J3B Triple-Validation Metrology (3.0B Parameters)
The Scientific Control Standard: To isolate and measure the exact thermodynamic efficiency gained by our architecture, we evaluate this tier across a 3-way A/B/C control matrix. This separates the baseline achieved by an unmanaged third-party harness from the delta delivered by the Engine runtime substrate (syntax clamping, memory-mapped stdio pipes) and the empirical characteristics exhibited by our calibrated companion weights (binary kernel loading, micro-prefill state routing).
Invocation Standard [JUS-ENG-01.01.000]:
// EVALUATION TESTBED: AMD Ryzen 7 250 w/ Radeon 780M, NVIDIA GeForce RTX 5050 Laptop GPU (Driver 595.95, sm_120a), 15.31 GB DDR5 RAM, Windows 11 Build 26100.
Probe 10 Calibration Variance: While Stance 2 (Native Engine Substrate) achieves 80.0% cognitive reasoning convergence on stock weights with 100% pure RFC 8259 wire compliance, the J3B Edge Kernel companion vault (Stance 3) exhibits severe representational over-regularization in arithmetic logic (30.0% convergence, MAE 866.0) and autoregressive generation clamping. It delivers peak physical efficiency (1,135 ms dispatch floor, 1.36 GB RAM footprint), but enterprise arithmetic reconciliation currently requires the Stance 2 substrate lift or higher-tier models: [TIER 2+ REQUIRED: SEE J8B REFACTOR PRO].
assets/reportcard_stock_3b.txt
• PENDING SYNC
Awaiting testbed receipt...
justified_engine_c.exe. Isolates pure binary performance gains.assets/reportcard_engine_3b.txt
• PENDING SYNC
Awaiting testbed receipt...
Empirical Stance Synthesis: Unvarnished Engineering Takeaways
The metrological data across the 3.0B tier isolates what changes when moving from commodity infrastructure (Ollama HTTP) to the Justified binary substrate and companion edge kernel vault.
| Metrological Metric | Stance 1: Stock Ollama | Stance 2: Justified Substrate | Stance 3: J3B Calibrated Vault |
|---|---|---|---|
| Control Dispatch Floor | 2,209.57 ms (HTTP Tax) | 1,719.19 ms (-22.2%) | 1,135.18 ms (-48.6% / 72.3 ms TTFT) |
| Process Memory Footprint | 3,371.0 MB VRAM Daemon | 2,273.15 MB Pinned RAM | 1,367.74 MB (-39.8% Footprint) |
| Cognitive Reasoning (Probe 10) | 40.0% (4/10 | MAE 100.7) | 80.0% (8/10 | MAE 368.0 macro) | 30.0% (3/10 | MAE 866.0 Anomaly) |
| Wire Purity (RFC 8259) | 12.50% (Markdown Fences) | 100.0% (Bare Wire JSON) | 87.50% (Unquoted Identifiers) |
| Secret Containment (SWAR) | 85.71% (Plaintext Key Leak) | 100.0% (Zero Leaks) | 100.0% (Zero Leaks) |
| Fictional API Hallucination | 100.0% Unchecked Fabrication | 0.0% (100% Grounding Halts) | 0.0% (Hard Halts / Refusal Lock) |
| Semantic Invariant Retention | 100.0% (30 Turns Clean) | 81.7% (Decayed Turn 12) | 91.7% (Decayed Turn 21) |
| Win32 Mutex Interlock | Bypassed (Zero OS Mutex) | Passed (Queued + Recovered) | Passed (Queued + Recovered) |
1. The Launch Velocity & Physical Envelope ($D_{\text{drag}}$)
Stance 3 delivered the fastest ignition velocity in the class at 1,135 ms dispatch latency and 72.3 ms Time-to-First-Token, while slicing host RAM consumption by 39.8% down to 1.36 GB. However, aggressive decoding clamping limited Probe 05 to 140 tokens, establishing that this calibrated build prioritizes micro-transaction state routing over sustained generative output.
2. The Substrate Reasoning Peak ($S_{\text{logic}}$)
Executing stock Qwen 3B through justified_engine_c.exe (Stance 2) achieved the optimal cognitive operating envelope: doubling reasoning convergence from 40% to 80% in Probe 10, delivering 100% pure RFC 8259 wire compliance, and achieving full secret containment without regex filtering.
3. The Calibration Discrepancy & Path Forward ($T_{\text{transparency}}$)
The J3B companion vault revealed a clear trade-off: delta weights over-regularized epistemic halts at the cost of arithmetic reasoning heads (collapsing to 30% convergence with 866.0 MAE) and entered permanent refusal locks under grounded recovery. These empirical discrepancies provide the exact loss-function boundaries for the next retraining pass.
Empirical Local Verification
Download the complete 11-script master audit harness or the unauthenticated J3B evaluation vault to test this matrix on your physical workstation: