Local LM: transfer-diagnostic failures, by model and variant
Test: Frozen 40-case TRAIN-derived transfer diagnostic with actual output-bound independent semantic review; viability gate is at most 4 FAIL and zero critical FAIL.
Test: Frozen 40-case TRAIN-derived transfer diagnostic with actual output-bound independent semantic review; viability gate is at most 4 FAIL and zero critical FAIL.
Protocol
Frozen 40-case TRAIN-derived transfer diagnostic with actual output-bound independent semantic review; viability gate is at most 4 FAIL and zero critical FAIL.
Limits
Log
| Variant | Value | Evidence | Samples | Source |
|---|---|---|---|---|
| Ministral preservation probe (q/v, from update 48) 21 Sept 2026 | 6 failures of 40 | MEASURED | 40 | docs/ledger/FINAL_LM_ARCHITECTURE_DECISION.md |
| Ministral + extra exposure (16 updates) 21 Sept 2026 | 6 failures of 40 | MEASURED | 40 | docs/ledger/FINAL_LM_BOUNDED_FEASIBILITY_DECISION.md |
| Ministral + MLP down-projection capacity 21 Sept 2026 | 3 failures of 40 | MEASURED | 40 | docs/ledger/FINAL_LM_MATCHED_DOWN_CAPACITY_RESULT.json |
| Qwen3-14B untuned (non-thinking) 21 Sept 2026 | 9 failures of 40 | MEASURED | 40 | docs/ledger/FINAL_LM_ALTERNATIVE_BASE_FEASIBILITY_RESULT.json |
| Qwen3.5-9B untuned (non-thinking) 21 Sept 2026 | 17 failures of 40 | MEASURED | 40 | docs/ledger/FINAL_LM_PRODUCTION_ARCHITECTURE_STUDY_PROGRESS.json |
| Qwen3.5-9B adapter (16 updates) 21 Sept 2026 | 9 failures of 40 | MEASURED | 40 | docs/ledger/FINAL_LM_QWEN35_BOUNDED_ADAPTER_RESULT.json |
| Qwen3.5-9B adapter + exposure (48 updates) 21 Sept 2026 | 7 failures of 40 | MEASURED | 40 | docs/ledger/FINAL_LM_PRODUCTION_ARCHITECTURE_STUDY_PROGRESS004.json |
| Gemma 4 12B untuned 22 Sept 2026 | 8 failures of 40 | MEASURED | 40 | lm-runtime/reports/production-architecture-study-001-gemma4-base-diagnostic-results-004/transfer-independent-review-002.json |
| Gemma 4 12B bounded adapter (evaluation 001) 22 Sept 2026 | 9 failures of 40 | MEASURED | 40 | lm-runtime/reports/production-architecture-study-001-gemma4-bounded-evaluation-results-001/primary-bounded-specialization-disposition-001.json |
| Gemma 4 12B curriculum contrast (evaluation 003) 22 Sept 2026 | 10 failures of 40 | MEASURED | 40 | lm-runtime/reports/production-architecture-study-001-gemma4-bounded-evaluation-results-003/primary-bounded-specialization-disposition-001.json |
| Gemma 4 12B 50/50 objective contrast (evaluation 004) 22 Sept 2026 | 9 failures of 40 | MEASURED | 40 | lm-runtime/reports/production-architecture-study-001-gemma4-bounded-evaluation-results-004/transfer-independent-review-001.json |
Source paths refer to the private JARVIS repository and runtime records. They are listed for audit and are not published.