scripts/iv/05_label_origin_test.py

Derived analysis output (scratchpad scripts/iv/05_label_origin_test.py).

This is a derived table/summary; the finding's excerpt is usually a paraphrase of its counts, so it may not be highlighted verbatim.

# Test: are "wrong" labels copied from the page being edited (visible text) or inherited from elsewhere?
import json,csv,re,collections,random
S='/tmp/claude-0/-root-swarm-hackathon/a8e4c17a-c034-48ae-9f11-70fc0a8841c9/scratchpad'
F={r['rev_id']:r for r in csv.DictReader(open(S+'/derived/iv/wiki_rev_features.csv'))}
def norm(s): return re.sub(r'[^a-z0-9]','',s.lower()).replace('openai','oai')
prevbody={}; res=collections.Counter(); allbodies=[]
revs=[json.loads(l) for l in open(S+'/wikilogs/revisions.jsonl')]; revs.sort(key=lambda r:(r['time'],r['rev_id']))
labels=[r['label'] for r in revs if r['label']]
random.seed(1)
for r in revs:
    f=F[r['rev_id']]; pb=prevbody.get(r['page_key'],'')
    if f['sigs'] and '|' not in f['sigs'] and r['label'] and not r['label'].startswith('['):
        kind='match' if norm(r['label'])==norm(f['sigs']) else 'mismatch'
        res[(kind,'label_in_prior_page_body', r['label'] in pb)]+=1
        rl=random.choice(labels); res[(kind,'random_label_in_prior_page_body', rl in pb)]+=1
    prevbody[r['page_key']]=r.get('body') or pb
for k in sorted(res): print(k,res[k])