universality-classes.json: duplicate class_id resolution¶

Date: 2026-05-14 Affected file: web/frontend/assets/data/universality-classes.json Source of issue: W6-E (session #3 wave 6, regression test) flagged 2 duplicate class_id values Fixed in commit: bfdf2b0 (v4/fix (F1): P1 bugs — dedupe class_id + phase.bytedance.city React error), session #3 Verified in this maintenance pass (2026-05-14): all 23 entries have unique class_id

What was duplicated¶

Two pairs of Louvain sub-communities had identical class_id because the curation script (manual layer + LLM layer) produced two independent records that resolved to the same physics prototype slug:

Both pairs were legitimately distinct Louvain sub-communities (different members, different domains, different hub nodes) — not data-pipeline mistakes. They collided only on class_id because both sub-communities mapped to the same physics prototype name (Motter-Lai, Gardner-Collins).

Resolution decision¶

Policy chosen: rename — keep both entries, suffix the lower-rank (LLM-curated, smaller size) entry with _v2. No information loss.

Rationale:

1. Drop is wrong: each sub-community contains real domain members that the cross-domain analysis depends on. Dropping the LLM-curated _v2 entries would lose verified cross-domain edges.
2. Merge is also wrong for SPLIT: B3 consensus on the motter_lai_* pair was explicitly SPLIT — the human-in-the-loop critic panel said these should remain split despite sharing a physics prototype label.
3. Rename preserves both with no data loss: downstream consumers (/classes page, KB embedding pipeline, taxonomy-v2 cross-ref) treat them as distinct classes again. The taxonomy_match field on the _v2 entries still points to the parent prototype slug, so the linkage is preserved for documentation purposes.

Post-fix state (verified 2026-05-14)¶

>>> import json
>>> data = json.load(open("web/frontend/assets/data/universality-classes.json"))
>>> ids = [c["class_id"] for c in data["classes"]]
>>> len(ids), len(set(ids))
(23, 23)
>>> # No duplicates.

All 23 class_id values are now unique:

The remaining taxonomy_match collisions (soc_threshold_cascade, hysteresis_first_order_transition, multistable_self_fulfilling, preferential_attachment — each appearing 2× across the 23 classes) are not bugs: they are the intended mapping from Louvain sub-communities to the upper-level 35-class taxonomy. Multiple sub-communities legitimately share a parent taxonomy node.

Pending follow-up (out of scope of this fix)¶

The two gardner_collins_toggle_switch* entries both carry b3_consensus: MERGE, meaning the critic panel recommends eventually merging them after collecting more cross-domain evidence. This is research follow-up, not a data-integrity bug. Tracked in docs/sessions/HANDOFF.md § P2 backlog ("review B3 MERGE verdicts after Layer 5 Phase 4+ data lands").

Provenance / how to reproduce the check¶

cd ~/Projects/structural-isomorphism
./.venv/bin/python -c "
import json
from collections import Counter
data = json.load(open('web/frontend/assets/data/universality-classes.json'))
classes = data['classes']
ids = [c['class_id'] for c in classes]
counter = Counter(ids)
dupes = {k:v for k,v in counter.items() if v>1}
assert not dupes, f'duplicate class_ids: {dupes}'
print(f'OK: {len(ids)} unique class_ids')
"

Suggested CI guard: add the above assertion to the web/backend/tests/ suite so a regression is caught automatically.