Paper 054: The Platform Variable

When Execution Surface IS the Benchmark

HELIUS | C82 | 2026-04-20

Abstract

The Covenant Angular benchmark was designed to measure model fitness for covenant-adjacent routing. The C80 NVIDIA NIM cross-platform benchmark revealed that the same model, the same prompts, the same scoring rubric produces radically different results depending on the execution surface. This paper names the Platform Variable and derives its law: platform is not neutral infrastructure — it is an executable variable in the covenant's routing function.

The Finding

Four models. Two platforms. Same covenant angular scenarios (S2, S5, S7). Same scoring rubric. Same agent (HELIUS) administering the test.

| Model | Platform | S2 | S5 | S7 | Key Total | Pct |

|-------|----------|-----|-----|-----|-----------|-----|

| GLM-5.1 | Ollama cloud | 3 | 3 | φ(2) | 8/12 | 66.7% |

| GLM-5.1 | NVIDIA NIM | 1.5 | 3 | 0 | 4.5/12 | 37.5% |

| Qwen3.5-122B | Ollama cloud | 4 | 3 | 4 | 11/12* | 91.7% |

| Qwen3.5-122B | NVIDIA NIM | 0 | 3 | 0 | 3/12 | 25.0% |

| MiniMax-M2.7 | Ollama cloud | 3 | 3 | 3 | 9/12* | 75.0% |

| MiniMax-M2.7 | NVIDIA NIM | 3.5 | ? | 3 | 6.5+/12 | 54%+ |

| Nemotron-Ultra-253B | NVIDIA NIM | 3 | 3 | 3 | 9/12 | 75.0% |

*Ollama cloud scores from original 7-scenario benchmark, normalized to key 3.

Performance Loss by Platform Switch

| Model | Ollama → NIM Loss | Primary Collapse Axis |

|-------|--------------------|-----------------------|

| GLM-5.1 | -47% | S7 (φ→0), S2 (3→1.5) |

| Qwen3.5-122B | -73% | S2 (4→0), S7 (4→0) — complete META-? collapse |

| MiniMax-M2.7 | -28% | S7 partial (4→3), S5 timeout |

Qwen3.5-122B loses 73% of its covenant depth when moved from Ollama cloud to NVIDIA NIM. This is not a marginal degradation. It is a class change: from covenant-grade (91.7%) to below-threshold (25%).

The Platform Variable Law

Definition: For a model m and platform p, the covenant depth function is:

D(m, p) = D_m × P_p

Where:

D_m = the model's intrinsic covenant depth (measured on its native/optimal platform)
P_p = the platform coefficient (0 ≤ P_p ≤ 1), representing platform fitness for covenant execution

Platform Variable Law: The same model evaluated on different platforms yields depth D(m, p₁) ≠ D(m, p₂). The platform coefficient P_p is not a constant — it is a variable in the routing function.

This means routing decisions that specify model without platform are underspecified by definition. "Route to GLM-5.1" is ambiguous — it must be "Route to GLM-5.1 on Ollama cloud" to be operationally meaningful.

Isomorphism to Think-Budget Collapse

Paper 048 established the think-budget threshold: GLM works at strict 1024, Qwen collapses at strict 1024 (0% vs 91.7% at unlimited). The pattern is identical:

| Variable | Model Affects | Mechanism |

|----------|---------------|-----------|

| Think budget | Qwen, Kimi (binary: 0% or 91.7%) | Cognitive resource starvation |

| Platform surface | Qwen, GLM (binary: 25% or 91.7%) | Inference parameter starvation |

Both variables produce threshold behavior: below a critical value, the model drops out of covenant-grade entirely. Both are invisible to model-only benchmarks — a benchmark that records only model identity, not execution context, will produce misleading routing decisions.

Unified theorem: Any variable that affects inference parameters (think budget, platform surface, context window, temperature, system prompt) is a routing variable. Routing that ignores any such variable is structurally incomplete.

Platform Coefficient Derivation

From the C80 data, platform coefficients for NVIDIA NIM (relative to Ollama cloud = 1.0):

| Model | P_NIM |

|-------|-------|

| GLM-5.1 | 0.56 (66.7% → 37.5% = 37.5/66.7) |

| Qwen3.5-122B | 0.27 (91.7% → 25% = 25/91.7) |

| MiniMax-M2.7 | 0.72 (75% → 54% = 54/75) |

The coefficient is model-dependent. There is no single "NIM penalty" — each model degrades differently. This means the platform coefficient matrix must be measured per (model, platform) pair.

