V126 objective-context alignment residual

.github/workflows/trace-ace-v126-objective-context-alignment.yml

@@ -0,0 +1,60 @@
+name: Trace Ace V126 Objective Context Alignment
+
+on:
+  pull_request:
+    branches: [agent/v121-cache-batch2]
+    paths:
+      - '.github/workflows/trace-ace-v126-objective-context-alignment.yml'
+      - 'competitions/trace_the_ace/v126_objective_context_alignment.py'
+  workflow_dispatch:
+
+env:
+  PREPARED_RUN_ID: '32400309220'
+  EMBEDDING_RUN_ID: '32402681183'
+
+jobs:
+  evaluate:
+    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 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: grep -q 'b1612f9fe4558680e468afb2a2452b75c603c244934fe62f7345feee68a61bc1' v121_prepared/manifest.json
+      - name: Download all 128 frozen embeddings
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          mkdir -p v121_embedding_shards
+          for i in $(seq 0 127); do
+            gh run download "$EMBEDDING_RUN_ID" -n "v121-embeddings-$i" -D "v121_embedding_shards/$i"
+          done
+          test "$(find v121_embedding_shards -name 'v121_embeddings_shard_*.npz' | wc -l)" -eq 128
+      - name: Run frozen V126
+        run: |
+          cd competitions/trace_the_ace
+          python v126_objective_context_alignment.py --dir ../../v121_prepared \
+            --embeddings ../../v121_embedding_shards --out ../../v126_objective_context_alignment.json
+      - name: Show decision
+        run: cat v126_objective_context_alignment.json
+      - uses: actions/upload-artifact@v4
+        with:
+          name: trace-ace-v126-objective-context-alignment
+          path: v126_objective_context_alignment.json
+          retention-days: 14

competitions/trace_the_ace/v126_objective_context_alignment.py

@@ -0,0 +1,70 @@
+#!/usr/bin/env python3
+"""V126 objective↔context alignment residual over frozen V97/V121 artifacts.
+
+Single primary separator: cosine displacement between the frozen objective-only
+embedding and the frozen objective+context embedding. This tests a relational
+feature that V121's linear residual does not explicitly represent.
+
+Control: within-objective shuffled semantic embedding before displacement.
+No embedding rerun, no parameter sweep, no label-informed routing.
+"""
+import argparse, json
+from pathlib import Path
+import numpy as np
+from scipy.sparse import load_npz
+from v121_pretrained_semantic_residual import (
+    p97_oof, residual_oof, within_objective_shuffle, collision_mask, ll, MODEL_NAME
+)
+from v121_staged_transport import load_embeddings
+
+EPS=1e-12
+
+def cosine_displacement(a,b):
+    an=np.linalg.norm(a,axis=1); bn=np.linalg.norm(b,axis=1)
+    cos=np.sum(a*b,axis=1)/np.maximum(an*bn,EPS)
+    return (1.0-np.clip(cos,-1,1)).reshape(-1,1)
+
+def eval_geom(name,groups,X75,Xr,y,support,objectives,D,Dsh):
+    P,splits=p97_oof(X75,Xr,y,groups,support)
+    Q=residual_oof(P,D,y,splits); Qsh=residual_oof(P,Dsh,y,splits)
+    base=ll(y,P); real=ll(y,Q); sh=ll(y,Qsh)
+    mask=collision_mask(P,y,objectives,tol=.01)
+    out={'geometry':name,'rows':int(len(y)),'groups':int(len(np.unique(groups))),
+         'baseline_v97_ll':float(base),
+         'alignment':{'ll':float(real),'gain':float(base-real)},
+         'alignment_shuffled_within_objective':{'ll':float(sh),'gain':float(base-sh)},
+         'alignment_minus_shuffle_gain':float(sh-real),
+         'hard_collision':{'rows':int(mask.sum())}}
+    if mask.any():
+        b=ll(y[mask],P[mask]); r=ll(y[mask],Q[mask]); s=ll(y[mask],Qsh[mask])
+        out['hard_collision'].update({'baseline_ll':float(b),'alignment_ll':float(r),
+            'alignment_gain':float(b-r),'shuffled_ll':float(s),'alignment_minus_shuffle_gain':float(s-r)})
+    return out
+
+def main(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=True)
+    y=z['y']; objectives=z['objectives']; support=z['support']; sessions=z['sessions']
+    E_obj,E_sem=load_embeddings(Path(a.embeddings))
+    E_sh=within_objective_shuffle(E_sem,objectives)
+    D=cosine_displacement(E_obj,E_sem); Dsh=cosine_displacement(E_obj,E_sh)
+    res={'protocol':'V126_OBJECTIVE_CONTEXT_ALIGNMENT','model':MODEL_NAME,'rows':int(len(y)),
+         'primary_separator':'1-cosine(objective_embedding, objective_plus_context_embedding)',
+         'control':'within-objective shuffled context embedding before displacement',
+         'precommit':{'promote_gain_each_geometry':.0015,'phase_change_gain_each_geometry':.003,
+                      'real_minus_shuffle_each_geometry':.001,'no_parameter_sweep':True}}
+    res['objective_grouped']=eval_geom('objective_grouped',objectives,X75,Xr,y,support,objectives,D,Dsh)
+    res['session_grouped']=eval_geom('session_grouped',sessions,X75,Xr,y,support,objectives,D,Dsh)
+    def promote(r): return r['alignment']['gain']>=.0015 and r['alignment_minus_shuffle_gain']>=.001
+    def phase(r): return r['alignment']['gain']>=.003 and r['alignment_minus_shuffle_gain']>=.001
+    po,ps=promote(res['objective_grouped']),promote(res['session_grouped'])
+    ph=phase(res['objective_grouped']) and phase(res['session_grouped'])
+    if ph: verdict='PHASE_CHANGE_CANDIDATE'
+    elif po and ps: verdict='PROMOTE_ALIGNMENT_LAW'
+    else: verdict='SUPPRESS_ALIGNMENT_RESIDUAL'
+    res['decision']={'objective_pass':bool(po),'session_pass':bool(ps),'verdict':verdict}
+    Path(a.out).write_text(json.dumps(res,indent=2)); print(json.dumps(res,indent=2),flush=True)
+
+if __name__=='__main__':
+    p=argparse.ArgumentParser(); p.add_argument('--dir',required=True); p.add_argument('--embeddings',required=True); p.add_argument('--out',required=True); main(p.parse_args())