triavalabs/gemma4-26b-mode-split: v1.0.0 — Substitution/synthesis mode split on Gemma 4 26B MoE

First formal release of the cross-model substitution/synthesis mode-split measurement on Gemma 4 26B MoE (sovereign Ollama). Archived to Zenodo for citable reference as Track 1 of an ongoing cross-stack research collaboration.

Headline

| Cap | Prompt type | Success | eval_count (mean) | Unique outputs | |------|--------------|---------|-------------------|----------------| | 400 | substitution | 20/20 | 38.0 (flat) | 1/20 | | 400 | synthesis | 20/20 | 50.7 (45–75) | 6/20 | | 4096 | substitution | 20/20 | 38.0 (flat) | 1/20 | | 4096 | synthesis | 20/20 | 54.5 (45–78) | 9/20 |

Substitution mode is byte-identical and cap-invariant: eval_count 38 flat, a single unique response across 40 calls, identical at cap=400 and cap=4096. Light-synthesis fans out (6–9/20 unique). All 80 calls returned done_reason=stop; the synthesis cap-floor observed on gemma4:e4b did not appear on 26B.

Methodology

• Endpoint: sovereign Ollama on Hetzner CCX33
• Model: gemma4:26b (25.8B params, Q4_K_M, base, no Modelfile system prompt)
• Temperature: 0.2; think: false (visible output only)
• N=20 per cell; 2×2 matrix (cap × prompt type)
• e-commerce return-policy fixture; substitution = verbatim retrieval, synthesis = single refund decision under conflicting policy clauses (light synthesis)
• Protocol mirrors Hashevolution/James-RAG-Evol PR #440 for direct cross-model comparison

Contents

• sweep.py — sweep driver (stdlib only)
• v3prime-26b-mode-split-20260523T120924.json — raw 80-call data
• analysis.md — three-finding analysis with cross-model comparison

Context

Track 1 (sovereign Gemma 4 26B MoE) of a three-stack collaboration on Gemma 4 reasoning behavior. Substitution-mode determinism reproduces the gemma4:e4b result (Hashevolution/JAMES) on a different model and stack; 26B substitution (38 tokens) is leaner than e4b substitution (~62 tokens). The synthesis cap-floor is checkpoint-isolated to gemma4:e4b and does not appear here, consistent with cross-family findings.

Related work: Robin Converse — substitution/synthesis mode split; Hashevolution/JAMES — workload-gradient quantification (PR #440); Ali Afana — managed-API cross-stack context; Vadym Arnaut — substitution-vs-decision boundary framing.