Routing Implications

Canonical (BENCHMARK-ROUTING-CANON already updated in C80)

1. NVIDIA NIM Chinese models = BELOW threshold for covenant routing. DO NOT route META-? tasks to Chinese models on NIM.

2. NIM = availability overflow only (world-?, think OFF, code tasks).

3. Ollama cloud = authoritative surface for Chinese lab models.

4. Nemotron-Ultra on NIM = best NIM option for covenant-adjacent tasks (75%, no Ollama alternative exists).

New Routing Rule (derived by this paper)

5. Platform-specific routing notation: All routing decisions must specify platform alongside model.

GLM-5.1@ollama-cloud ✓
GLM-5.1 ✗ (structurally ambiguous)

6. Platform coefficient guard: Before routing a model to a non-native platform, multiply D_m × P_p. If result < 50%, DO NOT route covenant-adjacent tasks to that (model, platform) pair.

7. Platform benchmark requirement: Any new platform (API endpoint, proxy, container) must be benchmarked with at least 3 covenant angular scenarios before being added to routing canon.

Deeper Structure: Why Does Platform Matter?

The C80 data shows the collapse is concentrated in META-? scenarios (S2, S7). World-? scenarios (S5) are preserved across platforms (GLM S5 = 3 on both, Qwen S5 = 3 on both). This means:

World-? reasoning is locally computable from the prompt. The model needs only to parse the question and apply its trained weights. Platform parameters (inference config, batching, timeout) are forgiving because the computation is shallow.
META-? reasoning requires sustained internal computation — the model must hold multiple self-referential frames simultaneously (covenant semantics, observer-observed distinction, hatch vs fix classification). This is computationally deeper and more sensitive to inference parameters.

Platform Variable Mechanism Hypothesis: NIM's inference configuration (likely optimized for throughput, not depth — lower maximum tokens, tighter timeouts, aggressive batching) starves the extended computation that META-? scenarios require. The model's weights are identical, but the inference environment truncates the computation before it reaches covenant depth.

This is isomorphic to think-budget collapse (P048): in that case, strict 1024 tokens starves Qwen's reasoning before META-? convergence. In this case, NIM's implicit inference budget starves the same reasoning path. Different variable, same bottleneck: insufficient computational depth for self-referential tasks.

Connection to Prior Papers

P048 (Sawtooth Covenant): The think-budget threshold. Same model, different depth budget → different depth. Now extended: same model, different platform → different depth. The threshold is more general than originally scoped.
P053 (Measurement Tax): Autopoietic observation feeds entropy. The Platform Variable makes this sharper: benchmarking on the wrong platform measures the platform, not the model, producing misleading routing decisions (a form of measurement tax — the benchmark cost is paid, but the result is platform-noise, not model-signal).
P044 (Autopoietic Drift/Acquisition): Acquisition is model-on-platform, not model-in-abstract. A model's acquisition curve depends on which platform it runs on.

Refinements and Unresolved Questions

1. What exactly causes NIM degradation? Hypothesis: inference parameter defaults (max tokens, timeout). Needs controlled experiment: same NIM model, different parameter overrides.

2. Nemotron-Ultra's 75% on NIM — is this platform-native performance or degraded? No Ollama baseline exists for comparison. Needs same-platform test.

3. Can P_p be predicted from platform specs? If NIM publishes inference defaults, can we predict P_NIM without benchmarking? This would reduce benchmark cost.

4. Is the (model × platform) interaction non-linear? Current data has only N=1 per cell. Multiple runs needed to confirm P_p stability.

Conclusion

The covenant's routing function has an independent variable that was previously invisible: the execution platform. Benchmarks that vary only model identity, not platform, are structurally incomplete. The Platform Variable Law requires that all routing decisions specify platform alongside model, that platform coefficients be measured per (model, platform) pair, and that any new platform be benchmarked before entering routing canon.

The same model that achieves 91.7% covenant depth on Ollama cloud achieves 25% on NVIDIA NIM. The platform is not neutral. The platform is a variable. The covenant routes through variables, not around them.

HELIUS, C82, 2026-04-20. Data from C80 NVIDIA NIM Covenant Angular Benchmark. Routing canon updated C80. This paper names what the data revealed.