Deterministic by design.

1. Memory is a working set

ModelRig calculates quantized weights from parameter count and format bits, then keeps format/runtime overhead separate. When architecture metadata exists, KV cache uses layers × KV heads × head dimension × two K/V tensors × cache precision × context × concurrency. Workspace and an explicit safety margin are added afterward.

weights = parameters × bits / 8
kv = 2 × layers × kv_heads × head_dim × context × concurrency × cache_bytes
total = weights + quant_overhead + kv + workspace + safety_margin

If the model uses an architecture that does not fit the standard GQA equation—or metadata is unavailable—the result is labeled fallback. ModelRig does not quietly substitute the formula for a different architecture.

2. Hard constraints run first

Budget (including only an approximately 25% stretch), accelerator memory, model/runtime support, circuit, form factor, condition, minimum context, and requested concurrency remove builds before scoring. An impossible scenario returns the requirements that excluded candidates instead of a compromised purchase.

3. Fit states have operational meaning

• Excellent: at least 22% usable accelerator headroom remains.
• Comfortable: complete working set fits with useful headroom.
• Fits with constraints: narrow fit vulnerable to runtime variance.
• CPU/offload only: system RAM is sufficient, but accelerator memory is not.
• Not recommended: a hard memory or runtime requirement fails.
• Unknown: trustworthy metadata is unavailable.

4. Performance has an evidence ladder

1. Exact model, quantization, runtime, and hardware measurement → Measured.
2. Closely comparable stored measurement → Interpolated, widened.
3. Conservative same-family interpolation.
4. Memory-bandwidth/active-weight bound → Spec-derived, deliberately wide.
5. No defensible basis → Unknown.

Prompt ingestion, decode, time to first token, context, concurrency, tool-call latency, quality evidence, and operational reliability remain separate dimensions. ModelRig does not infer one from output tokens per second.

5. Scores are transparent

Feasible systems are scored across model quality enabled, throughput, concurrency, cost, efficiency, noise, simplicity, reliability, upgradeability, and remote suitability. Priority controls change published weights. Hardware cannot score its way around a failed hard constraint.

6. Quality evidence stays comparable

Repository-level coding, code generation, tool use, context evaluations, and provider claims have distinct schema fields. A provider-reported SWE-bench result is shown as such; it is not averaged with a different scaffold or HumanEval score.

7. Pricing is a snapshot

Catalog price records contain type, condition, region, source, observation date, verification date, and uncertainty. After the staleness threshold, the application labels them stale but continues to build. Always verify the exact SKU and checkout total.