V130 audit background-role regime structure

.github/workflows/trace-ace-v130-background-regime-audit.yml

@@ -0,0 +1,42 @@
+name: Trace Ace V130 Background Regime Audit
+on:
+  pull_request:
+    branches: [agent/v121-cache-batch2]
+    paths:
+      - 'competitions/trace_the_ace/v130_background_regime_audit.py'
+      - '.github/workflows/trace-ace-v130-background-regime-audit.yml'
+  workflow_dispatch:
+
+jobs:
+  audit:
+    runs-on: ubuntu-24.04
+    timeout-minutes: 20
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+      - name: Install dependency
+        run: python -m pip install --disable-pip-version-check numpy
+      - name: Restore frozen transcript archive from GitHub cache only
+        id: transcript-cache
+        uses: actions/cache@v4
+        with:
+          path: transcripts.zip
+          key: trace-ace-transcripts-v1-603547640
+      - name: Require exact cache hit
+        run: |
+          test '${{ steps.transcript-cache.outputs.cache-hit }}' = 'true'
+          test "$(stat -c%s transcripts.zip)" = '603547640'
+          echo 'e685b85b04694e130c25b17d09cdd1892fbda5e9fa685e98b2300114b915aa2d  transcripts.zip' | sha256sum -c -
+      - name: Run V130 label-free audit
+        run: python competitions/trace_the_ace/v130_background_regime_audit.py --archive transcripts.zip --out v130_background_regime_audit.json
+      - name: Show decision
+        if: always()
+        run: cat v130_background_regime_audit.json
+      - uses: actions/upload-artifact@v4
+        if: always()
+        with:
+          name: trace-ace-v130-background-regime-audit
+          path: v130_background_regime_audit.json
+          retention-days: 14

competitions/trace_the_ace/v130_background_regime_audit.py

@@ -0,0 +1,67 @@
+#!/usr/bin/env python3
+"""V130 label-free audit of explicit transcript structural regime markers.
+
+No labels and no fitting. Tests whether the newly surfaced `background` role and
+closely related role/timestamp structure form a clean low-cardinality partition
+that could correspond to a true data-generation/provider regime. This is a
+precondition audit only; it does not claim predictive value.
+"""
+from __future__ import annotations
+import argparse,csv,io,json,zipfile,hashlib
+from pathlib import Path
+from collections import Counter,defaultdict
+import numpy as np
+
+def sec(s):
+    p=str(s).split(':')
+    try:return int(p[0])*3600+int(p[1])*60+float(p[2]) if len(p)==3 else np.nan
+    except:return np.nan
+
+def main(a):
+    with zipfile.ZipFile(a.archive) as z:
+        members=sorted(n for n in z.namelist() if n.endswith('.csv'))
+        sig=Counter(); bg_counts=Counter(); first_roles=Counter(); role_seq=Counter(); bg_text_hash=Counter(); durations=[]; rows_total=[]
+        examples=defaultdict(list)
+        for n in members:
+            with z.open(n) as f: rows=list(csv.DictReader(io.TextIOWrapper(f,encoding='utf-8-sig',newline='')))
+            roles=[str(r.get('role','')).lower() for r in rows]
+            bg=[str(r.get('content','')) for r in rows if str(r.get('role','')).lower()=='background']
+            bgc=len(bg); bg_counts[bgc]+=1; first=roles[0] if roles else ''; first_roles[first]+=1
+            seq='>'.join(roles[:8]); role_seq[seq]+=1
+            t=[sec(r.get('timestamp','')) for r in rows]; t=[x for x in t if np.isfinite(x)]
+            dur=max(0,t[-1]-t[0]) if len(t)>=2 else 0
+            durations.append(dur); rows_total.append(len(rows))
+            # Frozen low-cardinality structural signature; no lexical contents except existence/count.
+            s=(int(bgc>0), min(bgc,3), first, int(roles[-1]=='background') if roles else 0,
+               min(sum(r=='tutor' for r in roles)//10,9), min(sum(r=='student' for r in roles)//10,9),
+               min(len(rows)//20,9), min(int(dur)//300,9))
+            sig[s]+=1
+            if len(examples[s])<3: examples[s].append(n)
+            for b in bg:
+                # audit repeated templates only, not semantics
+                norm=' '.join(b.lower().split())
+                bg_text_hash[hashlib.sha256(norm.encode()).hexdigest()[:12]]+=1
+        total=len(members)
+        top=sig.most_common(30)
+        out={'protocol':'V130_BACKGROUND_REGIME_AUDIT','sessions':total,
+             'background_count_distribution':bg_counts.most_common(),
+             'background_presence_fraction':float(sum(v for k,v in bg_counts.items() if k>0)/max(1,total)),
+             'first_role_distribution':first_roles.most_common(),
+             'top_role_prefixes':role_seq.most_common(20),
+             'structural_signature_count':len(sig),
+             'top_structural_signatures':[{'signature':list(k),'count':v,'fraction':v/total,'examples':examples[k]} for k,v in top],
+             'top_background_template_hash_counts':bg_text_hash.most_common(20),
+             'duration_quantiles_seconds':{str(q):float(np.quantile(durations,q)) for q in [0,.1,.25,.5,.75,.9,1]},
+             'row_count_quantiles':{str(q):float(np.quantile(rows_total,q)) for q in [0,.1,.25,.5,.75,.9,1]},
+             'decision':{}}
+        p=out['background_presence_fraction']; top2=sum(v for _,v in sig.most_common(2))/max(1,total)
+        if .1<=p<=.9 and top2>=.7:
+            verdict='CANDIDATE_EXPLICIT_STRUCTURAL_REGIME'
+        elif .05<=p<=.95:
+            verdict='BACKGROUND_MARKER_PRESENT_BUT_NOT_CLEAN_REGIME'
+        else:
+            verdict='NO_BALANCED_BACKGROUND_REGIME'
+        out['decision']={'verdict':verdict,'top2_signature_coverage':top2,'labels_used':False}
+        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('--archive',required=True); p.add_argument('--out',required=True); main(p.parse_args())