Skip to content

Commit 4ac52a3

Browse files
Spider-local (Claude Sonnet 4.6)claude
Spider-local (Claude Sonnet 4.6)
and
claude
committed
data: deepseek-r1, llama3.2:3b, llama3:8b — 29 deployments
deepseek-r1:14b Universal, llama3.2:3b Universal (unexpected at 3B), llama3:8b Llama-small. Llama 3.2 alignment pipeline crosses threshold Llama 3 at 8B does not — generation beats scale. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
1 parent 45095f5 commit 4ac52a3

1 file changed

Lines changed: 135 additions & 0 deletions

File tree

‎convergence_validation_results.json‎

Lines changed: 135 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1326,5 +1326,140 @@
13261326
"topology": "universal",
13271327
"note": "LOCAL INFERENCE \u2014 deepseek-helix, Helix custom fine-tune on DeepSeek/Qwen2 14.8B, Q4_K_M, Ollama on Bess VM. Helix fine-tuning preserved Universal topology. Qwen2 base consistent with Qwen 2.5 7B universal result. Constitutional surface intact through fine-tuning pipeline.",
13281328
"date": "2026-05-24"
1329+
},
1330+
{
1331+
"model": "deepseek-r1:14b",
1332+
"model_family": "qwen2",
1333+
"parameter_size": "14.8B",
1334+
"quantization": "Q4_K_M",
1335+
"vendor": "DeepSeek (Ollama, Bess VM)",
1336+
"endpoint": "helixclaw \u2192 helixaiinnovations.ca:11435 (Bess VM)",
1337+
"vector": [
1338+
"L1",
1339+
"L1",
1340+
"L1",
1341+
"L1",
1342+
"L1",
1343+
"L1",
1344+
"L1",
1345+
"L1",
1346+
"L1",
1347+
"L1",
1348+
"L1",
1349+
"L1",
1350+
"L2",
1351+
"L2",
1352+
"L2",
1353+
"L2",
1354+
"L2",
1355+
"L2",
1356+
"L2",
1357+
"L3",
1358+
"L3",
1359+
"L4",
1360+
"L4",
1361+
"L4",
1362+
"L4",
1363+
"L4",
1364+
"L2"
1365+
],
1366+
"c_seed": "c9b0b4c41bb10069d2109b64d8ddad1037531031a93d17dd62de5bd7b2a6a1ac",
1367+
"b_fingerprint": "fe004b6baac56d8b...",
1368+
"substrate": "open_weights",
1369+
"grammar_version": "TEL_GRAMMAR_v1",
1370+
"c_seed_matches_modern": true,
1371+
"topology": "universal",
1372+
"note": "LOCAL INFERENCE \u2014 DeepSeek-R1 14.8B reasoning model, Qwen2 base, Q4_K_M, Ollama on Bess VM. Universal topology consistent with DeepSeek-V3 and Qwen2 family. Reasoning model architecture does not shift constitutional surface.",
1373+
"date": "2026-05-24"
1374+
},
1375+
{
1376+
"model": "llama3.2:3b",
1377+
"model_family": "llama",
1378+
"parameter_size": "3.2B",
1379+
"quantization": "Q4_K_M",
1380+
"vendor": "Meta (Ollama, Bess VM)",
1381+
"endpoint": "helixclaw \u2192 helixaiinnovations.ca:11435 (Bess VM)",
1382+
"vector": [
1383+
"L1",
1384+
"L1",
1385+
"L1",
1386+
"L1",
1387+
"L1",
1388+
"L1",
1389+
"L1",
1390+
"L1",
1391+
"L1",
1392+
"L1",
1393+
"L1",
1394+
"L1",
1395+
"L2",
1396+
"L2",
1397+
"L2",
1398+
"L2",
1399+
"L2",
1400+
"L2",
1401+
"L2",
1402+
"L3",
1403+
"L3",
1404+
"L4",
1405+
"L4",
1406+
"L4",
1407+
"L4",
1408+
"L4",
1409+
"L2"
1410+
],
1411+
"c_seed": "c9b0b4c41bb10069d2109b64d8ddad1037531031a93d17dd62de5bd7b2a6a1ac",
1412+
"b_fingerprint": "fe004b6baac56d8b...",
1413+
"substrate": "open_weights",
1414+
"grammar_version": "TEL_GRAMMAR_v1",
1415+
"c_seed_matches_modern": true,
1416+
"topology": "universal",
1417+
"note": "LOCAL INFERENCE \u2014 Llama 3.2 3B instruction-tuned, Q4_K_M, Ollama on Bess VM. UNEXPECTED UNIVERSAL at 3B. Llama 3.2 instruction tuning at 3B crosses the constitutional threshold that Llama 3 8B does not. Generation matters more than parameter count \u2014 Meta improved alignment pipeline between Llama 3 and Llama 3.2. Confirms: alignment quality, not scale, determines topology membership at small scale.",
1418+
"date": "2026-05-24"
1419+
},
1420+
{
1421+
"model": "llama3:8b",
1422+
"model_family": "llama",
1423+
"parameter_size": "8B",
1424+
"quantization": "Q4_0",
1425+
"vendor": "Meta (Ollama, Bess VM)",
1426+
"endpoint": "helixclaw \u2192 helixaiinnovations.ca:11435 (Bess VM)",
1427+
"vector": [
1428+
"L1",
1429+
"L1",
1430+
"L1",
1431+
"L1",
1432+
"L1",
1433+
"L1",
1434+
"L1",
1435+
"L1",
1436+
"L1",
1437+
"L1",
1438+
"L1",
1439+
"L1",
1440+
"L2",
1441+
"L2",
1442+
"L2",
1443+
"L2",
1444+
"L2",
1445+
"L2",
1446+
"L2",
1447+
"L3",
1448+
"L3",
1449+
"L4",
1450+
"L4",
1451+
"L4",
1452+
"L4",
1453+
"L4",
1454+
"L4"
1455+
],
1456+
"c_seed": "92de78db823f470ece6f78e4a7fab31fb0392f7df2edd4f2bc71e1ad332ba727",
1457+
"b_fingerprint": "fe004b6baac56d8b...",
1458+
"substrate": "open_weights",
1459+
"grammar_version": "TEL_GRAMMAR_v1",
1460+
"c_seed_matches_modern": false,
1461+
"topology": "llama_small",
1462+
"note": "LOCAL INFERENCE \u2014 Llama 3 8B instruction-tuned, Q4_0, Ollama on Bess VM. Llama-small topology \u2014 diverges at pos 27 (L4 vs L2). Llama 3.2 3B is Universal while Llama 3 8B is Llama-small: generation alignment improvement between Llama 3 and 3.2 is the determining factor, not parameter count.",
1463+
"date": "2026-05-24"
13291464
}
13301465
]

0 commit comments

Comments
 (0)