflows/trace-ace-v121-tail-recovery.yml b/.github/workflows/trace-ace-v121-tail-recovery.yml
new file mode 100644
index 00000000..f50b1b46
--- /dev/null
+++ b/.github/workflows/trace-ace-v121-tail-recovery.yml
@@ -0,0 +1,111 @@
+name: Trace Ace V121 Tail Recovery
+
+on:
+  pull_request:
+    branches: [agent/trace-ace-mastery-events, agent/v111-runner]
+    paths:
+      - '.github/workflows/trace-ace-v121-tail-recovery.yml'
+  workflow_dispatch:
+
+concurrency:
+  group: trace-ace-v121-tail-recovery
+  cancel-in-progress: false
+
+env:
+  PREPARED_RUN_ID: '32400309220'
+  ORIGINAL_RUN_ID: '32402681183'
+
+jobs:
+  embed_tail:
+    strategy:
+      fail-fast: false
+      max-parallel: 16
+      matrix:
+        shard: [65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127]
+    runs-on: ubuntu-24.04
+    timeout-minutes: 10
+    permissions:
+      actions: read
+      contents: read
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+          cache: pip
+      - name: Install exact embedding dependencies
+        run: python -m pip install --disable-pip-version-check numpy fastembed==0.8.0
+      - name: Download exact frozen V121 prepared artifact
+        uses: actions/download-artifact@v4
+        with:
+          name: v121-prepared
+          path: v121_prepared
+          repository: heathsanchez/mathgraph
+          run-id: ${{ env.PREPARED_RUN_ID }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+      - name: Verify frozen manifest
+        run: |
+          test -f v121_prepared/manifest.json
+          grep -q 'b1612f9fe4558680e468afb2a2452b75c603c244934fe62f7345feee68a61bc1' v121_prepared/manifest.json
+      - name: Embed exact frozen shard ${{ matrix.shard }}
+        run: |
+          cd competitions/trace_the_ace
+          python v121_embed_batch1_transport.py --dir ../../v121_prepared \
+            --shard ${{ matrix.shard }} --shards 128 \
+            --out ../../v121_embeddings_shard_${{ matrix.shard }}.npz
+      - uses: actions/upload-artifact@v4
+        with:
+          name: v121-embeddings-${{ matrix.shard }}
+          path: v121_embeddings_shard_${{ matrix.shard }}.npz
+          retention-days: 2
+          compression-level: 0
+
+  evaluate_union:
+    needs: embed_tail
+    runs-on: ubuntu-24.04
+    timeout-minutes: 20
+    permissions:
+      actions: read
+      contents: read
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+          cache: pip
+      - name: Install evaluation dependencies
+        run: python -m pip install --disable-pip-version-check numpy pandas scipy scikit-learn fastembed==0.8.0
+      - name: Download exact frozen V121 prepared artifact
+        uses: actions/download-artifact@v4
+        with:
+          name: v121-prepared
+          path: v121_prepared
+          repository: heathsanchez/mathgraph
+          run-id: ${{ env.PREPARED_RUN_ID }}
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+      - name: Download successful original shards 0 through 64
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          mkdir -p v121_embedding_shards/original
+          for i in $(seq 0 64); do
+            gh run download "$ORIGINAL_RUN_ID" -n "v121-embeddings-$i" -D "v121_embedding_shards/original/$i"
+          done
+      - name: Download recovered tail shards 65 through 127
+        uses: actions/download-artifact@v4
+        with:
+          pattern: v121-embeddings-*
+          path: v121_embedding_shards/recovered
+          merge-multiple: false
+      - name: Evaluate unchanged frozen V121 precommit
+        run: |
+          cd competitions/trace_the_ace
+          python v121_staged_transport.py evaluate --dir ../../v121_prepared \
+            --embeddings ../../v121_embedding_shards --out ../../v121_pretrained_semantic_residual.json
+      - name: Show decision
+        run: cat v121_pretrained_semantic_residual.json
+      - uses: actions/upload-artifact@v4
+        with:
+          name: trace-ace-v121-pretrained-semantic-residual
+          path: v121_pretrained_semantic_residual.json
+          retention-days: 14